📓 Uploaded Jupyter Notebook (.ipynb)
SONA Optimization
SONA is implemented and optimized
In [1]:
import os, time, json, numpy as np, torch
from PIL import Image
from diffusers import StableDiffusionPipeline Out[1]:
/home/hice1/yoh88/Out-of-distribution-Mukhopadhyay-gatech/kernel/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html from .autonotebook import tqdm as notebook_tqdm
In [2]:
os.environ["HF_HOME"] = "/storage/ice1/7/6/yoh88/hf_cache"
device = "cuda" if torch.cuda.is_available() else "cpu"
model_id = "Manojb/stable-diffusion-2-base"
print(f"Loading {model_id} onto {device}...")
t0 = time.perf_counter()
pipe = StableDiffusionPipeline.from_pretrained(
model_id,
cache_dir="/storage/ice1/7/6/yoh88/hf_cache",
torch_dtype=torch.float16 if device =="cuda" else torch.float32
).to(device)
if device == "cuda": torch.cuda.synchronize()
cold_start_time = time.perf_counter() - t0
print(f"Pipeline loaded successfully!")
print(f"Cold Start Load Time: {cold_start_time:.2f} seconds")
print(f"Peak VRAM Allocated: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB") Out[2]:
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Out[2]:
Loading Manojb/stable-diffusion-2-base onto cuda...
Out[2]:
Loading weights: 100%|█████████████████████████████████████████████████████████████████████| 372/372 [00:13<00:00, 27.95it/s] Loading pipeline components...: 100%|██████████████████████████████████████████████████████████| 6/6 [00:35<00:00, 5.97s/it]
Out[2]:
Pipeline loaded successfully! Cold Start Load Time: 36.47 seconds Peak VRAM Allocated: 2.60 GB
In [3]:
def profiled_generate_sona_outlier(
pipe,
input_image_path,
id_class_prompt,
ood_target_prompt,
stop_timestep=20,
guidance_scale=15.0,
threshold_percentile=0.10,
adaptive_mask=False,
device="cuda"
):
"""
Executes baseline SONA with fine-grained GPU event timing for performance analysis.
"""
# Create CUDA Events for precise timing
start_event = torch.cuda.Event(enable_timing=True)
vae_enc_event = torch.cuda.Event(enable_timing=True)
text_enc_event = torch.cuda.Event(enable_timing=True)
unet_loop_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
# -------------------------------------------------------------
# PHASE 1: Image Preprocessing & VAE Encode
# -------------------------------------------------------------
raw_img = Image.open(input_image_path).convert("RGB").resize((512, 512))
img_tensor = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.unet.dtype)
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
with torch.no_grad():
init_latents = pipe.vae.encode(img_tensor).latent_dist.sample(generator=generator)
init_latents = init_latents * pipe.vae.config.scaling_factor
vae_enc_event.record()
# -------------------------------------------------------------
# PHASE 2: Tokenize & Text Encoder Embeddings
# -------------------------------------------------------------
text_prompts = [id_class_prompt, ood_target_prompt, ""]
text_inputs = pipe.tokenizer(
text_prompts,
padding="max_length",
max_length=pipe.tokenizer.model_max_length,
return_tensors="pt"
)
with torch.no_grad():
embeddings = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
c_id = embeddings[0].unsqueeze(0)
c_ood = embeddings[1].unsqueeze(0)
c_uncond = embeddings[2].unsqueeze(0)
text_enc_event.record()
# -------------------------------------------------------------
# PHASE 3: Forward Noise Injection & Reverse Denoising Loop
# -------------------------------------------------------------
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
start_step_idx = 50 - stop_timestep
t_start = timesteps[start_step_idx]
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=init_latents.dtype)
z_t = pipe.scheduler.add_noise(init_latents, noise, t_start)
# Accumulators for inner loop breakdown
total_unet_time_ms = 0.0
total_math_time_ms = 0.0
u_start = torch.cuda.Event(enable_timing=True)
u_mid = torch.cuda.Event(enable_timing=True)
u_end = torch.cuda.Event(enable_timing=True)
with torch.no_grad():
for i in range(start_step_idx, len(timesteps)):
t = timesteps[i]
u_start.record()
# 3 Separate Sequential UNet forward calls
eps_uncond = pipe.unet(z_t, t, encoder_hidden_states=c_uncond).sample
eps_id_cond = pipe.unet(z_t, t, encoder_hidden_states=c_id).sample
eps_ood_cond = pipe.unet(z_t, t, encoder_hidden_states=c_ood).sample
u_mid.record()
# Vector math, dynamic quantile mask calculation, guidance assembly
psi_id = eps_id_cond - eps_uncond
psi_ood = eps_ood_cond - eps_uncond
mag_id = torch.abs(psi_id).mean(dim=1, keepdim=True)
mag_ood = torch.abs(psi_ood).mean(dim=1, keepdim=True)
if adaptive_mask:
max_val = mag_id.max()
active_pixels = (mag_id >= 0.15 * max_val).float()
coverage = active_pixels.mean().item()
t_percentile = np.clip(coverage * 1.2, 0.08, 0.40)
else:
t_percentile = threshold_percentile
eta_s = torch.quantile(mag_id.float().flatten(), 1.0 - t_percentile).to(mag_id.dtype)
m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
eta_n = torch.quantile(mag_id.float().flatten(), threshold_percentile).to(mag_id.dtype)
m_n_id = (mag_id < eta_n).to(psi_id.dtype)
if adaptive_mask:
max_ood = mag_ood.max()
active_ood = (mag_ood >= 0.15 * max_ood).float()
coverage_ood = active_ood.mean().item()
t_percentile_ood = np.clip(coverage_ood * 1.2, 0.08, 0.40)
else:
t_percentile_ood = threshold_percentile
eta_ood = torch.quantile(mag_ood.float().flatten(), 1.0 - t_percentile_ood).to(mag_ood.dtype)
m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
delta_id = -m_s_id * psi_id
delta_ood = m_s_ood * (1.0 - m_n_id) * psi_ood
delta_n = m_n_id * psi_id
eps_guided = eps_uncond + guidance_scale * (delta_id + delta_ood + delta_n)
z_t = pipe.scheduler.step(eps_guided, t, z_t).prev_sample
u_end.record()
torch.cuda.synchronize()
total_unet_time_ms += u_start.elapsed_time(u_mid)
total_math_time_ms += u_mid.elapsed_time(u_end)
unet_loop_event.record()
# -------------------------------------------------------------
# PHASE 4: VAE Decode Latents to Image
# -------------------------------------------------------------
with torch.no_grad():
decoded = pipe.vae.decode(z_t / pipe.vae.config.scaling_factor).sample
decoded = (decoded / 2 + 0.5).clamp(0, 1)
decoded = decoded.cpu().permute(0, 2, 3, 1).float().numpy()[0]
out_img = Image.fromarray((decoded * 255).astype(np.uint8)).resize(raw_img.size)
end_event.record()
torch.cuda.synchronize()
# Compute phase timings in milliseconds
metrics = {
"vae_encode_ms": start_event.elapsed_time(vae_enc_event),
"text_encode_ms": vae_enc_event.elapsed_time(text_enc_event),
"unet_inferences_ms": total_unet_time_ms,
"mask_and_scheduler_math_ms": total_math_time_ms,
"vae_decode_ms": unet_loop_event.elapsed_time(end_event),
"total_warm_start_ms": start_event.elapsed_time(end_event),
"steps_count": len(timesteps) - start_step_idx
}
return out_img, metrics
In [4]:
# -------------------------------------------------------------
# 1. Define Test Data & Prompts
# -------------------------------------------------------------
test_img_path = "/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG"
id_prompt = "a photo of a great white shark"
ood_prompt = "a photo of a harvester"
# -------------------------------------------------------------
# 2. Run 1 Untimed Warmup Iteration
# -------------------------------------------------------------
print("Running 1 Untimed Warmup Iteration...")
_, _ = profiled_generate_sona_outlier(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
print("Warmup complete\n")
# -------------------------------------------------------------
# 3. Run N Steady-State Benchmark Iterations
# -------------------------------------------------------------
NUM_RUNS = 5
print(f"Step 2: Running {NUM_RUNS} Steady-State Benchmark Iterations...")
metrics_list = []
for run in range(NUM_RUNS):
img, m = profiled_generate_sona_outlier(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
metrics_list.append(m)
print(f" Run {run+1}/{NUM_RUNS}: {m['total_warm_start_ms']:.2f} ms ({m['total_warm_start_ms']/1000:.3f} s)")
# -------------------------------------------------------------
# 4. Compute Averages & Display Benchmark Summary Table
# -------------------------------------------------------------
avg_metrics = {k: np.mean([m[k] for m in metrics_list]) for k in metrics_list[0].keys()}
total_ms = avg_metrics['total_warm_start_ms']
print("\n" + "="*70)
print(f"BASELINE SONA PROFILE BREAKDOWN (Averaged over {NUM_RUNS} runs)")
print("="*70)
print(f"1. VAE Encode (Image -> Latent) : {avg_metrics['vae_encode_ms']:8.2f} ms ({avg_metrics['vae_encode_ms']/total_ms*100:5.1f}%)")
print(f"2. Text Encoder (CLIP Embeddings) : {avg_metrics['text_encode_ms']:8.2f} ms ({avg_metrics['text_encode_ms']/total_ms*100:5.1f}%)")
print(f"3. UNet Inferences (3x per step) : {avg_metrics['unet_inferences_ms']:8.2f} ms ({avg_metrics['unet_inferences_ms']/total_ms*100:5.1f}%)")
print(f"4. Mask Math & Scheduler Steps : {avg_metrics['mask_and_scheduler_math_ms']:8.2f} ms ({avg_metrics['mask_and_scheduler_math_ms']/total_ms*100:5.1f}%)")
print(f"5. VAE Decode (Latent -> Image) : {avg_metrics['vae_decode_ms']:8.2f} ms ({avg_metrics['vae_decode_ms']/total_ms*100:5.1f}%)")
print("-" * 70)
print(f"⏱TOTAL WARM-START LATENCY PER IMAGE : {total_ms:8.2f} ms ({total_ms/1000:.3f} s)")
print("="*70)
Out[4]:
Running 1 Untimed Warmup Iteration... Warmup complete Step 2: Running 5 Steady-State Benchmark Iterations... Run 1/5: 1694.65 ms (1.695 s) Run 2/5: 1694.00 ms (1.694 s) Run 3/5: 1692.12 ms (1.692 s) Run 4/5: 1689.15 ms (1.689 s) Run 5/5: 1687.05 ms (1.687 s) ====================================================================== BASELINE SONA PROFILE BREAKDOWN (Averaged over 5 runs) ====================================================================== 1. VAE Encode (Image -> Latent) : 24.78 ms ( 1.5%) 2. Text Encoder (CLIP Embeddings) : 6.29 ms ( 0.4%) 3. UNet Inferences (3x per step) : 1600.08 ms ( 94.6%) 4. Mask Math & Scheduler Steps : 23.93 ms ( 1.4%) 5. VAE Decode (Latent -> Image) : 34.80 ms ( 2.1%) ---------------------------------------------------------------------- ⏱TOTAL WARM-START LATENCY PER IMAGE : 1691.39 ms (1.691 s) ======================================================================
In [5]:
def profiled_generate_sona_outlier_optimized1(
pipe,
input_image_path,
id_class_prompt,
ood_target_prompt,
stop_timestep=20,
guidance_scale=15.0,
threshold_percentile=0.10,
adaptive_mask=False,
device="cuda"
):
"""
Executes baseline SONA with fine-grained GPU event timing for performance analysis.
"""
# Create CUDA Events for precise timing
start_event = torch.cuda.Event(enable_timing=True)
vae_enc_event = torch.cuda.Event(enable_timing=True)
text_enc_event = torch.cuda.Event(enable_timing=True)
unet_loop_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
# -------------------------------------------------------------
# PHASE 1: Image Preprocessing & VAE Encode
# -------------------------------------------------------------
raw_img = Image.open(input_image_path).convert("RGB").resize((512, 512))
img_tensor = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.unet.dtype)
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
with torch.no_grad():
init_latents = pipe.vae.encode(img_tensor).latent_dist.sample(generator=generator)
init_latents = init_latents * pipe.vae.config.scaling_factor
vae_enc_event.record()
# -------------------------------------------------------------
# PHASE 2: Tokenize & Text Encoder Embeddings
# -------------------------------------------------------------
text_prompts = [id_class_prompt, ood_target_prompt, ""]
text_inputs = pipe.tokenizer(
text_prompts,
padding="max_length",
max_length=pipe.tokenizer.model_max_length,
return_tensors="pt"
)
with torch.no_grad():
embeddings = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
c_id = embeddings[0].unsqueeze(0)
c_ood = embeddings[1].unsqueeze(0)
c_uncond = embeddings[2].unsqueeze(0)
text_enc_event.record()
# -------------------------------------------------------------
# PHASE 3: Forward Noise Injection & Reverse Denoising Loop
# -------------------------------------------------------------
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
start_step_idx = 50 - stop_timestep
t_start = timesteps[start_step_idx]
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=init_latents.dtype)
z_t = pipe.scheduler.add_noise(init_latents, noise, t_start)
# z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
c_batched = torch.cat([c_uncond, c_id, c_ood], dim=0)
# Accumulators for inner loop breakdown
total_unet_time_ms = 0.0
total_math_time_ms = 0.0
u_start = torch.cuda.Event(enable_timing=True)
u_mid = torch.cuda.Event(enable_timing=True)
u_end = torch.cuda.Event(enable_timing=True)
with torch.no_grad():
for i in range(start_step_idx, len(timesteps)):
t = timesteps[i]
u_start.record()
# 3 Separate Sequential UNet forward calls
# eps_uncond = pipe.unet(z_t, t, encoder_hidden_states=c_uncond).sample
# eps_id_cond = pipe.unet(z_t, t, encoder_hidden_states=c_id).sample
# eps_ood_cond = pipe.unet(z_t, t, encoder_hidden_states=c_ood).sample
z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
eps_batched = pipe.unet(z_t_batched, t, encoder_hidden_states=c_batched).sample
eps_uncond, eps_id_cond, eps_ood_cond = eps_batched.chunk(3, dim=0)
u_mid.record()
# Vector math, dynamic quantile mask calculation, guidance assembly
psi_id = eps_id_cond - eps_uncond
psi_ood = eps_ood_cond - eps_uncond
mag_id = torch.abs(psi_id).mean(dim=1, keepdim=True)
mag_ood = torch.abs(psi_ood).mean(dim=1, keepdim=True)
if adaptive_mask:
max_val = mag_id.max()
active_pixels = (mag_id >= 0.15 * max_val).float()
coverage = active_pixels.mean().item()
t_percentile = np.clip(coverage * 1.2, 0.08, 0.40)
else:
t_percentile = threshold_percentile
eta_s = torch.quantile(mag_id.float().flatten(), 1.0 - t_percentile).to(mag_id.dtype)
m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
eta_n = torch.quantile(mag_id.float().flatten(), threshold_percentile).to(mag_id.dtype)
m_n_id = (mag_id < eta_n).to(psi_id.dtype)
if adaptive_mask:
max_ood = mag_ood.max()
active_ood = (mag_ood >= 0.15 * max_ood).float()
coverage_ood = active_ood.mean().item()
t_percentile_ood = np.clip(coverage_ood * 1.2, 0.08, 0.40)
else:
t_percentile_ood = threshold_percentile
eta_ood = torch.quantile(mag_ood.float().flatten(), 1.0 - t_percentile_ood).to(mag_ood.dtype)
m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
delta_id = -m_s_id * psi_id
delta_ood = m_s_ood * (1.0 - m_n_id) * psi_ood
delta_n = m_n_id * psi_id
eps_guided = eps_uncond + guidance_scale * (delta_id + delta_ood + delta_n)
z_t = pipe.scheduler.step(eps_guided, t, z_t).prev_sample
u_end.record()
torch.cuda.synchronize()
total_unet_time_ms += u_start.elapsed_time(u_mid)
total_math_time_ms += u_mid.elapsed_time(u_end)
unet_loop_event.record()
# -------------------------------------------------------------
# PHASE 4: VAE Decode Latents to Image
# -------------------------------------------------------------
with torch.no_grad():
decoded = pipe.vae.decode(z_t / pipe.vae.config.scaling_factor).sample
decoded = (decoded / 2 + 0.5).clamp(0, 1)
decoded = decoded.cpu().permute(0, 2, 3, 1).float().numpy()[0]
out_img = Image.fromarray((decoded * 255).astype(np.uint8)).resize(raw_img.size)
end_event.record()
torch.cuda.synchronize()
# Compute phase timings in milliseconds
metrics = {
"vae_encode_ms": start_event.elapsed_time(vae_enc_event),
"text_encode_ms": vae_enc_event.elapsed_time(text_enc_event),
"unet_inferences_ms": total_unet_time_ms,
"mask_and_scheduler_math_ms": total_math_time_ms,
"vae_decode_ms": unet_loop_event.elapsed_time(end_event),
"total_warm_start_ms": start_event.elapsed_time(end_event),
"steps_count": len(timesteps) - start_step_idx
}
return out_img, metrics
In [6]:
# -------------------------------------------------------------
# 1. Define Test Data & Prompts
# -------------------------------------------------------------
test_img_path = "/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG"
id_prompt = "a photo of a great white shark"
ood_prompt = "a photo of a harvester"
# -------------------------------------------------------------
# 2. Run 1 Untimed Warmup Iteration
# -------------------------------------------------------------
print("Running 1 Untimed Warmup Iteration...")
_, _ = profiled_generate_sona_outlier_optimized1(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
print("Warmup complete\n")
# -------------------------------------------------------------
# 3. Run N Steady-State Benchmark Iterations
# -------------------------------------------------------------
NUM_RUNS = 5
print(f"Step 2: Running {NUM_RUNS} Steady-State Benchmark Iterations...")
batched_metrics_list = []
for run in range(NUM_RUNS):
img_opt1, m_opt1 = profiled_generate_sona_outlier_optimized1(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
batched_metrics_list.append(m_opt1)
print(f" Run {run+1}/{NUM_RUNS}: {m_opt1['total_warm_start_ms']:.2f} ms ({m_opt1['total_warm_start_ms']/1000:.3f} s)")
# -------------------------------------------------------------
# 4. Compute Averages & Display Benchmark Summary Table
# -------------------------------------------------------------
avg_opt1 = {k: np.mean([m[k] for m in batched_metrics_list]) for k in batched_metrics_list[0].keys()}
total_ms = avg_opt1['total_warm_start_ms']
print("\n" + "="*70)
print(f"BASELINE SONA PROFILE BREAKDOWN (Averaged over {NUM_RUNS} runs)")
print("="*70)
print(f"1. VAE Encode (Image -> Latent) : {avg_opt1['vae_encode_ms']:8.2f} ms ({avg_opt1['vae_encode_ms']/total_ms*100:5.1f}%)")
print(f"2. Text Encoder (CLIP Embeddings) : {avg_opt1['text_encode_ms']:8.2f} ms ({avg_opt1['text_encode_ms']/total_ms*100:5.1f}%)")
print(f"3. UNet Inferences (3x per step) : {avg_opt1['unet_inferences_ms']:8.2f} ms ({avg_opt1['unet_inferences_ms']/total_ms*100:5.1f}%)")
print(f"4. Mask Math & Scheduler Steps : {avg_opt1['mask_and_scheduler_math_ms']:8.2f} ms ({avg_opt1['mask_and_scheduler_math_ms']/total_ms*100:5.1f}%)")
print(f"5. VAE Decode (Latent -> Image) : {avg_opt1['vae_decode_ms']:8.2f} ms ({avg_opt1['vae_decode_ms']/total_ms*100:5.1f}%)")
print("-" * 70)
print(f"⏱TOTAL WARM-START LATENCY PER IMAGE : {total_ms:8.2f} ms ({total_ms/1000:.3f} s)")
print("="*70)
Out[6]:
Running 1 Untimed Warmup Iteration... Warmup complete Step 2: Running 5 Steady-State Benchmark Iterations... Run 1/5: 685.48 ms (0.685 s) Run 2/5: 685.82 ms (0.686 s) Run 3/5: 685.93 ms (0.686 s) Run 4/5: 685.62 ms (0.686 s) Run 5/5: 685.93 ms (0.686 s) ====================================================================== BASELINE SONA PROFILE BREAKDOWN (Averaged over 5 runs) ====================================================================== 1. VAE Encode (Image -> Latent) : 24.85 ms ( 3.6%) 2. Text Encoder (CLIP Embeddings) : 6.17 ms ( 0.9%) 3. UNet Inferences (3x per step) : 611.17 ms ( 89.1%) 4. Mask Math & Scheduler Steps : 6.61 ms ( 1.0%) 5. VAE Decode (Latent -> Image) : 35.71 ms ( 5.2%) ---------------------------------------------------------------------- ⏱TOTAL WARM-START LATENCY PER IMAGE : 685.75 ms (0.686 s) ======================================================================
In [7]:
# Compare against Baseline (avg_metrics from Cell 3)
baseline_total = avg_metrics['total_warm_start_ms']
opt1_total = avg_opt1['total_warm_start_ms']
speedup = baseline_total / opt1_total
time_saved_ms = baseline_total - opt1_total
print("\n" + "="*75)
print(f"COMPARISON: BASELINE vs OPTIMIZATION 1 (Batched UNet Pass)")
print("="*75)
print(f"Metric │ Baseline │ Opt 1 (Batched) │ Reduction")
print("-" * 75)
print(f"UNet Inferences Latency │ {avg_metrics['unet_inferences_ms']:8.2f} ms │ {avg_opt1['unet_inferences_ms']:8.2f} ms │ {(1 - avg_opt1['unet_inferences_ms']/avg_metrics['unet_inferences_ms'])*100:5.1f}% faster")
print(f"Total Warm-Start Latency │ {baseline_total:8.2f} ms │ {opt1_total:8.2f} ms │ {(1 - opt1_total/baseline_total)*100:5.1f}% faster")
print("-" * 75)
print(f"⚡ OVERALL SPEEDUP FACTOR │ 1.00x │ {speedup:5.2f}x │ Time Saved: {time_saved_ms:.2f} ms/img")
print("="*75) Out[7]:
=========================================================================== COMPARISON: BASELINE vs OPTIMIZATION 1 (Batched UNet Pass) =========================================================================== Metric │ Baseline │ Opt 1 (Batched) │ Reduction --------------------------------------------------------------------------- UNet Inferences Latency │ 1600.08 ms │ 611.17 ms │ 61.8% faster Total Warm-Start Latency │ 1691.39 ms │ 685.75 ms │ 59.5% faster --------------------------------------------------------------------------- ⚡ OVERALL SPEEDUP FACTOR │ 1.00x │ 2.47x │ Time Saved: 1005.64 ms/img ===========================================================================
In [8]:
# Quality Check
baseline_arr = np.array(img).astype(np.float32)
opt1_arr = np.array(img_opt1).astype(np.float32)
abs_diff = np.abs(baseline_arr - opt1_arr)
max_pixel_diff = np.max(abs_diff)
mean_pixel_diff = np.mean(abs_diff)
mse = np.mean((baseline_arr - opt1_arr) ** 2)
if mse == 0:
psnr = float('inf')
else:
psnr = 20 * np.log10(255.0 / np.sqrt(mse))
print("="*65)
print("QUANTITATIVE IMAGE EQUIVALENCE VERIFICATION")
print("="*65)
print(f"Max Pixel Difference (0-255 scale) : {max_pixel_diff:.2f}")
print(f"Mean Pixel Difference (0-255 scale) : {mean_pixel_diff:.4f}")
print(f"Mean Squared Error (MSE) : {mse:.6f}")
print(f"PSNR (Signal-to-Noise Ratio) : {psnr:.2f} dB")
print("-" * 65) Out[8]:
================================================================= QUANTITATIVE IMAGE EQUIVALENCE VERIFICATION ================================================================= Max Pixel Difference (0-255 scale) : 44.00 Mean Pixel Difference (0-255 scale) : 1.1501 Mean Squared Error (MSE) : 3.970462 PSNR (Signal-to-Noise Ratio) : 42.14 dB -----------------------------------------------------------------
In [88]:
import matplotlib.pyplot as plt
fig, axes = plt.subplots(1, 2, figsize=(10, 5))
# Plot 1: Original Image
orig_img = Image.open(test_img_path).convert("RGB")
axes[0].imshow(orig_img)
axes[0].set_title("Original Input (A photo of a great shark)", fontsize=12)
axes[0].axis("off")
# Plot 2: Baseline SONA Output
axes[1].imshow(img)
axes[1].set_title(f"OOD (A photo of a harvester)", fontsize=12)
axes[1].axis("off")
plt.tight_layout()
plt.savefig('sona_image_comparisons.png', dpi=300, bbox_inches='tight')
plt.show() Out[88]:
In [10]:
prompt_cache = {}
def get_text_embedding(pipe, prompt_str, device="cuda"):
"""
Retrieves text embedding from cache if present; otherwise encodes and caches it
"""
if prompt_str not in prompt_cache:
text_inputs = pipe.tokenizer(
[prompt_str],
padding="max_length",
max_length=pipe.tokenizer.model_max_length,
return_tensors="pt"
)
with torch.no_grad():
emb = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
prompt_cache[prompt_str] = emb
return prompt_cache[prompt_str]
"""
Optimization 1 (batched UNET Pass) + optimization 2 (text embedding cache)
"""
def profiled_generate_sona_outlier_optimized2(
pipe,
input_image_path,
id_class_prompt,
ood_target_prompt,
stop_timestep=20,
guidance_scale=15.0,
threshold_percentile=0.10,
adaptive_mask=False,
device="cuda"
):
"""
Executes baseline SONA with fine-grained GPU event timing for performance analysis.
"""
# Create CUDA Events for precise timing
start_event = torch.cuda.Event(enable_timing=True)
vae_enc_event = torch.cuda.Event(enable_timing=True)
text_enc_event = torch.cuda.Event(enable_timing=True)
unet_loop_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
# -------------------------------------------------------------
# PHASE 1: Image Preprocessing & VAE Encode
# -------------------------------------------------------------
raw_img = Image.open(input_image_path).convert("RGB").resize((512, 512))
img_tensor = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.unet.dtype)
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
with torch.no_grad():
init_latents = pipe.vae.encode(img_tensor).latent_dist.sample(generator=generator)
init_latents = init_latents * pipe.vae.config.scaling_factor
vae_enc_event.record()
# -------------------------------------------------------------
# PHASE 2: Tokenize & Text Encoder Embeddings
# -------------------------------------------------------------
# text_prompts = [id_class_prompt, ood_target_prompt, ""]
# text_inputs = pipe.tokenizer(
# text_prompts,
# padding="max_length",
# max_length=pipe.tokenizer.model_max_length,
# return_tensors="pt"
# )
# with torch.no_grad():
# embeddings = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
c_id = get_text_embedding(pipe, id_class_prompt, device=device)
c_ood = get_text_embedding(pipe, ood_target_prompt, device=device)
c_uncond = get_text_embedding(pipe, "", device=device)
text_enc_event.record()
# -------------------------------------------------------------
# PHASE 3: Forward Noise Injection & Reverse Denoising Loop
# -------------------------------------------------------------
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
start_step_idx = 50 - stop_timestep
t_start = timesteps[start_step_idx]
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=init_latents.dtype)
z_t = pipe.scheduler.add_noise(init_latents, noise, t_start)
# z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
c_batched = torch.cat([c_uncond, c_id, c_ood], dim=0)
# Accumulators for inner loop breakdown
total_unet_time_ms = 0.0
total_math_time_ms = 0.0
u_start = torch.cuda.Event(enable_timing=True)
u_mid = torch.cuda.Event(enable_timing=True)
u_end = torch.cuda.Event(enable_timing=True)
with torch.no_grad():
for i in range(start_step_idx, len(timesteps)):
t = timesteps[i]
u_start.record()
# 3 Separate Sequential UNet forward calls
# eps_uncond = pipe.unet(z_t, t, encoder_hidden_states=c_uncond).sample
# eps_id_cond = pipe.unet(z_t, t, encoder_hidden_states=c_id).sample
# eps_ood_cond = pipe.unet(z_t, t, encoder_hidden_states=c_ood).sample
z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
eps_batched = pipe.unet(z_t_batched, t, encoder_hidden_states=c_batched).sample
eps_uncond, eps_id_cond, eps_ood_cond = eps_batched.chunk(3, dim=0)
u_mid.record()
# Vector math, dynamic quantile mask calculation, guidance assembly
psi_id = eps_id_cond - eps_uncond
psi_ood = eps_ood_cond - eps_uncond
mag_id = torch.abs(psi_id).mean(dim=1, keepdim=True)
mag_ood = torch.abs(psi_ood).mean(dim=1, keepdim=True)
if adaptive_mask:
max_val = mag_id.max()
active_pixels = (mag_id >= 0.15 * max_val).float()
coverage = active_pixels.mean().item()
t_percentile = np.clip(coverage * 1.2, 0.08, 0.40)
else:
t_percentile = threshold_percentile
eta_s = torch.quantile(mag_id.float().flatten(), 1.0 - t_percentile).to(mag_id.dtype)
m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
eta_n = torch.quantile(mag_id.float().flatten(), threshold_percentile).to(mag_id.dtype)
m_n_id = (mag_id < eta_n).to(psi_id.dtype)
if adaptive_mask:
max_ood = mag_ood.max()
active_ood = (mag_ood >= 0.15 * max_ood).float()
coverage_ood = active_ood.mean().item()
t_percentile_ood = np.clip(coverage_ood * 1.2, 0.08, 0.40)
else:
t_percentile_ood = threshold_percentile
eta_ood = torch.quantile(mag_ood.float().flatten(), 1.0 - t_percentile_ood).to(mag_ood.dtype)
m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
delta_id = -m_s_id * psi_id
delta_ood = m_s_ood * (1.0 - m_n_id) * psi_ood
delta_n = m_n_id * psi_id
eps_guided = eps_uncond + guidance_scale * (delta_id + delta_ood + delta_n)
z_t = pipe.scheduler.step(eps_guided, t, z_t).prev_sample
u_end.record()
torch.cuda.synchronize()
total_unet_time_ms += u_start.elapsed_time(u_mid)
total_math_time_ms += u_mid.elapsed_time(u_end)
unet_loop_event.record()
# -------------------------------------------------------------
# PHASE 4: VAE Decode Latents to Image
# -------------------------------------------------------------
with torch.no_grad():
decoded = pipe.vae.decode(z_t / pipe.vae.config.scaling_factor).sample
decoded = (decoded / 2 + 0.5).clamp(0, 1)
decoded = decoded.cpu().permute(0, 2, 3, 1).float().numpy()[0]
out_img = Image.fromarray((decoded * 255).astype(np.uint8)).resize(raw_img.size)
end_event.record()
torch.cuda.synchronize()
# Compute phase timings in milliseconds
metrics = {
"vae_encode_ms": start_event.elapsed_time(vae_enc_event),
"text_encode_ms": vae_enc_event.elapsed_time(text_enc_event),
"unet_inferences_ms": total_unet_time_ms,
"mask_and_scheduler_math_ms": total_math_time_ms,
"vae_decode_ms": unet_loop_event.elapsed_time(end_event),
"total_warm_start_ms": start_event.elapsed_time(end_event),
"steps_count": len(timesteps) - start_step_idx
}
return out_img, metrics
In [11]:
prompt_cache.clear()
# -------------------------------------------------------------
# 1. Define Test Data & Prompts
# -------------------------------------------------------------
test_img_path = "/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG"
id_prompt = "a photo of a great white shark"
ood_prompt = "a photo of a harvester"
# -------------------------------------------------------------
# 2. Run 1 Untimed Warmup Iteration
# -------------------------------------------------------------
print("Running 1 Untimed Warmup Iteration...")
_, _ = profiled_generate_sona_outlier_optimized2(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
print("Warmup complete\n")
# -------------------------------------------------------------
# 3. Run N Steady-State Benchmark Iterations
# -------------------------------------------------------------
NUM_RUNS = 5
print(f"Step 2: Running {NUM_RUNS} Steady-State Benchmark Iterations...")
opt2_metrics_list = []
for run in range(NUM_RUNS):
img_opt2, m_opt2 = profiled_generate_sona_outlier_optimized2(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
opt2_metrics_list.append(m_opt2)
print(f" Run {run+1}/{NUM_RUNS}: {m_opt2['total_warm_start_ms']:.2f} ms ({m_opt2['total_warm_start_ms']/1000:.3f} s)")
# -------------------------------------------------------------
# 4. Compute Averages & Display Benchmark Summary Table
# -------------------------------------------------------------
avg_opt2 = {k: np.mean([m[k] for m in opt2_metrics_list]) for k in opt2_metrics_list[0].keys()}
total_ms = avg_opt2['total_warm_start_ms']
print("\n" + "="*70)
print(f"BASELINE SONA PROFILE BREAKDOWN (Averaged over {NUM_RUNS} runs)")
print("="*70)
print(f"1. VAE Encode (Image -> Latent) : {avg_opt2['vae_encode_ms']:8.2f} ms ({avg_opt2['vae_encode_ms']/total_ms*100:5.1f}%)")
print(f"2. Text Encoder (CLIP Embeddings) : {avg_opt2['text_encode_ms']:8.2f} ms ({avg_opt2['text_encode_ms']/total_ms*100:5.1f}%)")
print(f"3. UNet Inferences (3x per step) : {avg_opt2['unet_inferences_ms']:8.2f} ms ({avg_opt2['unet_inferences_ms']/total_ms*100:5.1f}%)")
print(f"4. Mask Math & Scheduler Steps : {avg_opt2['mask_and_scheduler_math_ms']:8.2f} ms ({avg_opt2['mask_and_scheduler_math_ms']/total_ms*100:5.1f}%)")
print(f"5. VAE Decode (Latent -> Image) : {avg_opt2['vae_decode_ms']:8.2f} ms ({avg_opt2['vae_decode_ms']/total_ms*100:5.1f}%)")
print("-" * 70)
print(f"⏱TOTAL WARM-START LATENCY PER IMAGE : {total_ms:8.2f} ms ({total_ms/1000:.3f} s)")
print("="*70)
Out[11]:
Running 1 Untimed Warmup Iteration...
Out[11]:
Warmup complete Step 2: Running 5 Steady-State Benchmark Iterations... Run 1/5: 684.25 ms (0.684 s) Run 2/5: 684.05 ms (0.684 s) Run 3/5: 684.29 ms (0.684 s) Run 4/5: 682.98 ms (0.683 s) Run 5/5: 683.51 ms (0.684 s) ====================================================================== BASELINE SONA PROFILE BREAKDOWN (Averaged over 5 runs) ====================================================================== 1. VAE Encode (Image -> Latent) : 25.20 ms ( 3.7%) 2. Text Encoder (CLIP Embeddings) : 0.00 ms ( 0.0%) 3. UNet Inferences (3x per step) : 614.98 ms ( 89.9%) 4. Mask Math & Scheduler Steps : 6.62 ms ( 1.0%) 5. VAE Decode (Latent -> Image) : 35.88 ms ( 5.2%) ---------------------------------------------------------------------- ⏱TOTAL WARM-START LATENCY PER IMAGE : 683.82 ms (0.684 s) ======================================================================
In [12]:
# Benchmark comparison table across Baseline, Opt 1, and Opt 2
print("\n" + "="*80)
print(f"3-WAY COMPARISON: BASELINE vs OPT 1 (BATCHED) vs OPT 2 (BATCHED + CACHED)")
print("="*80)
print(f"Phase Latency │ Baseline │ Opt 1 (Batched) │ Opt 2 (Cached)")
print("-" * 80)
print(f"Text Encoder Latency │ {avg_metrics['text_encode_ms']:8.2f} ms │ {avg_opt1['text_encode_ms']:13.2f} ms │ {avg_opt2['text_encode_ms']:12.2f} ms")
print(f"UNet Inferences Latency │ {avg_metrics['unet_inferences_ms']:8.2f} ms │ {avg_opt1['unet_inferences_ms']:13.2f} ms │ {avg_opt2['unet_inferences_ms']:12.2f} ms")
print(f"Total Warm-Start Latency │ {baseline_total:8.2f} ms │ {opt1_total:13.2f} ms │ {avg_opt2['total_warm_start_ms']:12.2f} ms")
print("-" * 80)
speedup_opt2 = baseline_total / avg_opt2['total_warm_start_ms']
print(f"OVERALL SPEEDUP FACTOR │ 1.00x │ {speedup:13.2f}x │ {speedup_opt2:12.2f}x")
print("="*80) Out[12]:
================================================================================ 3-WAY COMPARISON: BASELINE vs OPT 1 (BATCHED) vs OPT 2 (BATCHED + CACHED) ================================================================================ Phase Latency │ Baseline │ Opt 1 (Batched) │ Opt 2 (Cached) -------------------------------------------------------------------------------- Text Encoder Latency │ 6.29 ms │ 6.17 ms │ 0.00 ms UNet Inferences Latency │ 1600.08 ms │ 611.17 ms │ 614.98 ms Total Warm-Start Latency │ 1691.39 ms │ 685.75 ms │ 683.82 ms -------------------------------------------------------------------------------- OVERALL SPEEDUP FACTOR │ 1.00x │ 2.47x │ 2.47x ================================================================================
In [13]:
"""
Optimization 1 (batched UNET Pass) + optimization 2 (text embedding cache) + optimization 3 (inference)
"""
def profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path,
id_class_prompt,
ood_target_prompt,
stop_timestep=20,
guidance_scale=15.0,
threshold_percentile=0.10,
adaptive_mask=False,
device="cuda"
):
"""
Executes baseline SONA with fine-grained GPU event timing for performance analysis.
"""
# Create CUDA Events for precise timing
start_event = torch.cuda.Event(enable_timing=True)
vae_enc_event = torch.cuda.Event(enable_timing=True)
text_enc_event = torch.cuda.Event(enable_timing=True)
unet_loop_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
# -------------------------------------------------------------
# PHASE 1: Image Preprocessing & VAE Encode
# -------------------------------------------------------------
raw_img = Image.open(input_image_path).convert("RGB").resize((512, 512))
img_tensor = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
# img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.unet.dtype)
img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.text_encoder.dtype)
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
# opt 3
with torch.inference_mode():
init_latents = pipe.vae.encode(img_tensor).latent_dist.sample(generator=generator)
init_latents = init_latents * pipe.vae.config.scaling_factor
vae_enc_event.record()
# -------------------------------------------------------------
# PHASE 2: Tokenize & Text Encoder Embeddings
# -------------------------------------------------------------
# text_prompts = [id_class_prompt, ood_target_prompt, ""]
# text_inputs = pipe.tokenizer(
# text_prompts,
# padding="max_length",
# max_length=pipe.tokenizer.model_max_length,
# return_tensors="pt"
# )
# with torch.no_grad():
# embeddings = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
c_id = get_text_embedding(pipe, id_class_prompt, device=device)
c_ood = get_text_embedding(pipe, ood_target_prompt, device=device)
c_uncond = get_text_embedding(pipe, "", device=device)
text_enc_event.record()
# -------------------------------------------------------------
# PHASE 3: Forward Noise Injection & Reverse Denoising Loop
# -------------------------------------------------------------
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
start_step_idx = 50 - stop_timestep
t_start = timesteps[start_step_idx]
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(42)
noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=init_latents.dtype)
z_t = pipe.scheduler.add_noise(init_latents, noise, t_start)
# z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
c_batched = torch.cat([c_uncond, c_id, c_ood], dim=0)
# Accumulators for inner loop breakdown
total_unet_time_ms = 0.0
total_math_time_ms = 0.0
u_start = torch.cuda.Event(enable_timing=True)
u_mid = torch.cuda.Event(enable_timing=True)
u_end = torch.cuda.Event(enable_timing=True)
with torch.inference_mode():
for i in range(start_step_idx, len(timesteps)):
t = timesteps[i]
u_start.record()
# 3 Separate Sequential UNet forward calls
# eps_uncond = pipe.unet(z_t, t, encoder_hidden_states=c_uncond).sample
# eps_id_cond = pipe.unet(z_t, t, encoder_hidden_states=c_id).sample
# eps_ood_cond = pipe.unet(z_t, t, encoder_hidden_states=c_ood).sample
z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
eps_batched = pipe.unet(z_t_batched, t, encoder_hidden_states=c_batched).sample
eps_uncond, eps_id_cond, eps_ood_cond = eps_batched.chunk(3, dim=0)
u_mid.record()
# Vector math, dynamic quantile mask calculation, guidance assembly
psi_id = eps_id_cond - eps_uncond
psi_ood = eps_ood_cond - eps_uncond
mag_id = torch.abs(psi_id).mean(dim=1, keepdim=True)
mag_ood = torch.abs(psi_ood).mean(dim=1, keepdim=True)
if adaptive_mask:
max_val = mag_id.max()
active_pixels = (mag_id >= 0.15 * max_val).float()
coverage = active_pixels.mean().item()
t_percentile = np.clip(coverage * 1.2, 0.08, 0.40)
else:
t_percentile = threshold_percentile
eta_s = torch.quantile(mag_id.float().flatten(), 1.0 - t_percentile).to(mag_id.dtype)
m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
eta_n = torch.quantile(mag_id.float().flatten(), threshold_percentile).to(mag_id.dtype)
m_n_id = (mag_id < eta_n).to(psi_id.dtype)
if adaptive_mask:
max_ood = mag_ood.max()
active_ood = (mag_ood >= 0.15 * max_ood).float()
coverage_ood = active_ood.mean().item()
t_percentile_ood = np.clip(coverage_ood * 1.2, 0.08, 0.40)
else:
t_percentile_ood = threshold_percentile
eta_ood = torch.quantile(mag_ood.float().flatten(), 1.0 - t_percentile_ood).to(mag_ood.dtype)
m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
delta_id = -m_s_id * psi_id
delta_ood = m_s_ood * (1.0 - m_n_id) * psi_ood
delta_n = m_n_id * psi_id
eps_guided = eps_uncond + guidance_scale * (delta_id + delta_ood + delta_n)
z_t = pipe.scheduler.step(eps_guided, t, z_t).prev_sample
u_end.record()
torch.cuda.synchronize()
total_unet_time_ms += u_start.elapsed_time(u_mid)
total_math_time_ms += u_mid.elapsed_time(u_end)
unet_loop_event.record()
# -------------------------------------------------------------
# PHASE 4: VAE Decode Latents to Image
# -------------------------------------------------------------
with torch.no_grad():
decoded = pipe.vae.decode(z_t / pipe.vae.config.scaling_factor).sample
decoded = (decoded / 2 + 0.5).clamp(0, 1)
decoded = decoded.cpu().permute(0, 2, 3, 1).float().numpy()[0]
out_img = Image.fromarray((decoded * 255).astype(np.uint8)).resize(raw_img.size)
end_event.record()
torch.cuda.synchronize()
# Compute phase timings in milliseconds
metrics = {
"vae_encode_ms": start_event.elapsed_time(vae_enc_event),
"text_encode_ms": vae_enc_event.elapsed_time(text_enc_event),
"unet_inferences_ms": total_unet_time_ms,
"mask_and_scheduler_math_ms": total_math_time_ms,
"vae_decode_ms": unet_loop_event.elapsed_time(end_event),
"total_warm_start_ms": start_event.elapsed_time(end_event),
"steps_count": len(timesteps) - start_step_idx
}
return out_img, metrics
In [14]:
prompt_cache.clear()
# -------------------------------------------------------------
# 1. Define Test Data & Prompts
# -------------------------------------------------------------
test_img_path = "/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG"
id_prompt = "a photo of a great white shark"
ood_prompt = "a photo of a harvester"
# -------------------------------------------------------------
# 2. Run 1 Untimed Warmup Iteration
# -------------------------------------------------------------
print("Running 1 Untimed Warmup Iteration...")
_, _ = profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
print("Warmup complete\n")
# -------------------------------------------------------------
# 3. Run N Steady-State Benchmark Iterations
# -------------------------------------------------------------
NUM_RUNS = 5
print(f"Step 2: Running {NUM_RUNS} Steady-State Benchmark Iterations...")
opt3_metrics_list = []
for run in range(NUM_RUNS):
img_opt3, m_opt3 = profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
opt3_metrics_list.append(m_opt3)
print(f" Run {run+1}/{NUM_RUNS}: {m_opt3['total_warm_start_ms']:.2f} ms ({m_opt3['total_warm_start_ms']/1000:.3f} s)")
# -------------------------------------------------------------
# 4. Compute Averages & Display Benchmark Summary Table
# -------------------------------------------------------------
avg_opt3 = {k: np.mean([m[k] for m in opt3_metrics_list]) for k in opt3_metrics_list[0].keys()}
total_ms = avg_opt3['total_warm_start_ms']
print("\n" + "="*70)
print(f"BASELINE SONA PROFILE BREAKDOWN (Averaged over {NUM_RUNS} runs)")
print("="*70)
print(f"1. VAE Encode (Image -> Latent) : {avg_opt3['vae_encode_ms']:8.2f} ms ({avg_opt3['vae_encode_ms']/total_ms*100:5.1f}%)")
print(f"2. Text Encoder (CLIP Embeddings) : {avg_opt3['text_encode_ms']:8.2f} ms ({avg_opt3['text_encode_ms']/total_ms*100:5.1f}%)")
print(f"3. UNet Inferences (3x per step) : {avg_opt3['unet_inferences_ms']:8.2f} ms ({avg_opt3['unet_inferences_ms']/total_ms*100:5.1f}%)")
print(f"4. Mask Math & Scheduler Steps : {avg_opt3['mask_and_scheduler_math_ms']:8.2f} ms ({avg_opt3['mask_and_scheduler_math_ms']/total_ms*100:5.1f}%)")
print(f"5. VAE Decode (Latent -> Image) : {avg_opt3['vae_decode_ms']:8.2f} ms ({avg_opt3['vae_decode_ms']/total_ms*100:5.1f}%)")
print("-" * 70)
print(f"⏱TOTAL WARM-START LATENCY PER IMAGE : {total_ms:8.2f} ms ({total_ms/1000:.3f} s)")
print("="*70)
Out[14]:
Running 1 Untimed Warmup Iteration...
Out[14]:
Warmup complete Step 2: Running 5 Steady-State Benchmark Iterations... Run 1/5: 669.39 ms (0.669 s) Run 2/5: 669.23 ms (0.669 s) Run 3/5: 669.54 ms (0.670 s) Run 4/5: 670.27 ms (0.670 s) Run 5/5: 669.49 ms (0.669 s) ====================================================================== BASELINE SONA PROFILE BREAKDOWN (Averaged over 5 runs) ====================================================================== 1. VAE Encode (Image -> Latent) : 22.96 ms ( 3.4%) 2. Text Encoder (CLIP Embeddings) : 0.00 ms ( 0.0%) 3. UNet Inferences (3x per step) : 602.75 ms ( 90.0%) 4. Mask Math & Scheduler Steps : 6.72 ms ( 1.0%) 5. VAE Decode (Latent -> Image) : 35.95 ms ( 5.4%) ---------------------------------------------------------------------- ⏱TOTAL WARM-START LATENCY PER IMAGE : 669.58 ms (0.670 s) ======================================================================
In [15]:
print("\n" + "="*85)
print(f"4-WAY COMPARISON: BASELINE vs OPT 1 vs OPT 2 vs OPT 3 (INFERENCE MODE)")
print("="*85)
print(f"Phase Latency │ Baseline │ Opt 1 │ Opt 2 │ Opt 3 (Inference Mode)")
print("-" * 85)
print(f"UNet Inferences Latency │ {avg_metrics['unet_inferences_ms']:8.2f} ms │ {avg_opt1['unet_inferences_ms']:9.2f} ms │ {avg_opt2['unet_inferences_ms']:9.2f} ms │ {avg_opt3['unet_inferences_ms']:10.2f} ms")
print(f"Total Warm-Start Latency │ {baseline_total:8.2f} ms │ {opt1_total:9.2f} ms │ {avg_opt2['total_warm_start_ms']:9.2f} ms │ {avg_opt3['total_warm_start_ms']:10.2f} ms")
print("-" * 85)
speedup_opt3 = baseline_total / avg_opt3['total_warm_start_ms']
print(f"OVERALL SPEEDUP FACTOR │ 1.00x │ {speedup:9.2f}x │ {speedup_opt2:9.2f}x │ {speedup_opt3:10.2f}x")
print("="*85) Out[15]:
===================================================================================== 4-WAY COMPARISON: BASELINE vs OPT 1 vs OPT 2 vs OPT 3 (INFERENCE MODE) ===================================================================================== Phase Latency │ Baseline │ Opt 1 │ Opt 2 │ Opt 3 (Inference Mode) ------------------------------------------------------------------------------------- UNet Inferences Latency │ 1600.08 ms │ 611.17 ms │ 614.98 ms │ 602.75 ms Total Warm-Start Latency │ 1691.39 ms │ 685.75 ms │ 683.82 ms │ 669.58 ms ------------------------------------------------------------------------------------- OVERALL SPEEDUP FACTOR │ 1.00x │ 2.47x │ 2.47x │ 2.53x =====================================================================================
In [16]:
## Compilation
import torch
print("Compiling pipe.unet with PyTorch 2.0 TorchInductor (mode='reduce-overhead')...")
t_comp_start = time.perf_counter()
# pipe.unet = torch.compile(pipe.unet, mode="max-autotune", options={"triton.cudagraphs": True})
pipe.unet = torch.compile(pipe.unet,
options={
"triton.cudagraphs": True,
"max_autotune": True})
if device == "cuda":
torch.cuda.synchronize()
print(f"UNet marked for compilation. (Setup time: {time.perf_counter() - t_comp_start:.2f} s)") Out[16]:
Compiling pipe.unet with PyTorch 2.0 TorchInductor (mode='reduce-overhead')... UNet marked for compilation. (Setup time: 1.41 s)
In [17]:
prompt_cache.clear()
# -------------------------------------------------------------
# 1. Define Test Data & Prompts
# -------------------------------------------------------------
test_img_path = "/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG"
id_prompt = "a photo of a great white shark"
ood_prompt = "a photo of a harvester"
# -------------------------------------------------------------
# 2. Run 1 Untimed Warmup Iteration
# -------------------------------------------------------------
print("Running 1 Untimed Warmup Iteration...")
t_warm_start = time.perf_counter()
_, _ = profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
t_warm_end = time.perf_counter()
print(f"JIT Compilation & Warmup Complete! (Compilation Warmup took: {t_warm_end - t_warm_start:.2f} s)\n")
# -------------------------------------------------------------
# 3. Run N Steady-State Benchmark Iterations
# -------------------------------------------------------------
NUM_RUNS = 5
print(f"Step 2: Running {NUM_RUNS} Steady-State Benchmark Iterations...")
opt4_metrics_list = []
for run in range(NUM_RUNS):
img_opt4, m_opt4 = profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt=ood_prompt,
stop_timestep=23,
guidance_scale=10.0,
threshold_percentile=0.2
)
opt4_metrics_list.append(m_opt4)
print(f" Run {run+1}/{NUM_RUNS}: {m_opt4['total_warm_start_ms']:.2f} ms ({m_opt4['total_warm_start_ms']/1000:.3f} s)")
# -------------------------------------------------------------
# 4. Compute Averages & Display Benchmark Summary Table
# -------------------------------------------------------------
avg_opt4 = {k: np.mean([m[k] for m in opt4_metrics_list]) for k in opt4_metrics_list[0].keys()}
total_ms = avg_opt4['total_warm_start_ms']
print("\n" + "="*70)
print(f"BASELINE SONA PROFILE BREAKDOWN (Averaged over {NUM_RUNS} runs)")
print("="*70)
print(f"1. VAE Encode (Image -> Latent) : {avg_opt4['vae_encode_ms']:8.2f} ms ({avg_opt4['vae_encode_ms']/total_ms*100:5.1f}%)")
print(f"2. Text Encoder (CLIP Embeddings) : {avg_opt4['text_encode_ms']:8.2f} ms ({avg_opt4['text_encode_ms']/total_ms*100:5.1f}%)")
print(f"3. UNet Inferences (3x per step) : {avg_opt4['unet_inferences_ms']:8.2f} ms ({avg_opt4['unet_inferences_ms']/total_ms*100:5.1f}%)")
print(f"4. Mask Math & Scheduler Steps : {avg_opt4['mask_and_scheduler_math_ms']:8.2f} ms ({avg_opt4['mask_and_scheduler_math_ms']/total_ms*100:5.1f}%)")
print(f"5. VAE Decode (Latent -> Image) : {avg_opt4['vae_decode_ms']:8.2f} ms ({avg_opt4['vae_decode_ms']/total_ms*100:5.1f}%)")
print("-" * 70)
print(f"TOTAL WARM-START LATENCY PER IMAGE : {total_ms:8.2f} ms ({total_ms/1000:.3f} s)")
print("="*70)
Out[17]:
Running 1 Untimed Warmup Iteration...
Out[17]:
/home/hice1/yoh88/Out-of-distribution-Mukhopadhyay-gatech/kernel/lib/python3.10/site-packages/torch/_inductor/select_algorithm.py:4628: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly. To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()
current_out_size = out_base.storage().size()
E0801 16:59:19.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:19.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:19.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:21.347000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:21.347000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:21.347000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:24.214000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:24.214000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:24.214000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:25.070000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:25.070000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:25.070000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_311", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.0870399996638298, "best_triton_pos": 0}
AUTOTUNE addmm(12288x320, 12288x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_311 0.0870 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_307 0.0911 ms 95.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_314 0.0911 ms 95.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_317 0.0952 ms 91.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_310 0.1014 ms 85.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_309 0.1044 ms 83.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.1126 ms 77.3%
triton_mm_316 0.1126 ms 77.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_304 0.1147 ms 75.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_302 0.1178 ms 73.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 9.2044 seconds and 0.0003 seconds precompiling for 21 choices
E0801 16:59:27.684000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:27.684000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:27.684000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:29.282000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:29.282000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:29.282000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:32.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:32.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:32.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:32.515000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:32.515000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:32.515000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_623", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.03174399957060814, "best_triton_pos": 0}
AUTOTUNE addmm(12288x320, 12288x320, 320x320)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_623 0.0317 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_628 0.0327 ms 97.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_626 0.0328 ms 96.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_619 0.0338 ms 93.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_625 0.0338 ms 93.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_621 0.0348 ms 91.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_629 0.0348 ms 91.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_622 0.0358 ms 88.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_614 0.0399 ms 79.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_615 0.0430 ms 73.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.0955 seconds and 0.0002 seconds precompiling for 21 choices
E0801 16:59:35.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:35.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:35.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:36.873000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:36.873000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:36.873000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:39.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:39.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:39.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:39.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:39.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:39.481000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_938", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.05939200147986412, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x2560, 2560x640)
strides: [0, 1], [2560, 1], [1, 2560]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_938 0.0594 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0655 ms 90.6%
triton_mm_932 0.0666 ms 89.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_931 0.0788 ms 75.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_935 0.0799 ms 74.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_928 0.0932 ms 63.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_937 0.0952 ms 62.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_934 0.1024 ms 58.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_930 0.1085 ms 54.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_925 0.1106 ms 53.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 6.5795 seconds and 0.0002 seconds precompiling for 21 choices
E0801 16:59:42.215000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:42.215000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:42.215000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:44.027000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:44.027000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:44.027000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:46.617000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:46.617000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:46.617000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:47.148000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:47.148000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:47.148000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_1244", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.01945599913597107, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x640, 640x640)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_1244 0.0195 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1250 0.0195 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1240 0.0236 ms 82.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1243 0.0287 ms 67.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1247 0.0307 ms 63.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1236 0.0369 ms 52.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1249 0.0369 ms 52.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1234 0.0379 ms 51.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_1235 0.0379 ms 51.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1246 0.0379 ms 51.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.3399 seconds and 0.0002 seconds precompiling for 21 choices
E0801 16:59:49.906000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:49.906000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:49.906000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:51.301000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:51.301000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:51.301000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:53.492000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:53.492000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:53.492000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:53.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:53.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:53.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "addmm", "best_time": 0.06761600077152252, "best_triton_pos": 1, "best_triton_time": 0.0870399996638298, "best_triton_kernel": "triton_mm_1549", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8"}
AUTOTUNE addmm(768x1280, 768x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
addmm 0.0676 ms 100.0%
triton_mm_1549 0.0870 ms 77.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1553 0.0891 ms 75.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1559 0.1055 ms 64.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1546 0.1116 ms 60.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_1552 0.1116 ms 60.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1556 0.1126 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1555 0.1300 ms 52.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1551 0.1311 ms 51.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1544 0.1331 ms 50.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.4195 seconds and 0.0002 seconds precompiling for 21 choices
E0801 16:59:56.988000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:56.988000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:56.988000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 16:59:58.502000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 16:59:58.502000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 16:59:58.502000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:01.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:01.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:01.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_1861", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8", "best_time": 0.02457600086927414, "best_triton_pos": 0}
AUTOTUNE addmm(768x1280, 768x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_1861 0.0246 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
addmm 0.0266 ms 92.3%
triton_mm_1865 0.0266 ms 92.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1858 0.0338 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_1871 0.0338 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1856 0.0379 ms 64.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1855 0.0389 ms 63.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_1857 0.0400 ms 61.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1864 0.0430 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1868 0.0430 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7920 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:00:05.296000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:05.296000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:05.296000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:06.749000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:06.749000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:06.749000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:09.030000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:09.030000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:09.030000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:09.459000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:09.459000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:09.459000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2251", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.03484800085425377, "best_triton_pos": 0}
AUTOTUNE addmm(192x1280, 192x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_2251 0.0348 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
addmm 0.0440 ms 79.1%
triton_mm_2248 0.0655 ms 53.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2254 0.0779 ms 44.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_2249 0.0788 ms 44.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2250 0.0891 ms 39.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2258 0.0891 ms 39.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2257 0.1014 ms 34.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2261 0.1023 ms 34.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2264 0.1024 ms 34.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.1988 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 16, "num_triton_choices": 15, "best_kernel": "triton_convolution2d_5421", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.035840000957250595, "best_triton_pos": 0}
AUTOTUNE convolution(3x320x64x64, 4x320x3x3)
strides: [1310720, 1, 20480, 320], [2880, 1, 960, 320]
dtypes: torch.float16, torch.float16
triton_convolution2d_5421 0.0358 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5422 0.0369 ms 97.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5426 0.0379 ms 94.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
convolution 0.0399 ms 89.7%
triton_convolution2d_5427 0.0399 ms 89.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5420 0.0440 ms 81.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_5425 0.0502 ms 71.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5428 0.0584 ms 61.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5424 0.0594 ms 60.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5423 0.0645 ms 55.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 4.1525 seconds and 0.0002 seconds precompiling for 16 choices
E0801 17:00:17.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:17.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:17.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:18.519000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:18.519000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:18.519000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_33", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.006144000217318535, "best_triton_pos": 0}
AUTOTUNE addmm(3x1280, 3x320, 320x1280)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_33 0.0061 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_30 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_36 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_37 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_40 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_45 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_31 0.0082 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_32 0.0082 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_43 0.0091 ms 67.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
triton_mm_38 0.0092 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 4.1049 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:00:22.357000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:22.357000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:22.357000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:23.525000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:23.525000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:23.525000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_51", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.011264000087976456, "best_triton_pos": 0}
AUTOTUNE addmm(3x1280, 3x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_51 0.0113 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_55 0.0123 ms 91.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_63 0.0133 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_48 0.0143 ms 78.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_54 0.0154 ms 73.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_58 0.0154 ms 73.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0174 ms 64.7%
triton_mm_49 0.0205 ms 55.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_50 0.0205 ms 55.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_61 0.0236 ms 47.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 4.6724 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "triton_convolution2d_8", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.014336000196635723, "best_triton_pos": 0}
AUTOTUNE convolution(3x4x64x64, 320x4x3x3)
strides: [16384, 1, 256, 4], [36, 1, 12, 4]
dtypes: torch.float16, torch.float16
triton_convolution2d_8 0.0143 ms 100.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_9 0.0143 ms 100.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1 0.0154 ms 93.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7 0.0154 ms 93.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_4 0.0164 ms 87.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_3 0.0184 ms 77.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5 0.0195 ms 73.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6 0.0195 ms 73.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11 0.0205 ms 70.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
convolution 0.0214 ms 66.9%
SingleProcess AUTOTUNE benchmarking takes 4.1413 seconds and 0.0002 seconds precompiling for 13 choices
E0801 17:00:50.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:50.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:50.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:50.851000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:50.851000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:50.851000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.14131200313568115, "best_triton_pos": 1, "best_triton_time": 0.15462400019168854, "best_triton_kernel": "triton_convolution2d_18", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(3x320x64x64, 320x320x3x3)
strides: [1310720, 1, 20480, 320], [2880, 1, 960, 320]
dtypes: torch.float16, torch.float16
convolution 0.1413 ms 100.0%
triton_convolution2d_18 0.1546 ms 91.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_21 0.1556 ms 90.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_20 0.1761 ms 80.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_19 0.1802 ms 78.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_16 0.1864 ms 75.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_15 0.2221 ms 63.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13 0.2232 ms 63.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_17 0.2243 ms 63.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_23 0.2263 ms 62.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.9585 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:00:55.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:55.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:55.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:00:56.635000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:56.635000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:56.635000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_69", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.0071680000983178616, "best_triton_pos": 0}
AUTOTUNE addmm(3x320, 3x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_69 0.0072 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_73 0.0072 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_81 0.0092 ms 77.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_66 0.0102 ms 70.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_67 0.0102 ms 70.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_72 0.0102 ms 70.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_68 0.0112 ms 63.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_76 0.0113 ms 63.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_79 0.0123 ms 58.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
addmm 0.0133 ms 53.8%
SingleProcess AUTOTUNE benchmarking takes 4.5472 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:00:58.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:00:58.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:00:58.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:00.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:00.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:00.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:02.274000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:02.274000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:02.274000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:02.700000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:02.700000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:02.700000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_109", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.030719999223947525, "best_triton_pos": 0}
AUTOTUNE mm(12288x320, 320x320)
strides: [320, 1], [1, 320]
dtypes: torch.float16, torch.float16
triton_mm_109 0.0307 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_113 0.0307 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_107 0.0317 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_111 0.0317 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_116 0.0317 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_114 0.0328 ms 93.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_117 0.0328 ms 93.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_110 0.0348 ms 88.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_115 0.0358 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
mm 0.0369 ms 83.3%
SingleProcess AUTOTUNE benchmarking takes 6.0628 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:01:05.090000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:05.090000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:05.090000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:06.846000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:06.846000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:06.846000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:09.136000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:09.136000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:09.136000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:09.720000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:09.720000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:09.720000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_224", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.008191999979317188, "best_triton_pos": 0}
AUTOTUNE mm(231x1024, 1024x320)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_224 0.0082 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
mm 0.0133 ms 61.5%
triton_mm_221 0.0143 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_222 0.0154 ms 53.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_223 0.0154 ms 53.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_227 0.0154 ms 53.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_231 0.0164 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_234 0.0174 ms 47.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_230 0.0184 ms 44.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_233 0.0195 ms 42.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.0173 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:01:12.131000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:12.131000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:12.131000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:13.760000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:13.760000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:13.760000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:16.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:16.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:16.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:17.040000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:17.040000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:17.040000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_291", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.1382399946451187, "best_triton_pos": 0}
AUTOTUNE addmm(12288x2560, 12288x320, 320x2560)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_291 0.1382 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_296 0.1382 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_297 0.1413 ms 97.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_289 0.1434 ms 96.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_293 0.1444 ms 95.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_290 0.1495 ms 92.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_294 0.1526 ms 90.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_287 0.1638 ms 84.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_286 0.1720 ms 80.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_295 0.2017 ms 68.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.3189 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:01:22.227000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:22.227000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:22.227000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:22.708000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:22.708000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:22.708000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_641", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.04198399931192398, "best_triton_pos": 0}
AUTOTUNE convolution(3x320x64x64, 320x320x3x3)
strides: [1310720, 1, 20480, 320], [2880, 1, 960, 320]
dtypes: torch.float16, torch.float16
triton_convolution2d_641 0.0420 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
convolution 0.0430 ms 97.6%
triton_convolution2d_639 0.0635 ms 66.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_640 0.0840 ms 50.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_644 0.0891 ms 47.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_643 0.1096 ms 38.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_636 0.1116 ms 37.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_642 0.1147 ms 36.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_638 0.1270 ms 33.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_648 0.1270 ms 33.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.9076 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:01:29.096000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:29.096000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:29.096000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:29.573000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:29.573000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:29.573000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_656", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8", "best_time": 0.06348799914121628, "best_triton_pos": 0}
AUTOTUNE convolution(3x320x32x32, 640x320x3x3)
strides: [327680, 1, 10240, 320], [2880, 1, 960, 320]
dtypes: torch.float16, torch.float16
triton_convolution2d_656 0.0635 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_658 0.0717 ms 88.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_657 0.0799 ms 79.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_661 0.0860 ms 73.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
convolution 0.0973 ms 65.3%
triton_convolution2d_653 0.0993 ms 63.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_659 0.1044 ms 60.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_655 0.1065 ms 59.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_652 0.1116 ms 56.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_660 0.1230 ms 51.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.8616 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:01:34.224000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:34.224000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:34.224000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:35.368000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:35.368000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:35.368000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_670", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.008191999979317188, "best_triton_pos": 0}
AUTOTUNE addmm(3x640, 3x1280, 1280x640)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_670 0.0082 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_674 0.0092 ms 88.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_682 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_667 0.0113 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_673 0.0123 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_677 0.0123 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0143 ms 57.1%
triton_mm_668 0.0143 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_669 0.0143 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_680 0.0154 ms 53.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 4.5479 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:01:37.643000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:37.643000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:37.643000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:39.257000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:39.257000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:39.257000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:42.024000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:42.024000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:42.024000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:42.522000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:42.522000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:42.522000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_695", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.013311999849975109, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x320, 320x640)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_695 0.0133 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_701 0.0143 ms 92.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_694 0.0174 ms 76.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_698 0.0184 ms 72.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_691 0.0195 ms 68.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_693 0.0225 ms 59.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_700 0.0225 ms 59.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_697 0.0236 ms 56.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_685 0.0236 ms 56.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_686 0.0256 ms 52.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1529 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:01:47.852000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:47.852000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:47.852000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:48.343000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:48.343000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:48.343000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_711", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8", "best_time": 0.11980800330638885, "best_triton_pos": 0}
AUTOTUNE convolution(3x640x32x32, 640x640x3x3)
strides: [655360, 1, 20480, 640], [5760, 1, 1920, 640]
dtypes: torch.float16, torch.float16
triton_convolution2d_711 0.1198 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_713 0.1372 ms 87.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_716 0.1464 ms 81.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_712 0.1526 ms 78.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
convolution 0.1792 ms 66.9%
triton_convolution2d_708 0.1946 ms 61.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_714 0.2007 ms 59.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_710 0.2018 ms 59.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_707 0.2150 ms 55.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_719 0.2161 ms 55.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1385 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:01:51.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:51.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:51.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:53.540000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:53.540000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:53.540000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:56.164000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:56.164000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:56.164000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:01:56.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:56.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:56.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_732", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.01945599913597107, "best_triton_pos": 0}
AUTOTUNE mm(3072x640, 640x640)
strides: [640, 1], [1, 640]
dtypes: torch.float16, torch.float16
triton_mm_732 0.0195 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_738 0.0195 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_728 0.0236 ms 82.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_731 0.0287 ms 67.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_735 0.0297 ms 65.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_737 0.0358 ms 54.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_723 0.0369 ms 52.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_724 0.0369 ms 52.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
mm 0.0379 ms 51.4%
triton_mm_722 0.0379 ms 51.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 6.9836 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:01:59.086000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:01:59.086000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:01:59.086000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:00.857000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:00.857000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:00.857000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:03.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:03.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:03.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:03.785000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:03.785000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:03.785000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_845", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.009216000325977802, "best_triton_pos": 0}
AUTOTUNE mm(231x1024, 1024x640)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_845 0.0092 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
mm 0.0123 ms 75.0%
triton_mm_842 0.0164 ms 56.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_843 0.0174 ms 52.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_848 0.0184 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_844 0.0195 ms 47.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_852 0.0205 ms 45.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_858 0.0205 ms 45.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_851 0.0215 ms 42.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_855 0.0225 ms 40.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1372 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:02:06.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:06.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:06.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:08.260000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:08.260000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:08.260000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:10.947000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:10.947000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:10.947000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:11.483000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:11.483000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:11.483000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_917", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.11673600226640701, "best_triton_pos": 0}
AUTOTUNE addmm(3072x5120, 3072x640, 640x5120)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_917 0.1167 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_918 0.1188 ms 98.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_912 0.1219 ms 95.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_910 0.1280 ms 91.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_911 0.1311 ms 89.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_914 0.1372 ms 85.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_915 0.1393 ms 83.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_908 0.1444 ms 80.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
addmm 0.1618 ms 72.2%
triton_mm_907 0.1679 ms 69.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.6959 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:02:16.921000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:16.921000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:16.921000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:17.412000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:17.412000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:17.412000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_1262", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.06860800087451935, "best_triton_pos": 0}
AUTOTUNE convolution(3x640x32x32, 640x640x3x3)
strides: [655360, 1, 20480, 640], [5760, 1, 1920, 640]
dtypes: torch.float16, torch.float16
triton_convolution2d_1262 0.0686 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1261 0.0717 ms 95.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
convolution 0.0952 ms 72.0%
triton_convolution2d_1260 0.1157 ms 59.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_1265 0.1280 ms 53.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1257 0.1516 ms 45.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_1263 0.1956 ms 35.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1259 0.2028 ms 33.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1256 0.2058 ms 33.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1258 0.2130 ms 32.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.2840 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:02:24.233000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:24.233000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:24.233000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:24.726000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:24.726000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:24.726000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_1279", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.07680000364780426, "best_triton_pos": 0}
AUTOTUNE convolution(3x640x16x16, 1280x640x3x3)
strides: [163840, 1, 10240, 640], [5760, 1, 1920, 640]
dtypes: torch.float16, torch.float16
triton_convolution2d_1279 0.0768 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1278 0.0809 ms 94.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
convolution 0.0973 ms 78.9%
triton_convolution2d_1277 0.1116 ms 68.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_1282 0.1311 ms 58.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1274 0.1546 ms 49.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_1280 0.1915 ms 40.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1276 0.1935 ms 39.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1286 0.1956 ms 39.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1273 0.2099 ms 36.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.3370 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:02:28.563000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:28.563000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:28.563000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:30.432000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:30.432000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:30.432000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:33.080000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:33.080000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:33.080000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:33.618000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:33.618000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:33.618000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_1312", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8", "best_time": 0.015359999611973763, "best_triton_pos": 0}
AUTOTUNE addmm(768x1280, 768x640, 640x1280)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_1312 0.0154 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1316 0.0164 ms 93.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0184 ms 83.3%
triton_mm_1322 0.0195 ms 78.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1306 0.0225 ms 68.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_1309 0.0225 ms 68.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_1307 0.0246 ms 62.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1308 0.0246 ms 62.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1315 0.0256 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1319 0.0256 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.5041 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:02:39.106000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:39.106000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:39.106000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:39.596000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:39.596000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:39.596000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.11366400122642517, "best_triton_pos": 1, "best_triton_time": 0.13926400244235992, "best_triton_kernel": "triton_convolution2d_1333", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(3x1280x16x16, 1280x1280x3x3)
strides: [327680, 1, 20480, 1280], [11520, 1, 3840, 1280]
dtypes: torch.float16, torch.float16
convolution 0.1137 ms 100.0%
triton_convolution2d_1333 0.1393 ms 81.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1334 0.1475 ms 77.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1332 0.1987 ms 57.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_1337 0.2458 ms 46.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1329 0.2673 ms 42.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_1330 0.3533 ms 32.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1331 0.3645 ms 31.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1328 0.3707 ms 30.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1335 0.3748 ms 30.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.3654 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:02:43.640000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:43.640000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:43.640000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:45.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:45.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:45.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:47.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:47.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:47.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:48.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:48.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:48.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.021503999829292297, "best_triton_pos": 1, "best_triton_time": 0.02457600086927414, "best_triton_kernel": "triton_mm_1349", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8"}
AUTOTUNE mm(768x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
mm 0.0215 ms 100.0%
triton_mm_1349 0.0246 ms 87.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1353 0.0256 ms 84.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1346 0.0328 ms 65.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_1359 0.0328 ms 65.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1344 0.0379 ms 56.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1343 0.0399 ms 53.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_1345 0.0399 ms 53.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1352 0.0430 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1356 0.0430 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4839 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:02:50.913000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:50.913000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:50.913000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:52.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:52.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:52.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:55.061000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:55.061000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:55.061000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:02:55.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:55.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:55.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.01228800043463707, "best_triton_pos": 1, "best_triton_time": 0.015359999611973763, "best_triton_kernel": "triton_mm_1466", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4"}
AUTOTUNE mm(231x1024, 1024x1280)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
mm 0.0123 ms 100.0%
triton_mm_1466 0.0154 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_1463 0.0174 ms 70.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_1469 0.0205 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1464 0.0225 ms 54.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_1473 0.0225 ms 54.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1479 0.0246 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1472 0.0266 ms 46.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1476 0.0266 ms 46.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1465 0.0276 ms 44.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1781 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:02:58.728000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:02:58.728000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:02:58.728000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:00.778000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:00.778000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:00.778000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:03.505000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:03.505000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:03.505000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:04.374000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:04.374000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:04.374000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_1539", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.12294399738311768, "best_triton_pos": 0}
AUTOTUNE addmm(768x10240, 768x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_1539 0.1229 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1538 0.1249 ms 98.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.1321 ms 93.1%
triton_mm_1533 0.1321 ms 93.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1531 0.1372 ms 89.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1532 0.1475 ms 83.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1536 0.1475 ms 83.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_1529 0.1485 ms 82.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_1535 0.1536 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_1528 0.1905 ms 64.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.7100 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:03:09.964000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:09.964000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:09.964000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:10.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:10.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:10.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.07577600330114365, "best_triton_pos": 1, "best_triton_time": 0.11878400295972824, "best_triton_kernel": "triton_convolution2d_1883", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(3x1280x16x16, 1280x1280x3x3)
strides: [327680, 1, 20480, 1280], [11520, 1, 3840, 1280]
dtypes: torch.float16, torch.float16
convolution 0.0758 ms 100.0%
triton_convolution2d_1883 0.1188 ms 63.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1882 0.1229 ms 61.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1881 0.1997 ms 37.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_1886 0.2345 ms 32.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1878 0.2632 ms 28.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_1880 0.3636 ms 20.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1877 0.3768 ms 20.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1879 0.3860 ms 19.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1884 0.4076 ms 18.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4443 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:03:17.289000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:17.289000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:17.289000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:17.816000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:17.816000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:17.816000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.06752000004053116, "best_triton_pos": 1, "best_triton_time": 0.12390399724245071, "best_triton_kernel": "triton_convolution2d_1899", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(3x1280x8x8, 1280x1280x3x3)
strides: [81920, 1, 10240, 1280], [11520, 1, 3840, 1280]
dtypes: torch.float16, torch.float16
convolution 0.0675 ms 100.0%
triton_convolution2d_1899 0.1239 ms 54.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1900 0.1352 ms 50.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_1898 0.1915 ms 35.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_1903 0.2355 ms 28.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1895 0.2601 ms 26.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_1896 0.3512 ms 19.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1894 0.3625 ms 18.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1897 0.3645 ms 18.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_1907 0.3718 ms 18.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.3579 seconds and 0.0003 seconds precompiling for 18 choices
E0801 17:03:21.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:21.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:21.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:23.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:23.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:23.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:26.072000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:26.072000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:26.072000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:26.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:26.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:26.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2051", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.01228800043463707, "best_triton_pos": 0}
AUTOTUNE mm(192x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
triton_mm_2051 0.0123 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
mm 0.0143 ms 85.7%
triton_mm_2048 0.0205 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2054 0.0246 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_2058 0.0276 ms 44.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2049 0.0297 ms 41.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2064 0.0297 ms 41.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2050 0.0317 ms 38.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2057 0.0339 ms 36.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2061 0.0358 ms 34.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.5292 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:03:30.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:30.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:30.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:32.201000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:32.201000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:32.201000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:34.978000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:34.978000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:34.978000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:35.864000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:35.864000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:35.864000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2131", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.01228800043463707, "best_triton_pos": 0}
AUTOTUNE addmm(192x1280, 192x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_2131 0.0123 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
addmm 0.0183 ms 67.0%
triton_mm_2128 0.0226 ms 54.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2134 0.0266 ms 46.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_2129 0.0297 ms 41.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2138 0.0297 ms 41.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2144 0.0307 ms 40.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2130 0.0317 ms 38.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2137 0.0348 ms 35.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2141 0.0360 ms 34.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 9.1144 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:03:38.842000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:38.842000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:38.842000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:40.912000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:40.912000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:40.912000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:43.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:43.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:43.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:44.571000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:44.571000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:44.571000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2238", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.06143999844789505, "best_triton_pos": 0}
AUTOTUNE addmm(192x10240, 192x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_2238 0.0614 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2231 0.0676 ms 90.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
addmm 0.0727 ms 84.5%
triton_mm_2234 0.0727 ms 84.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_2228 0.0737 ms 83.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2236 0.0748 ms 82.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2237 0.0768 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2240 0.0778 ms 78.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2243 0.0788 ms 77.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2244 0.0788 ms 77.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.7010 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:03:50.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:50.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:50.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:50.741000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:50.741000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:50.741000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.131071999669075, "best_triton_pos": 1, "best_triton_time": 0.21503999829292297, "best_triton_kernel": "triton_convolution2d_2347", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(3x2560x8x8, 1280x2560x3x3)
strides: [163840, 1, 20480, 2560], [23040, 1, 7680, 2560]
dtypes: torch.float16, torch.float16
convolution 0.1311 ms 100.0%
triton_convolution2d_2347 0.2150 ms 61.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_2348 0.2171 ms 60.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_2346 0.3626 ms 36.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_2351 0.4669 ms 28.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2343 0.4936 ms 26.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_2344 0.6799 ms 19.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2342 0.7096 ms 18.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2345 0.7158 ms 18.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2355 0.7342 ms 17.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.6246 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:03:54.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:54.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:54.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:56.355000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:56.355000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:56.355000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:58.754000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:58.754000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:58.754000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:03:59.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:03:59.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:03:59.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2378", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.020479999482631683, "best_triton_pos": 0}
AUTOTUNE addmm(192x1280, 192x2560, 2560x1280)
strides: [0, 1], [2560, 1], [1, 2560]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_2378 0.0205 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
addmm 0.0287 ms 71.4%
triton_mm_2375 0.0369 ms 55.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2381 0.0471 ms 43.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_2376 0.0522 ms 39.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2385 0.0543 ms 37.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2391 0.0594 ms 34.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2377 0.0635 ms 32.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2384 0.0696 ms 29.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2388 0.0696 ms 29.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.9942 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:04:05.299000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:05.299000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:05.299000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:05.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:05.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:05.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.21094399690628052, "best_triton_pos": 1, "best_triton_time": 0.2539519965648651, "best_triton_kernel": "triton_convolution2d_2580", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(3x2560x16x16, 1280x2560x3x3)
strides: [655360, 1, 40960, 2560], [23040, 1, 7680, 2560]
dtypes: torch.float16, torch.float16
convolution 0.2109 ms 100.0%
triton_convolution2d_2580 0.2540 ms 83.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_2581 0.2540 ms 83.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_2579 0.3728 ms 56.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_2584 0.4792 ms 44.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2576 0.5079 ms 41.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_2577 0.6799 ms 31.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2578 0.7137 ms 29.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2575 0.7219 ms 29.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_2582 0.7301 ms 28.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.1284 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:04:10.138000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:10.138000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:10.138000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:11.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:11.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:11.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:14.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:14.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:14.054000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:14.511000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:14.511000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:14.511000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_2614", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8", "best_time": 0.04608000069856644, "best_triton_pos": 0}
AUTOTUNE addmm(768x1280, 768x2560, 2560x1280)
strides: [0, 1], [2560, 1], [1, 2560]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_2614 0.0461 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
addmm 0.0512 ms 90.0%
triton_mm_2618 0.0533 ms 86.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2611 0.0594 ms 77.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_2624 0.0645 ms 71.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_2609 0.0727 ms 63.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_2617 0.0758 ms 60.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2621 0.0758 ms 60.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_2608 0.0788 ms 58.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_2610 0.0799 ms 57.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1612 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:04:20.533000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:20.533000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:20.533000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:21.068000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:21.068000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:21.068000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_3205", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.19353599846363068, "best_triton_pos": 0}
AUTOTUNE convolution(3x1920x16x16, 1280x1920x3x3)
strides: [491520, 1, 30720, 1920], [17280, 1, 5760, 1920]
dtypes: torch.float16, torch.float16
triton_convolution2d_3205 0.1935 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3204 0.1976 ms 97.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
convolution 0.2099 ms 92.2%
triton_convolution2d_3203 0.2857 ms 67.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_3208 0.3707 ms 52.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3200 0.3840 ms 50.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_3201 0.5192 ms 37.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3202 0.5376 ms 36.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3199 0.5448 ms 35.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3206 0.5519 ms 35.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.9775 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:04:25.863000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:25.863000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:25.863000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:27.372000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:27.372000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:27.372000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:29.743000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:29.743000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:29.743000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:30.197000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:30.197000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:30.197000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_3238", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8", "best_time": 0.03673600032925606, "best_triton_pos": 0}
AUTOTUNE addmm(768x1280, 768x1920, 1920x1280)
strides: [0, 1], [1920, 1], [1, 1920]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_3238 0.0367 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_3242 0.0399 ms 92.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0410 ms 89.7%
triton_mm_3235 0.0440 ms 83.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_3248 0.0502 ms 73.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3233 0.0553 ms 66.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_3241 0.0573 ms 64.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3232 0.0584 ms 62.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_3245 0.0594 ms 61.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3234 0.0624 ms 58.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.6553 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:04:36.067000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:36.067000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:36.067000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:36.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:36.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:36.592000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.418720006942749, "best_triton_pos": 1, "best_triton_time": 0.42188799381256104, "best_triton_kernel": "triton_convolution2d_3518", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(3x1280x32x32, 1280x1280x3x3)
strides: [1310720, 1, 40960, 1280], [11520, 1, 3840, 1280]
dtypes: torch.float16, torch.float16
convolution 0.4187 ms 100.0%
triton_convolution2d_3518 0.4219 ms 99.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3515 0.4495 ms 93.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_3524 0.4751 ms 88.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3523 0.4936 ms 84.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3519 0.4997 ms 83.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3511 0.5048 ms 82.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3517 0.5089 ms 82.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3513 0.5263 ms 79.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3514 0.5263 ms 79.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7792 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:04:43.832000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:43.832000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:43.832000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:44.353000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:44.353000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:44.353000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.3184640109539032, "best_triton_pos": 1, "best_triton_time": 0.32256001234054565, "best_triton_kernel": "triton_convolution2d_3532", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8"}
AUTOTUNE convolution(3x1920x32x32, 640x1920x3x3)
strides: [1966080, 1, 61440, 1920], [17280, 1, 5760, 1920]
dtypes: torch.float16, torch.float16
convolution 0.3185 ms 100.0%
triton_convolution2d_3532 0.3226 ms 98.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_3534 0.3820 ms 83.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3537 0.4014 ms 79.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3533 0.4188 ms 76.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3529 0.5294 ms 60.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_3531 0.5468 ms 58.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3528 0.5519 ms 57.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3535 0.5693 ms 55.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3540 0.5908 ms 53.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7282 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:04:49.101000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:49.101000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:49.101000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:50.613000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:50.613000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:50.613000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:52.992000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:52.992000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:52.992000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:53.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:53.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:53.445000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_3577", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.04403200000524521, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x1920, 1920x640)
strides: [0, 1], [1920, 1], [1, 1920]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_3577 0.0440 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3571 0.0451 ms 97.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3570 0.0625 ms 70.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3574 0.0635 ms 69.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3567 0.0707 ms 62.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_3576 0.0788 ms 55.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3564 0.0850 ms 51.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_3573 0.0870 ms 50.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3563 0.0881 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_3562 0.0901 ms 48.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7083 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:04:59.382000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:59.382000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:59.382000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:04:59.910000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:04:59.910000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:04:59.910000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.22937600314617157, "best_triton_pos": 1, "best_triton_time": 0.2303999960422516, "best_triton_kernel": "triton_convolution2d_3844", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8"}
AUTOTUNE convolution(3x1280x32x32, 640x1280x3x3)
strides: [1310720, 1, 40960, 1280], [11520, 1, 3840, 1280]
dtypes: torch.float16, torch.float16
convolution 0.2294 ms 100.0%
triton_convolution2d_3844 0.2304 ms 99.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_3846 0.2509 ms 91.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3849 0.2714 ms 84.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3845 0.2796 ms 82.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_3841 0.3645 ms 62.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_3843 0.3820 ms 60.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3840 0.3830 ms 59.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3847 0.3912 ms 58.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_3852 0.4004 ms 57.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.8544 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:05:04.376000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:04.376000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:04.376000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:06.015000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:06.015000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:06.015000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:08.819000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:08.819000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:08.819000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:09.717000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:09.717000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:09.717000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_3889", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.030719999223947525, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x1280, 1280x640)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_3889 0.0307 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3883 0.0317 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3882 0.0430 ms 71.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3879 0.0461 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_3886 0.0461 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_3876 0.0614 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_3875 0.0625 ms 49.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_3874 0.0635 ms 48.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_3888 0.0645 ms 47.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_3873 0.0707 ms 43.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.4032 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:05:15.534000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:15.534000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:15.534000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:16.056000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:16.056000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:16.056000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.17612800002098083, "best_triton_pos": 1, "best_triton_time": 0.17715199291706085, "best_triton_kernel": "triton_convolution2d_4156", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8"}
AUTOTUNE convolution(3x960x32x32, 640x960x3x3)
strides: [983040, 1, 30720, 960], [8640, 1, 2880, 960]
dtypes: torch.float16, torch.float16
convolution 0.1761 ms 100.0%
triton_convolution2d_4156 0.1772 ms 99.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_4158 0.1925 ms 91.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4161 0.2171 ms 81.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4157 0.2232 ms 78.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4153 0.2806 ms 62.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_4155 0.2908 ms 60.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4152 0.2949 ms 59.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4159 0.2980 ms 59.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4164 0.3267 ms 53.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.6931 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:05:20.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:20.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:20.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:21.815000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:21.815000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:21.815000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:24.691000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:24.691000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:24.691000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:25.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:25.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:25.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_4195", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.02457600086927414, "best_triton_pos": 0}
AUTOTUNE addmm(3072x640, 3072x960, 960x640)
strides: [0, 1], [960, 1], [1, 960]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_4195 0.0246 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4201 0.0246 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4191 0.0317 ms 77.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_4194 0.0348 ms 70.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4198 0.0358 ms 68.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4186 0.0481 ms 51.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_4187 0.0481 ms 51.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_4200 0.0492 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4188 0.0501 ms 49.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_4185 0.0532 ms 46.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.0854 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:05:31.163000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:31.163000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:31.163000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:31.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:31.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:31.687000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.41574400663375854, "best_triton_pos": 1, "best_triton_time": 0.4700160026550293, "best_triton_kernel": "triton_convolution2d_4476", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(3x640x64x64, 640x640x3x3)
strides: [2621440, 1, 40960, 640], [5760, 1, 1920, 640]
dtypes: torch.float16, torch.float16
convolution 0.4157 ms 100.0%
triton_convolution2d_4476 0.4700 ms 88.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4468 0.4710 ms 88.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_4471 0.4731 ms 87.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4470 0.5007 ms 83.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4469 0.5120 ms 81.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4464 0.5150 ms 80.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4467 0.5283 ms 78.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4473 0.5652 ms 73.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4466 0.6349 ms 65.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4977 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:05:38.656000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:38.656000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:38.656000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:39.182000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:39.182000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:39.182000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.3491840064525604, "best_triton_pos": 1, "best_triton_time": 0.4034560024738312, "best_triton_kernel": "triton_convolution2d_4484", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(3x960x64x64, 320x960x3x3)
strides: [3932160, 1, 61440, 960], [8640, 1, 2880, 960]
dtypes: torch.float16, torch.float16
convolution 0.3492 ms 100.0%
triton_convolution2d_4484 0.4035 ms 86.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4487 0.4301 ms 81.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4486 0.4700 ms 74.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4485 0.4947 ms 70.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_4479 0.5775 ms 60.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_4481 0.5898 ms 59.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4490 0.6001 ms 58.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4483 0.6246 ms 55.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4493 0.6359 ms 54.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4683 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:05:43.245000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:43.245000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:43.245000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:44.853000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:44.853000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:44.853000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:47.714000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:47.714000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:47.714000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:48.524000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:48.524000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:48.524000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_4524", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.06656000018119812, "best_triton_pos": 0}
AUTOTUNE addmm(12288x320, 12288x960, 960x320)
strides: [0, 1], [960, 1], [1, 960]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_4524 0.0666 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4520 0.0717 ms 92.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_4527 0.0727 ms 91.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4530 0.0737 ms 90.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4523 0.0778 ms 85.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4522 0.0788 ms 84.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4529 0.0850 ms 78.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4526 0.0891 ms 74.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0931 ms 71.5%
triton_mm_4515 0.0952 ms 69.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.0318 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:05:54.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:54.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:54.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:05:54.598000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:54.598000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:54.598000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.2457599937915802, "best_triton_pos": 1, "best_triton_time": 0.2877439856529236, "best_triton_kernel": "triton_convolution2d_4796", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(3x640x64x64, 320x640x3x3)
strides: [2621440, 1, 40960, 640], [5760, 1, 1920, 640]
dtypes: torch.float16, torch.float16
convolution 0.2458 ms 100.0%
triton_convolution2d_4796 0.2877 ms 85.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4799 0.2918 ms 84.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4798 0.3277 ms 75.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_4797 0.3379 ms 72.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_4794 0.3492 ms 70.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_4802 0.3830 ms 64.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4795 0.4035 ms 60.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4793 0.4055 ms 60.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_4791 0.4065 ms 60.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.3755 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:05:58.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:05:58.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:05:58.470000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:06:00.400000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:06:00.400000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:06:00.400000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:06:03.119000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:06:03.119000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:06:03.119000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
E0801 17:06:03.672000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Runtime error during autotuning:
E0801 17:06:03.672000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:06:03.672000 2856711 torch/_inductor/select_algorithm.py:4888] [0/0] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_4836", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.04915200173854828, "best_triton_pos": 0}
AUTOTUNE addmm(12288x320, 12288x640, 640x320)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_4836 0.0492 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4839 0.0522 ms 94.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4832 0.0532 ms 92.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_4842 0.0543 ms 90.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4841 0.0563 ms 87.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4834 0.0604 ms 81.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4835 0.0604 ms 81.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_4838 0.0604 ms 81.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_4827 0.0655 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_4828 0.0655 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7547 seconds and 0.0002 seconds precompiling for 21 choices
Out[17]:
JIT Compilation & Warmup Complete! (Compilation Warmup took: 541.46 s) Step 2: Running 5 Steady-State Benchmark Iterations... Run 1/5: 714.04 ms (0.714 s) Run 2/5: 579.69 ms (0.580 s) Run 3/5: 579.85 ms (0.580 s) Run 4/5: 580.08 ms (0.580 s) Run 5/5: 580.36 ms (0.580 s) ====================================================================== BASELINE SONA PROFILE BREAKDOWN (Averaged over 5 runs) ====================================================================== 1. VAE Encode (Image -> Latent) : 50.14 ms ( 8.3%) 2. Text Encoder (CLIP Embeddings) : 0.00 ms ( 0.0%) 3. UNet Inferences (3x per step) : 511.30 ms ( 84.3%) 4. Mask Math & Scheduler Steps : 6.97 ms ( 1.1%) 5. VAE Decode (Latent -> Image) : 37.24 ms ( 6.1%) ---------------------------------------------------------------------- TOTAL WARM-START LATENCY PER IMAGE : 606.80 ms (0.607 s) ======================================================================
In [18]:
print(f"5-WAY COMPARISON: BASELINE vs OPT 1 vs OPT 2 vs OPT 3 vs OPT 4 (TORCH.COMPILE)")
print("="*95)
print(f"Phase Latency │ Baseline │ Opt 1 │ Opt 2 │ Opt 3 │ Opt 4 (torch.compile)")
print("-" * 95)
print(f"UNet Inferences Latency │ {avg_metrics['unet_inferences_ms']:8.2f} ms │ {avg_opt1['unet_inferences_ms']:9.2f} ms │ {avg_opt2['unet_inferences_ms']:9.2f} ms │ {avg_opt3['unet_inferences_ms']:9.2f} ms │ {avg_opt4['unet_inferences_ms']:10.2f} ms")
print(f"Total Warm-Start Latency │ {baseline_total:8.2f} ms │ {opt1_total:9.2f} ms │ {avg_opt2['total_warm_start_ms']:9.2f} ms │ {avg_opt3['total_warm_start_ms']:9.2f} ms │ {avg_opt4['total_warm_start_ms']:10.2f} ms")
print("-" * 95)
speedup_opt4 = baseline_total / avg_opt4['total_warm_start_ms']
print(f"OVERALL SPEEDUP FACTOR │ 1.00x │ {speedup:9.2f}x │ {speedup_opt2:9.2f}x │ {speedup_opt3:9.2f}x │ {speedup_opt4:10.2f}x")
print("="*95) Out[18]:
5-WAY COMPARISON: BASELINE vs OPT 1 vs OPT 2 vs OPT 3 vs OPT 4 (TORCH.COMPILE) =============================================================================================== Phase Latency │ Baseline │ Opt 1 │ Opt 2 │ Opt 3 │ Opt 4 (torch.compile) ----------------------------------------------------------------------------------------------- UNet Inferences Latency │ 1600.08 ms │ 611.17 ms │ 614.98 ms │ 602.75 ms │ 511.30 ms Total Warm-Start Latency │ 1691.39 ms │ 685.75 ms │ 683.82 ms │ 669.58 ms │ 606.80 ms ----------------------------------------------------------------------------------------------- OVERALL SPEEDUP FACTOR │ 1.00x │ 2.47x │ 2.47x │ 2.53x │ 2.79x ===============================================================================================
In [19]:
"""
Optimization 1 (batched UNET Pass) + optimization 2 (text embedding cache) + optimization 3 (inference)
+ optimization 4 (compilation) + optimization 5 (image batch process)
"""
def profiled_generate_sona_outlier_optimized5(
pipe,
input_image_paths, # List
id_class_prompts, # List
ood_target_prompts, # List
stop_timestep=20,
guidance_scale=15.0,
threshold_percentile=0.10,
adaptive_mask=False,
device="cuda",
seed=42,
):
"""
Executes baseline SONA with fine-grained GPU event timing for performance analysis.
"""
B = len(input_image_paths)
# Create CUDA Events for precise timing
start_event = torch.cuda.Event(enable_timing=True)
# vae_enc_event = torch.cuda.Event(enable_timing=True)
# text_enc_event = torch.cuda.Event(enable_timing=True)
# unet_loop_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
# -------------------------------------------------------------
# PHASE 1: Image Preprocessing & VAE Encode
# -------------------------------------------------------------
# Optimization 5: Batch image processing
img_tensors = []
for path in input_image_paths:
raw_img = Image.open(path).convert("RGB").resize((512, 512))
arr = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
img_tensors.append(torch.from_numpy(arr).permute(2, 0, 1))
img_batch = torch.stack(img_tensors, dim=0).to(device, dtype=pipe.text_encoder.dtype)
# raw_img = Image.open(input_image_path).convert("RGB").resize((512, 512))
# img_tensor = np.array(raw_img).astype(np.float32) / 127.5 - 1.0
# img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.unet.dtype)
# img_tensor = torch.from_numpy(img_tensor).permute(2, 0, 1).unsqueeze(0).to(device, dtype=pipe.text_encoder.dtype)
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(seed)
# opt 3
with torch.inference_mode():
init_latents = pipe.vae.encode(img_batch).latent_dist.sample(generator=generator)
init_latents = init_latents * pipe.vae.config.scaling_factor
# vae_enc_event.record()
# -------------------------------------------------------------
# PHASE 2: Tokenize & Text Encoder Embeddings
# -------------------------------------------------------------
# text_prompts = [id_class_prompt, ood_target_prompt, ""]
# text_inputs = pipe.tokenizer(
# text_prompts,
# padding="max_length",
# max_length=pipe.tokenizer.model_max_length,
# return_tensors="pt"
# )
# with torch.no_grad():
# embeddings = pipe.text_encoder(text_inputs.input_ids.to(device))[0]
# Optimization 5: Batch image processing
c_id_list, c_ood_list, c_uncond_list = [], [], []
c_uncond_single = get_text_embedding(pipe, "", device=device)
for i in range(B):
c_id_list.append(get_text_embedding(pipe, id_class_prompts[i], device=device))
c_ood_list.append(get_text_embedding(pipe, ood_target_prompts[i], device=device))
c_uncond_list.append(c_uncond_single)
# c_id = get_text_embedding(pipe, id_class_prompt, device=device)
# c_ood = get_text_embedding(pipe, ood_target_prompt, device=device)
# c_uncond = get_text_embedding(pipe, "", device=device)
# text_enc_event.record()
c_uncond_b = torch.cat(c_uncond_list, dim=0)
c_id_b = torch.cat(c_id_list, dim=0)
c_ood_b = torch.cat(c_ood_list, dim=0)
c_batched = torch.cat([c_uncond_b, c_id_b, c_ood_b], dim=0)
# -------------------------------------------------------------
# PHASE 3: Forward Noise Injection & Reverse Denoising Loop
# -------------------------------------------------------------
pipe.scheduler.set_timesteps(50)
timesteps = pipe.scheduler.timesteps
start_step_idx = 50 - stop_timestep
t_start = timesteps[start_step_idx]
# deterministic image generation
generator = torch.Generator(device=device).manual_seed(seed)
noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=init_latents.dtype)
z_t = pipe.scheduler.add_noise(init_latents, noise, t_start)
# z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
# c_batched = torch.cat([c_uncond, c_id, c_ood], dim=0)
# Accumulators for inner loop breakdown
# total_unet_time_ms = 0.0
# total_math_time_ms = 0.0
# u_start = torch.cuda.Event(enable_timing=True)
# u_mid = torch.cuda.Event(enable_timing=True)
# u_end = torch.cuda.Event(enable_timing=True)
with torch.inference_mode():
for i in range(start_step_idx, len(timesteps)):
t = timesteps[i]
# u_start.record()
# 3 Separate Sequential UNet forward calls
# eps_uncond = pipe.unet(z_t, t, encoder_hidden_states=c_uncond).sample
# eps_id_cond = pipe.unet(z_t, t, encoder_hidden_states=c_id).sample
# eps_ood_cond = pipe.unet(z_t, t, encoder_hidden_states=c_ood).sample
z_t_batched = torch.cat([z_t, z_t, z_t], dim=0)
eps_batched = pipe.unet(z_t_batched, t, encoder_hidden_states=c_batched).sample
eps_uncond, eps_id_cond, eps_ood_cond = eps_batched.chunk(3, dim=0)
# u_mid.record()
# Vector math, dynamic quantile mask calculation, guidance assembly
psi_id = eps_id_cond - eps_uncond
psi_ood = eps_ood_cond - eps_uncond
mag_id = torch.abs(psi_id).mean(dim=1, keepdim=True)
mag_ood = torch.abs(psi_ood).mean(dim=1, keepdim=True)
if adaptive_mask:
max_val = mag_id.max()
active_pixels = (mag_id >= 0.15 * max_val).float()
coverage = active_pixels.mean().item()
t_percentile = np.clip(coverage * 1.2, 0.08, 0.40)
else:
t_percentile = threshold_percentile
# eta_s = torch.quantile(mag_id.float().flatten(), 1.0 - t_percentile).to(mag_id.dtype)
# m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
# eta_n = torch.quantile(mag_id.float().flatten(), threshold_percentile).to(mag_id.dtype)
# m_n_id = (mag_id < eta_n).to(psi_id.dtype)
mag_id_flat = mag_id.float().view(B, -1)
eta_s = torch.quantile(mag_id_flat, 1.0 - t_percentile, dim=1).view(B, 1, 1, 1).to(mag_id.dtype)
m_s_id = (mag_id >= eta_s).to(psi_id.dtype)
eta_n = torch.quantile(mag_id_flat, t_percentile, dim=1).view(B, 1, 1, 1).to(mag_id.dtype)
m_n_id = (mag_id <= eta_n).to(psi_id.dtype)
if adaptive_mask:
max_ood = mag_ood.max()
active_ood = (mag_ood >= 0.15 * max_ood).float()
coverage_ood = active_ood.mean().item()
t_percentile_ood = np.clip(coverage_ood * 1.2, 0.08, 0.40)
else:
t_percentile_ood = threshold_percentile
# eta_ood = torch.quantile(mag_ood.float().flatten(), 1.0 - t_percentile_ood).to(mag_ood.dtype)
# m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
mag_ood_flat = mag_ood.float().view(B, -1)
eta_ood = torch.quantile(mag_ood_flat, 1.0 - t_percentile_ood, dim=1).view(B, 1, 1, 1).to(mag_ood.dtype)
m_s_ood = (mag_ood >= eta_ood).to(psi_ood.dtype)
delta_id = -m_s_id * psi_id
delta_ood = m_s_ood * (1.0 - m_n_id) * psi_ood
delta_n = m_n_id * psi_id
eps_guided = eps_uncond + guidance_scale * (delta_id + delta_ood + delta_n)
z_t = pipe.scheduler.step(eps_guided, t, z_t).prev_sample
# u_end.record()
torch.cuda.synchronize()
# total_unet_time_ms += u_start.elapsed_time(u_mid)
# total_math_time_ms += u_mid.elapsed_time(u_end)
# unet_loop_event.record()
# -------------------------------------------------------------
# PHASE 4: VAE Decode Latents to Image
# -------------------------------------------------------------
with torch.no_grad():
decoded = pipe.vae.decode(z_t / pipe.vae.config.scaling_factor).sample
decoded = (decoded / 2 + 0.5).clamp(0, 1)
decoded = decoded.cpu().permute(0, 2, 3, 1).float().numpy()
# out_img = Image.fromarray((decoded * 255).astype(np.uint8)).resize(raw_img.size)
out_imgs = []
for i in range(B):
img_np = (decoded[i] * 255).astype(np.uint8)
out_imgs.append(Image.fromarray(img_np).resize((512, 512)))
end_event.record()
torch.cuda.synchronize()
# Compute phase timings in milliseconds
# metrics = {
# "vae_encode_ms": start_event.elapsed_time(vae_enc_event),
# "text_encode_ms": vae_enc_event.elapsed_time(text_enc_event),
# "unet_inferences_ms": total_unet_time_ms,
# "mask_and_scheduler_math_ms": total_math_time_ms,
# "vae_decode_ms": unet_loop_event.elapsed_time(end_event),
# "total_warm_start_ms": start_event.elapsed_time(end_event),
# "steps_count": len(timesteps) - start_step_idx
# }
total_batch_ms = start_event.elapsed_time(end_event)
per_image_effective_ms = total_batch_ms / B
# return out_img, metrics,
return out_imgs, total_batch_ms, per_image_effective_ms
In [20]:
batch_paths = [test_img_path] * 32
batch_id_prompts = [id_prompt] * 32
batch_ood_prompts = [ood_prompt] * 32
for B_size in [1, 2, 4, 8, 16, 32]:
print(f"\n Testing Micro-Batch Size B={B_size}...")
sub_paths = batch_paths[:B_size]
sub_id = batch_id_prompts[:B_size]
sub_ood = batch_ood_prompts[:B_size]
# Reset peak memory tracker
torch.cuda.reset_peak_memory_stats()
# 1. Warmup run specifically for this batch size shape
_, _, _ = profiled_generate_sona_outlier_optimized5(
pipe, sub_paths, sub_id, sub_ood, stop_timestep=23
)
# 2. Benchmark runs
r_total_ms, r_per_img_ms = [], []
for run in range(3):
_, total_ms, per_img_ms = profiled_generate_sona_outlier_optimized5(
pipe, sub_paths, sub_id, sub_ood, stop_timestep=23
)
r_total_ms.append(total_ms)
r_per_img_ms.append(per_img_ms)
avg_total = np.mean(r_total_ms)
avg_per_img = np.mean(r_per_img_ms)
eff_speedup = baseline_total / avg_per_img
peak_vram = torch.cuda.max_memory_allocated() / 1e9
print(f" Batch B={B_size} Summary:")
print(f" Total Batch Time : {avg_total:8.2f} ms")
print(f" Effective Per-Image Time : {avg_per_img:8.2f} ms")
print(f" Peak VRAM Memory : {peak_vram:8.2f} GB")
print(f" Effective Speedup Factor : {eff_speedup:8.2f}x Faster!")
Out[20]:
Testing Micro-Batch Size B=1...
Batch B=1 Summary:
Total Batch Time : 584.24 ms
Effective Per-Image Time : 584.24 ms
Peak VRAM Memory : 3.25 GB
Effective Speedup Factor : 2.90x Faster!
Testing Micro-Batch Size B=2...
Out[20]:
E0801 17:08:43.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:43.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:43.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:45.350000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:45.350000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:45.350000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:48.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:48.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:48.193000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:49.149000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:49.149000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:49.149000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_5740", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.16076800227165222, "best_triton_pos": 0}
AUTOTUNE addmm(24576x320, 24576x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5740 0.1608 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5736 0.1618 ms 99.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_5746 0.1679 ms 95.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5738 0.1894 ms 84.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5739 0.1894 ms 84.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5743 0.1915 ms 84.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5742 0.1966 ms 81.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5745 0.2028 ms 79.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.2191 ms 73.4%
triton_mm_5731 0.2222 ms 72.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 9.5988 seconds and 0.0008 seconds precompiling for 21 choices
E0801 17:08:52.664000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:52.664000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:52.664000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:54.253000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:54.253000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:54.253000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:56.699000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:56.699000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:56.699000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:08:57.171000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:08:57.171000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:08:57.171000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6364", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.1228799968957901, "best_triton_pos": 0}
AUTOTUNE addmm(6144x640, 6144x2560, 2560x640)
strides: [0, 1], [2560, 1], [1, 2560]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6364 0.1229 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6358 0.1249 ms 98.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6363 0.1260 ms 97.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.1270 ms 96.8%
triton_mm_6360 0.1403 ms 87.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6361 0.1454 ms 84.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6357 0.1464 ms 83.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6356 0.1608 ms 76.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6362 0.1608 ms 76.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_6354 0.1638 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4537 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:09:00.241000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:00.241000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:00.241000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:01.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:01.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:01.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:04.360000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:04.360000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:04.360000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:04.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:04.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:04.841000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "addmm", "best_time": 0.11468800157308578, "best_triton_pos": 1, "best_triton_time": 0.11468800157308578, "best_triton_kernel": "triton_mm_6982", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4"}
AUTOTUNE addmm(1536x1280, 1536x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
addmm 0.1147 ms 100.0%
triton_mm_6982 0.1147 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6976 0.1198 ms 95.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6979 0.1495 ms 76.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6975 0.1536 ms 74.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6981 0.1567 ms 73.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6978 0.1597 ms 71.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6974 0.1618 ms 70.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6972 0.1802 ms 63.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6971 0.1946 ms 58.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.1845 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.17100800573825836, "best_triton_pos": 1, "best_triton_time": 0.6144000291824341, "best_triton_kernel": "triton_convolution2d_7410", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(6x1280x8x8, 1280x1280x3x3)
strides: [81920, 64, 8, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.1710 ms 100.0%
triton_convolution2d_7410 0.6144 ms 27.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7405 0.6369 ms 26.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7403 0.7055 ms 24.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=512, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_7407 0.7178 ms 23.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7408 0.9216 ms 18.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_7402 0.9544 ms 17.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7404 0.9923 ms 17.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7415 1.6589 ms 10.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7409 1.7439 ms 9.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 10.8768 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:09:19.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:19.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:19.023000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:20.629000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:20.629000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:20.629000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:23.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:23.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:23.103000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:23.572000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:23.572000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:23.572000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "addmm", "best_time": 0.053247999399900436, "best_triton_pos": 1, "best_triton_time": 0.060416001826524734, "best_triton_kernel": "triton_mm_7674", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4"}
AUTOTUNE addmm(384x1280, 384x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
addmm 0.0532 ms 100.0%
triton_mm_7674 0.0604 ms 88.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_7677 0.0778 ms 68.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_7671 0.0799 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_7681 0.0860 ms 61.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7672 0.0911 ms 58.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7673 0.0973 ms 54.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7684 0.1065 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_7687 0.1065 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7680 0.1075 ms 49.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.0764 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 16, "num_triton_choices": 15, "best_kernel": "triton_convolution2d_10808", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4", "best_time": 0.08601599931716919, "best_triton_pos": 0}
AUTOTUNE convolution(6x320x64x64, 4x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_10808 0.0860 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_10805 0.0870 ms 98.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_10809 0.0901 ms 95.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_10811 0.0922 ms 93.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10807 0.0932 ms 92.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_10810 0.0932 ms 92.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10804 0.0973 ms 88.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10815 0.1014 ms 84.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10806 0.1034 ms 83.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10801 0.1065 ms 80.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 5.0178 seconds and 0.0002 seconds precompiling for 16 choices
E0801 17:09:32.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:32.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:32.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:33.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:33.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:33.604000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_5465", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.006144000217318535, "best_triton_pos": 0}
AUTOTUNE addmm(6x1280, 6x320, 320x1280)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5465 0.0061 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5466 0.0061 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5459 0.0062 ms 98.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_5462 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_5469 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5474 0.0072 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5460 0.0082 ms 75.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5461 0.0082 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_5471 0.0082 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5472 0.0082 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 4.5489 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:09:37.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:37.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:37.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:09:39.178000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:09:39.178000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:09:39.178000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_5484", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.011168000288307667, "best_triton_pos": 0}
AUTOTUNE addmm(6x1280, 6x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5484 0.0112 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5480 0.0113 ms 99.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_5492 0.0124 ms 90.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5477 0.0133 ms 83.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_5483 0.0143 ms 78.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5487 0.0143 ms 77.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0164 ms 68.2%
triton_mm_5478 0.0195 ms 57.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5479 0.0206 ms 54.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_5490 0.0215 ms 51.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 5.2043 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "triton_convolution2d_5433", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4", "best_time": 0.029632000252604485, "best_triton_pos": 0}
AUTOTUNE convolution(6x4x64x64, 320x4x3x3)
strides: [16384, 4096, 64, 1], [36, 9, 3, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_5433 0.0296 ms 100.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_5430 0.0317 ms 93.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_5436 0.0317 ms 93.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_5438 0.0317 ms 93.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5432 0.0328 ms 90.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5435 0.0328 ms 90.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5431 0.0338 ms 87.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_5434 0.0348 ms 85.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5437 0.0348 ms 85.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5439 0.0348 ms 85.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 4.0781 seconds and 0.0002 seconds precompiling for 13 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.29183998703956604, "best_triton_pos": 1, "best_triton_time": 0.7905279994010925, "best_triton_kernel": "triton_convolution2d_5442", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x320x64x64, 320x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.2918 ms 100.0%
triton_convolution2d_5442 0.7905 ms 36.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_5447 0.8172 ms 35.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5443 1.1930 ms 24.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_5444 1.3158 ms 22.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5448 1.3261 ms 22.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_5451 1.3588 ms 21.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5450 1.6466 ms 17.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_5456 1.7418 ms 16.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_5454 2.1770 ms 13.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.7270 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:10:14.770000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:14.770000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:14.770000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:16.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:16.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:16.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_5498", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.0071680000983178616, "best_triton_pos": 0}
AUTOTUNE addmm(6x320, 6x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5498 0.0072 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_5502 0.0072 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5510 0.0092 ms 77.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5501 0.0102 ms 70.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5495 0.0113 ms 63.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_5496 0.0113 ms 63.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_5497 0.0113 ms 63.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
addmm 0.0123 ms 58.3%
triton_mm_5505 0.0123 ms 58.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5508 0.0123 ms 58.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 5.1323 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:10:18.404000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:18.404000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:18.404000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:19.983000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:19.983000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:19.983000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:22.434000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:22.434000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:22.434000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:22.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:22.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:22.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_5540", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.05632000043988228, "best_triton_pos": 0}
AUTOTUNE mm(24576x320, 320x320)
strides: [320, 1], [1, 320]
dtypes: torch.float16, torch.float16
triton_mm_5540 0.0563 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5536 0.0573 ms 98.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_5546 0.0573 ms 98.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5543 0.0604 ms 93.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5539 0.0614 ms 91.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5535 0.0645 ms 87.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_5545 0.0645 ms 87.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
mm 0.0655 ms 85.9%
triton_mm_5538 0.0655 ms 85.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5544 0.0676 ms 83.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.8561 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:10:25.482000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:25.482000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:25.482000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:27.238000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:27.238000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:27.238000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:30.251000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:30.251000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:30.251000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:30.783000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:30.783000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:30.783000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_5620", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.05734400078654289, "best_triton_pos": 0}
AUTOTUNE addmm(24576x320, 24576x320, 320x320)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5620 0.0573 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5616 0.0584 ms 98.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_5618 0.0614 ms 93.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5623 0.0614 ms 93.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5619 0.0635 ms 90.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5626 0.0635 ms 90.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5625 0.0666 ms 86.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5622 0.0686 ms 83.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5612 0.0707 ms 81.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5611 0.0717 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.8709 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:10:33.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:33.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:33.472000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:35.135000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:35.135000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:35.135000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:37.418000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:37.418000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:37.418000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:38.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:38.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:38.065000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_5653", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.009216000325977802, "best_triton_pos": 0}
AUTOTUNE mm(462x1024, 1024x320)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_5653 0.0092 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
mm 0.0123 ms 75.0%
triton_mm_5650 0.0164 ms 56.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_5652 0.0174 ms 52.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5651 0.0184 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_5656 0.0184 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_5660 0.0205 ms 45.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5659 0.0215 ms 42.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5666 0.0215 ms 42.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5663 0.0225 ms 40.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.2797 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:10:40.705000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:40.705000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:40.705000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:42.488000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:42.488000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:42.488000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:45.558000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:45.558000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:45.558000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:10:46.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:10:46.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:10:46.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_5725", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.2519040107727051, "best_triton_pos": 0}
AUTOTUNE addmm(24576x2560, 24576x320, 320x2560)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_5725 0.2519 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5726 0.2632 ms 95.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5720 0.2673 ms 94.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5718 0.2724 ms 92.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5722 0.2775 ms 90.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_5719 0.2847 ms 88.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5723 0.2908 ms 86.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_5716 0.3215 ms 78.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_5715 0.3256 ms 77.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_5724 0.3697 ms 68.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.0099 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.1709119975566864, "best_triton_pos": 1, "best_triton_time": 0.3819519877433777, "best_triton_kernel": "triton_convolution2d_6062", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x320x64x64, 320x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.1709 ms 100.0%
triton_convolution2d_6062 0.3820 ms 44.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6067 0.4536 ms 37.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6070 0.5253 ms 32.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6064 0.6195 ms 27.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6068 0.6400 ms 26.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6074 0.7137 ms 23.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6072 0.7158 ms 23.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6073 0.7168 ms 23.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6063 0.7209 ms 23.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=1, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.7007 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.22118400037288666, "best_triton_pos": 1, "best_triton_time": 0.40857601165771484, "best_triton_kernel": "triton_convolution2d_6084", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x320x32x32, 640x320x3x3)
strides: [327680, 1024, 32, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.2212 ms 100.0%
triton_convolution2d_6084 0.4086 ms 54.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6079 0.4280 ms 51.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6080 0.6031 ms 36.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_6088 0.6298 ms 35.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6085 0.6513 ms 34.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6081 0.6646 ms 33.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6093 0.7619 ms 29.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6087 0.8274 ms 26.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6089 1.0988 ms 20.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.3404 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:11:11.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:11.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:11.907000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:13.159000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:13.159000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 110592 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:13.159000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 18, "best_kernel": "triton_mm_6099", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2", "best_time": 0.008191999979317188, "best_triton_pos": 0}
AUTOTUNE addmm(6x640, 6x1280, 1280x640)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6099 0.0082 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_6103 0.0082 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6111 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6096 0.0113 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=2
triton_mm_6102 0.0113 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6098 0.0123 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=2
triton_mm_6106 0.0123 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=16, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0133 ms 61.5%
triton_mm_6097 0.0133 ms 61.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6109 0.0143 ms 57.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=16, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 5.0159 seconds and 0.0002 seconds precompiling for 19 choices
E0801 17:11:18.271000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:18.271000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:18.271000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:18.768000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:18.768000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:18.768000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.025599999353289604, "best_triton_pos": 1, "best_triton_time": 0.025599999353289604, "best_triton_kernel": "triton_convolution2d_6120", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8"}
AUTOTUNE convolution(6x320x32x32, 640x320x1x1)
strides: [327680, 1024, 32, 1], [320, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.0256 ms 100.0%
triton_convolution2d_6120 0.0256 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_6121 0.0276 ms 92.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_6122 0.0276 ms 92.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_6123 0.0276 ms 92.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6128 0.0276 ms 92.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6119 0.0287 ms 89.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6125 0.0317 ms 80.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6116 0.0348 ms 73.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6124 0.0358 ms 71.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.7201 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.27033600211143494, "best_triton_pos": 1, "best_triton_time": 0.7874559760093689, "best_triton_kernel": "triton_convolution2d_6136", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x640x32x32, 640x640x3x3)
strides: [655360, 1024, 32, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.2703 ms 100.0%
triton_convolution2d_6136 0.7875 ms 34.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6131 0.8223 ms 32.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6132 1.1540 ms 23.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_6140 1.2216 ms 22.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6137 1.2687 ms 21.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6133 1.2758 ms 21.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6145 1.4070 ms 19.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6139 1.6241 ms 16.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6141 2.1494 ms 12.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.5121 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:11:32.806000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:32.806000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:32.806000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:34.590000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:34.590000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:34.590000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:37.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:37.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:37.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:37.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:37.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:37.843000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6163", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.03276799991726875, "best_triton_pos": 0}
AUTOTUNE mm(6144x640, 640x640)
strides: [640, 1], [1, 640]
dtypes: torch.float16, torch.float16
triton_mm_6163 0.0328 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6158 0.0348 ms 94.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6164 0.0369 ms 88.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6160 0.0399 ms 82.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
mm 0.0420 ms 78.0%
triton_mm_6156 0.0420 ms 78.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6157 0.0430 ms 76.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6154 0.0440 ms 74.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6161 0.0481 ms 68.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6162 0.0645 ms 50.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.4461 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:11:40.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:40.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:40.480000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:42.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:42.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:42.437000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:45.240000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:45.240000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:45.240000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:45.820000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:45.820000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:45.820000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6243", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.03379200026392937, "best_triton_pos": 0}
AUTOTUNE addmm(6144x640, 6144x640, 640x640)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6243 0.0338 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6238 0.0358 ms 94.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6244 0.0399 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6240 0.0420 ms 80.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6237 0.0430 ms 78.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6234 0.0451 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6241 0.0492 ms 68.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6236 0.0502 ms 67.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0584 ms 57.9%
triton_mm_6230 0.0666 ms 50.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.9730 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:11:48.512000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:48.512000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:48.512000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:50.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:50.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:50.200000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:52.527000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:52.527000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:52.527000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:53.176000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:53.176000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:53.176000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.01228800043463707, "best_triton_pos": 1, "best_triton_time": 0.014336000196635723, "best_triton_kernel": "triton_mm_6271", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4"}
AUTOTUNE mm(462x1024, 1024x640)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
mm 0.0123 ms 100.0%
triton_mm_6271 0.0143 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6268 0.0174 ms 70.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_6274 0.0195 ms 63.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6270 0.0195 ms 63.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6269 0.0205 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6278 0.0225 ms 54.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6277 0.0246 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6281 0.0256 ms 48.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6284 0.0276 ms 44.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.3539 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:11:56.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:56.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:56.049000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:11:58.126000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:11:58.126000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:11:58.126000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:01.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:01.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:01.084000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:01.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:01.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:01.679000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6343", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.21401600539684296, "best_triton_pos": 0}
AUTOTUNE addmm(6144x5120, 6144x640, 640x5120)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6343 0.2140 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6344 0.2222 ms 96.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6338 0.2335 ms 91.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6336 0.2447 ms 87.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6337 0.2519 ms 85.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6340 0.2529 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6341 0.2611 ms 82.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6334 0.2867 ms 74.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6342 0.2888 ms 74.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
addmm 0.3072 ms 69.7%
SingleProcess AUTOTUNE benchmarking takes 8.4972 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.130048006772995, "best_triton_pos": 1, "best_triton_time": 0.5427200198173523, "best_triton_kernel": "triton_convolution2d_6688", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(6x640x32x32, 640x640x3x3)
strides: [655360, 1024, 32, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.1300 ms 100.0%
triton_convolution2d_6688 0.5427 ms 24.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6686 0.7188 ms 18.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6680 0.7363 ms 17.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6682 0.7475 ms 17.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6683 0.7526 ms 17.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6685 0.8684 ms 15.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6684 1.0691 ms 12.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6687 1.1315 ms 11.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6693 1.2104 ms 10.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 11.7219 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.17510400712490082, "best_triton_pos": 1, "best_triton_time": 0.40959998965263367, "best_triton_kernel": "triton_convolution2d_6702", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x640x16x16, 1280x640x3x3)
strides: [163840, 256, 16, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.1751 ms 100.0%
triton_convolution2d_6702 0.4096 ms 42.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6697 0.5110 ms 34.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6703 0.6687 ms 26.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6699 0.6871 ms 25.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6705 0.8110 ms 21.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6698 0.9974 ms 17.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_6706 1.0936 ms 16.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6711 1.1366 ms 15.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6707 1.1397 ms 15.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.7097 seconds and 0.0003 seconds precompiling for 18 choices
E0801 17:12:29.569000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:29.569000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:29.569000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:30.162000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:30.162000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:30.162000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_6738", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8", "best_time": 0.01945599913597107, "best_triton_pos": 0}
AUTOTUNE convolution(6x640x16x16, 1280x640x1x1)
strides: [163840, 256, 16, 1], [640, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_6738 0.0195 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_6740 0.0225 ms 86.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_6739 0.0236 ms 82.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_6743 0.0256 ms 76.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0328 ms 59.4%
triton_convolution2d_6746 0.0348 ms 55.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6747 0.0358 ms 54.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6735 0.0379 ms 51.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_6737 0.0389 ms 50.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_6741 0.0389 ms 50.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.1925 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.3142400085926056, "best_triton_pos": 1, "best_triton_time": 0.7751680016517639, "best_triton_kernel": "triton_convolution2d_6754", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x1280x16x16, 1280x1280x3x3)
strides: [327680, 256, 16, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.3142 ms 100.0%
triton_convolution2d_6754 0.7752 ms 40.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6749 0.9851 ms 31.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_6755 1.2585 ms 25.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_6751 1.2749 ms 24.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6757 1.5575 ms 20.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_6750 2.0244 ms 15.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_6758 2.1115 ms 14.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6759 2.1637 ms 14.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_6763 2.2170 ms 14.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.7098 seconds and 0.0003 seconds precompiling for 18 choices
E0801 17:12:44.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:44.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:44.897000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:46.707000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:46.707000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:46.707000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:48.764000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:48.764000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:48.764000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:49.501000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:49.501000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:49.501000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6776", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.03174399957060814, "best_triton_pos": 0}
AUTOTUNE mm(1536x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
triton_mm_6776 0.0317 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6782 0.0328 ms 96.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6772 0.0420 ms 75.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6775 0.0522 ms 60.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6779 0.0543 ms 58.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6769 0.0584 ms 54.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6767 0.0635 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6768 0.0645 ms 49.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
mm 0.0655 ms 48.4%
triton_mm_6766 0.0686 ms 46.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.4528 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:12:52.576000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:52.576000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:52.576000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:54.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:54.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:54.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:56.763000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:56.763000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:56.763000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:12:57.565000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:12:57.565000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:12:57.565000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6856", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.03174399957060814, "best_triton_pos": 0}
AUTOTUNE addmm(1536x1280, 1536x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6856 0.0317 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6862 0.0338 ms 93.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6852 0.0410 ms 77.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6855 0.0522 ms 60.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6859 0.0553 ms 57.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6849 0.0614 ms 51.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6847 0.0635 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6848 0.0645 ms 49.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6846 0.0686 ms 46.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_6854 0.0707 ms 44.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.0598 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:13:00.283000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:00.283000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:00.283000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:01.962000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:04.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:04.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:04.262000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:04.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:04.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:04.908000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.018432000651955605, "best_triton_pos": 1, "best_triton_time": 0.020479999482631683, "best_triton_kernel": "triton_mm_6886", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4"}
AUTOTUNE mm(462x1024, 1024x1280)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
mm 0.0184 ms 100.0%
triton_mm_6886 0.0205 ms 90.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_6892 0.0215 ms 85.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_6889 0.0225 ms 81.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_6896 0.0236 ms 78.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6887 0.0266 ms 69.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_6895 0.0276 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6902 0.0276 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6899 0.0317 ms 58.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6888 0.0328 ms 56.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.3400 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:13:08.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:08.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:08.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:10.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:10.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:10.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:12.830000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:12.830000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:12.830000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:13.623000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:13.623000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:13.623000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_6962", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.20582400262355804, "best_triton_pos": 0}
AUTOTUNE addmm(1536x10240, 1536x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_6962 0.2058 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6961 0.2089 ms 98.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6956 0.2273 ms 90.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6958 0.2386 ms 86.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.2407 ms 85.5%
triton_mm_6960 0.2447 ms 84.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_6955 0.2478 ms 83.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6954 0.2488 ms 82.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_6959 0.2519 ms 81.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_6952 0.2734 ms 75.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.7093 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.18329599499702454, "best_triton_pos": 1, "best_triton_time": 0.8120319843292236, "best_triton_kernel": "triton_convolution2d_7301", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x1280x16x16, 1280x1280x3x3)
strides: [327680, 256, 16, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.1833 ms 100.0%
triton_convolution2d_7301 0.8120 ms 22.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7306 1.0220 ms 17.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7305 1.3066 ms 14.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7300 1.4633 ms 12.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7304 1.4776 ms 12.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_7303 1.6998 ms 10.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7302 1.9098 ms 9.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7309 2.1402 ms 8.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7310 2.1432 ms 8.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.9880 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:13:28.451000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:28.451000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:28.451000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:30.230000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:30.230000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:30.230000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:32.235000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:32.235000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:32.235000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:32.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:32.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:32.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.015359999611973763, "best_triton_pos": 1, "best_triton_time": 0.01945599913597107, "best_triton_kernel": "triton_mm_7474", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4"}
AUTOTUNE mm(384x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
mm 0.0154 ms 100.0%
triton_mm_7474 0.0195 ms 78.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_7471 0.0236 ms 65.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_7477 0.0236 ms 65.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_7472 0.0256 ms 60.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7481 0.0276 ms 55.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7480 0.0307 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_7484 0.0316 ms 48.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_7473 0.0317 ms 48.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7487 0.0317 ms 48.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 7.3051 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:13:35.959000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:35.959000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:35.959000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:37.608000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:37.608000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:37.608000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:40.057000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:40.057000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:40.057000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:40.855000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:40.855000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:40.855000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "addmm", "best_time": 0.018432000651955605, "best_triton_pos": 1, "best_triton_time": 0.020479999482631683, "best_triton_kernel": "triton_mm_7554", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4"}
AUTOTUNE addmm(384x1280, 384x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
addmm 0.0184 ms 100.0%
triton_mm_7554 0.0205 ms 90.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_7557 0.0246 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_7551 0.0256 ms 72.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_7561 0.0297 ms 62.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7552 0.0348 ms 52.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7567 0.0348 ms 52.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7553 0.0358 ms 51.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_7560 0.0369 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_7564 0.0389 ms 47.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.8761 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:13:43.932000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:43.932000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:43.932000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:46.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:46.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:46.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:48.546000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:48.546000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:48.546000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:13:49.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:13:49.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:13:49.334000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_7661", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.07680000364780426, "best_triton_pos": 0}
AUTOTUNE addmm(384x10240, 384x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_7661 0.0768 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7667 0.0788 ms 97.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7666 0.0819 ms 93.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7664 0.0870 ms 88.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_7660 0.0881 ms 87.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
addmm 0.0891 ms 86.2%
triton_mm_7657 0.0922 ms 83.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_7663 0.0922 ms 83.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7659 0.1034 ms 74.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_7665 0.1065 ms 72.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.4653 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.3328000009059906, "best_triton_pos": 1, "best_triton_time": 1.1612160205841064, "best_triton_kernel": "triton_convolution2d_7771", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4"}
AUTOTUNE convolution(6x2560x8x8, 1280x2560x3x3)
strides: [163840, 64, 8, 1], [23040, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.3328 ms 100.0%
triton_convolution2d_7771 1.1612 ms 28.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7764 1.2431 ms 26.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=512, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_7766 1.2452 ms 26.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7768 1.4162 ms 23.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7769 1.8299 ms 18.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_7763 1.8565 ms 17.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7765 1.9743 ms 16.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7776 3.4079 ms 9.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7770 3.4611 ms 9.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 10.4709 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:14:04.529000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:04.529000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:04.529000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:14:04.994000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:04.994000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:04.994000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_7805", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4", "best_time": 0.03481600061058998, "best_triton_pos": 0}
AUTOTUNE convolution(6x2560x8x8, 1280x2560x1x1)
strides: [163840, 64, 8, 1], [2560, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_7805 0.0348 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_7806 0.0348 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_7804 0.0614 ms 56.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_7809 0.0655 ms 53.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0840 ms 41.5%
triton_convolution2d_7801 0.0952 ms 36.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_7812 0.0963 ms 36.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_7813 0.1024 ms 34.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_7803 0.1055 ms 33.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_7800 0.1065 ms 32.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.3154 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.6082559823989868, "best_triton_pos": 1, "best_triton_time": 1.5134719610214233, "best_triton_kernel": "triton_convolution2d_7992", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x2560x16x16, 1280x2560x3x3)
strides: [655360, 256, 16, 1], [23040, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.6083 ms 100.0%
triton_convolution2d_7992 1.5135 ms 40.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7987 1.9251 ms 31.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_7993 2.5426 ms 23.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_7989 2.5846 ms 23.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7995 2.6051 ms 23.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_7988 3.9117 ms 15.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_7996 4.2342 ms 14.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_7997 4.3172 ms 14.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8001 4.4083 ms 13.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.8611 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:14:21.872000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:21.872000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:21.872000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:14:22.339000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:22.339000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:22.339000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_8028", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8", "best_time": 0.06451199948787689, "best_triton_pos": 0}
AUTOTUNE convolution(6x2560x16x16, 1280x2560x1x1)
strides: [655360, 256, 16, 1], [2560, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_8028 0.0645 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_8030 0.0707 ms 91.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_8029 0.0717 ms 90.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_8033 0.0748 ms 86.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0861 ms 74.9%
triton_convolution2d_8025 0.1004 ms 64.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_8024 0.1055 ms 61.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8027 0.1065 ms 60.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8037 0.1065 ms 60.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8036 0.1075 ms 60.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.4703 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.4700160026550293, "best_triton_pos": 1, "best_triton_time": 1.1520320177078247, "best_triton_kernel": "triton_convolution2d_8610", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x1920x16x16, 1280x1920x3x3)
strides: [491520, 256, 16, 1], [17280, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.4700 ms 100.0%
triton_convolution2d_8610 1.1520 ms 40.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8605 1.4244 ms 33.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_8611 1.8729 ms 25.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_8607 1.8893 ms 24.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8613 2.0388 ms 23.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_8606 2.9420 ms 16.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_8614 3.1140 ms 15.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8615 3.2840 ms 14.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8619 3.3229 ms 14.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.7019 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:14:39.009000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:39.009000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:39.009000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:14:39.460000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:14:39.460000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:14:39.460000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_8646", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8", "best_time": 0.04915200173854828, "best_triton_pos": 0}
AUTOTUNE convolution(6x1920x16x16, 1280x1920x1x1)
strides: [491520, 256, 16, 1], [1920, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_8646 0.0492 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_8648 0.0553 ms 88.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_8647 0.0573 ms 85.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_8651 0.0614 ms 80.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0748 ms 65.8%
triton_convolution2d_8654 0.0829 ms 59.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8655 0.0829 ms 59.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8642 0.0870 ms 56.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8643 0.0870 ms 56.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_8645 0.0901 ms 54.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.3467 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.8980479836463928, "best_triton_pos": 1, "best_triton_time": 2.9757440090179443, "best_triton_kernel": "triton_convolution2d_8914", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x1280x32x32, 1280x1280x3x3)
strides: [1310720, 1024, 32, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.8980 ms 100.0%
triton_convolution2d_8914 2.9757 ms 30.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_8919 3.1171 ms 28.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8915 4.5373 ms 19.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_8923 4.6101 ms 19.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8916 4.6612 ms 19.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8920 5.1548 ms 17.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_8928 5.4108 ms 16.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8922 6.4973 ms 13.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_8924 8.0517 ms 11.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.1547 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.7403519749641418, "best_triton_pos": 1, "best_triton_time": 2.279423952102661, "best_triton_kernel": "triton_convolution2d_8936", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x1920x32x32, 640x1920x3x3)
strides: [1966080, 1024, 32, 1], [17280, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.7404 ms 100.0%
triton_convolution2d_8936 2.2794 ms 32.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8931 2.2866 ms 32.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_8932 3.4212 ms 21.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_8940 3.4806 ms 21.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8933 3.6035 ms 20.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8937 3.6925 ms 20.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_8945 4.0960 ms 18.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_8939 4.1738 ms 17.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_8941 6.1082 ms 12.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.0472 seconds and 0.0003 seconds precompiling for 18 choices
E0801 17:15:07.689000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:07.689000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:07.689000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:15:08.142000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:08.142000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:08.142000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.09728000313043594, "best_triton_pos": 1, "best_triton_time": 0.09728000313043594, "best_triton_kernel": "triton_convolution2d_8980", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x1920x32x32, 640x1920x1x1)
strides: [1966080, 1024, 32, 1], [1920, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.0973 ms 100.0%
triton_convolution2d_8980 0.0973 ms 100.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8972 0.1004 ms 96.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_8975 0.1004 ms 96.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8971 0.1044 ms 93.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8968 0.1065 ms 91.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8977 0.1137 ms 85.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_8974 0.1219 ms 79.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_8966 0.1249 ms 77.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_8973 0.1249 ms 77.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 6.4347 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.4915199875831604, "best_triton_pos": 1, "best_triton_time": 1.5421439409255981, "best_triton_kernel": "triton_convolution2d_9245", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x1280x32x32, 640x1280x3x3)
strides: [1310720, 1024, 32, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.4915 ms 100.0%
triton_convolution2d_9245 1.5421 ms 31.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9240 1.5821 ms 31.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_9241 2.2886 ms 21.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_9249 2.3552 ms 20.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9242 2.4617 ms 20.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9246 2.4852 ms 19.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_9254 2.7218 ms 18.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9248 3.2461 ms 15.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_9250 4.1103 ms 12.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.7713 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:15:24.833000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:24.833000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:24.833000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:15:25.287000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:25.287000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:25.287000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_9289", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8", "best_time": 0.06860800087451935, "best_triton_pos": 0}
AUTOTUNE convolution(6x1280x32x32, 640x1280x1x1)
strides: [1310720, 1024, 32, 1], [1280, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_9289 0.0686 ms 100.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0727 ms 94.4%
triton_convolution2d_9284 0.0737 ms 93.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9281 0.0758 ms 90.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_9280 0.0768 ms 89.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9277 0.0809 ms 84.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9286 0.0810 ms 84.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9283 0.0881 ms 77.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_9282 0.0901 ms 76.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_9275 0.0922 ms 74.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 6.3590 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.39929598569869995, "best_triton_pos": 1, "best_triton_time": 1.1581439971923828, "best_triton_kernel": "triton_convolution2d_9554", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x960x32x32, 640x960x3x3)
strides: [983040, 1024, 32, 1], [8640, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.3993 ms 100.0%
triton_convolution2d_9554 1.1581 ms 34.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9549 1.1694 ms 34.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_9550 1.7265 ms 23.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_9558 1.7736 ms 22.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9551 1.8217 ms 21.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9555 1.8381 ms 21.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_9563 2.1248 ms 18.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9557 2.4381 ms 16.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_9559 3.1191 ms 12.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 10.6409 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:15:44.129000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:44.129000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:44.129000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:15:44.636000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:15:44.636000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:15:44.636000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_9598", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8", "best_time": 0.05427199974656105, "best_triton_pos": 0}
AUTOTUNE convolution(6x960x32x32, 640x960x1x1)
strides: [983040, 1024, 32, 1], [960, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_9598 0.0543 ms 100.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9590 0.0573 ms 94.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_9593 0.0594 ms 91.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.0604 ms 89.8%
triton_convolution2d_9589 0.0655 ms 82.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9591 0.0686 ms 79.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_9595 0.0707 ms 76.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9586 0.0717 ms 75.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9592 0.0717 ms 75.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_9584 0.0788 ms 68.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.7501 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.8775680065155029, "best_triton_pos": 1, "best_triton_time": 2.8282880783081055, "best_triton_kernel": "triton_convolution2d_9863", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x640x64x64, 640x640x3x3)
strides: [2621440, 4096, 64, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.8776 ms 100.0%
triton_convolution2d_9863 2.8283 ms 31.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9858 3.0331 ms 28.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_9859 3.9670 ms 22.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_9867 4.7902 ms 18.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9864 4.7985 ms 18.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_9860 5.0391 ms 17.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9872 5.5695 ms 15.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9866 6.0375 ms 14.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_9868 8.1009 ms 10.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.5090 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.7690240144729614, "best_triton_pos": 1, "best_triton_time": 2.1729280948638916, "best_triton_kernel": "triton_convolution2d_9875", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x960x64x64, 320x960x3x3)
strides: [3932160, 4096, 64, 1], [8640, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.7690 ms 100.0%
triton_convolution2d_9875 2.1729 ms 35.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_9880 2.3521 ms 32.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9876 3.3453 ms 23.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_9877 3.6127 ms 21.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9881 3.8072 ms 20.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_9884 3.8216 ms 20.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9889 4.9142 ms 15.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_9883 4.9152 ms 15.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_9887 6.0959 ms 12.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.1779 seconds and 0.0003 seconds precompiling for 18 choices
E0801 17:16:13.594000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:16:13.594000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:16:13.594000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:16:14.113000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:16:14.113000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:16:14.113000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.11776000261306763, "best_triton_pos": 1, "best_triton_time": 0.1536639928817749, "best_triton_kernel": "triton_convolution2d_9915", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(6x960x64x64, 320x960x1x1)
strides: [3932160, 4096, 64, 1], [960, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.1178 ms 100.0%
triton_convolution2d_9915 0.1537 ms 76.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9910 0.1587 ms 74.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_9925 0.1628 ms 72.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9920 0.1638 ms 71.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9916 0.1690 ms 69.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_9912 0.1700 ms 69.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9918 0.1700 ms 69.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_9914 0.1782 ms 66.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_9921 0.1782 ms 66.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.7891 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.5263360142707825, "best_triton_pos": 1, "best_triton_time": 1.5144959688186646, "best_triton_kernel": "triton_convolution2d_10184", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(6x640x64x64, 320x640x3x3)
strides: [2621440, 4096, 64, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.5263 ms 100.0%
triton_convolution2d_10184 1.5145 ms 34.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_10189 1.6026 ms 32.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10185 2.2436 ms 23.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_10190 2.5774 ms 20.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_10186 2.6348 ms 20.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10193 2.6552 ms 19.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10198 3.1826 ms 16.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10192 3.2768 ms 16.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_10195 4.3090 ms 12.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 11.4234 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:16:32.503000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:16:32.503000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:16:32.503000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
E0801 17:16:33.141000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Runtime error during autotuning:
E0801 17:16:33.141000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:16:33.141000 2856711 torch/_inductor/select_algorithm.py:4888] [0/1] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.08806400001049042, "best_triton_pos": 1, "best_triton_time": 0.10751999914646149, "best_triton_kernel": "triton_convolution2d_10227", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4"}
AUTOTUNE convolution(6x640x64x64, 320x640x1x1)
strides: [2621440, 4096, 64, 1], [640, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.0881 ms 100.0%
triton_convolution2d_10227 0.1075 ms 81.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_10219 0.1096 ms 80.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_10224 0.1096 ms 80.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_10225 0.1126 ms 78.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_10234 0.1126 ms 78.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_10221 0.1137 ms 77.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_10226 0.1147 ms 76.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_10229 0.1147 ms 76.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_10233 0.1189 ms 74.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.5638 seconds and 0.0003 seconds precompiling for 19 choices
Out[20]:
Batch B=2 Summary:
Total Batch Time : 1059.22 ms
Effective Per-Image Time : 529.61 ms
Peak VRAM Memory : 3.69 GB
Effective Speedup Factor : 3.19x Faster!
Testing Micro-Batch Size B=4...
Batch B=4 Summary:
Total Batch Time : 2060.32 ms
Effective Per-Image Time : 515.08 ms
Peak VRAM Memory : 4.77 GB
Effective Speedup Factor : 3.28x Faster!
Testing Micro-Batch Size B=8...
Batch B=8 Summary:
Total Batch Time : 4231.64 ms
Effective Per-Image Time : 528.96 ms
Peak VRAM Memory : 6.93 GB
Effective Speedup Factor : 3.20x Faster!
Testing Micro-Batch Size B=16...
Batch B=16 Summary:
Total Batch Time : 8599.95 ms
Effective Per-Image Time : 537.50 ms
Peak VRAM Memory : 11.25 GB
Effective Speedup Factor : 3.15x Faster!
Testing Micro-Batch Size B=32...
Out[20]:
E0801 17:19:29.364000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:29.364000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:29.364000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:29.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:29.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:29.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:29.803000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:29.803000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:29.803000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:29.804000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:29.804000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:29.804000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11139", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 2.1125121116638184, "best_triton_pos": 0}
AUTOTUNE addmm(393216x320, 393216x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11139 2.1125 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11136 2.3419 ms 90.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11133 2.3941 ms 88.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11129 2.4259 ms 87.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11138 2.4371 ms 86.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11135 2.4842 ms 85.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11131 2.5610 ms 82.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11132 2.6051 ms 81.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11128 2.6348 ms 80.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11137 2.6583 ms 79.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.0163 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:19:30.845000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:30.845000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:30.845000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:31.005000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:31.005000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:31.005000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:31.264000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:31.264000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:31.264000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:31.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:31.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:31.265000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11761", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 1.6629760265350342, "best_triton_pos": 0}
AUTOTUNE addmm(98304x640, 98304x2560, 2560x640)
strides: [0, 1], [2560, 1], [1, 2560]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11761 1.6630 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11760 1.7080 ms 97.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11759 1.8964 ms 87.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_11755 1.8995 ms 87.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 2.0070 ms 82.9%
triton_mm_11753 2.0163 ms 82.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11757 2.0634 ms 80.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11758 2.1463 ms 77.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11754 2.1740 ms 76.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11751 2.3378 ms 71.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9803 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:19:32.246000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:32.246000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:32.246000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:32.402000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:32.402000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:32.402000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:32.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:32.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:32.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:32.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:32.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:32.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_12383", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 1.5544320344924927, "best_triton_pos": 0}
AUTOTUNE addmm(24576x1280, 24576x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_12383 1.5544 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12382 1.6128 ms 96.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 1.6640 ms 93.4%
triton_mm_12381 1.7510 ms 88.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_12377 1.8022 ms 86.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12379 1.9118 ms 81.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12375 1.9190 ms 81.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12380 2.0029 ms 77.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12376 2.0357 ms 76.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12373 2.2272 ms 69.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9623 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.8878080248832703, "best_triton_pos": 1, "best_triton_time": 2.9716479778289795, "best_triton_kernel": "triton_convolution2d_12809", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x1280x8x8, 1280x1280x3x3)
strides: [81920, 64, 8, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 0.8878 ms 100.0%
triton_convolution2d_12809 2.9716 ms 29.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_12814 3.1427 ms 28.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12810 4.4780 ms 19.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_12818 4.6469 ms 19.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12811 4.7104 ms 18.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12815 5.0811 ms 17.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12823 5.3770 ms 16.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12817 6.0212 ms 14.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_12819 8.1500 ms 10.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.8503 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:19:35.564000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:35.564000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:35.564000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:35.704000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:35.704000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:35.704000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:35.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:35.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:35.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:35.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:35.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:35.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "addmm", "best_time": 0.40038400888442993, "best_triton_pos": 1, "best_triton_time": 0.4362240135669708, "best_triton_kernel": "triton_mm_13096", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4"}
AUTOTUNE addmm(6144x1280, 6144x5120, 5120x1280)
strides: [0, 1], [5120, 1], [1, 5120]
dtypes: torch.float16, torch.float16, torch.float16
addmm 0.4004 ms 100.0%
triton_mm_13096 0.4362 ms 91.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13095 0.4454 ms 89.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13090 0.4751 ms 84.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13094 0.4884 ms 82.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_13088 0.5038 ms 79.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13092 0.5048 ms 79.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13089 0.5325 ms 75.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_13093 0.5345 ms 74.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_13086 0.5816 ms 68.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8029 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 16, "num_triton_choices": 15, "best_kernel": "triton_convolution2d_16245", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8", "best_time": 0.8243200182914734, "best_triton_pos": 0}
AUTOTUNE convolution(96x320x64x64, 4x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_16245 0.8243 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_16242 0.8847 ms 93.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_16246 0.8847 ms 93.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_16241 0.9206 ms 89.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_16244 0.9257 ms 89.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_16239 0.9564 ms 86.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_16250 0.9574 ms 86.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_16240 1.0465 ms 78.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_16243 1.0762 ms 76.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_16236 1.0813 ms 76.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.7727 seconds and 0.0002 seconds precompiling for 16 choices
E0801 17:19:39.610000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:39.610000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:39.610000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:41.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:41.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:41.377000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:44.309000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:44.309000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:44.309000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:44.850000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:44.850000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:44.850000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_10849", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.008191999979317188, "best_triton_pos": 0}
AUTOTUNE addmm(96x1280, 96x320, 320x1280)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_10849 0.0082 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_10852 0.0092 ms 88.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_10846 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=False, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_10856 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10858 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10859 0.0102 ms 80.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_10862 0.0103 ms 79.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10847 0.0113 ms 72.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
addmm 0.0123 ms 66.7%
triton_mm_10848 0.0123 ms 66.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.7256 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:19:48.204000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:48.204000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:48.204000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:49.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:49.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:49.879000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:52.375000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:52.375000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:52.375000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:19:53.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:19:53.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:19:53.188000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_10869", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.012223999947309494, "best_triton_pos": 0}
AUTOTUNE addmm(96x1280, 96x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_10869 0.0122 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_10866 0.0225 ms 54.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_10872 0.0236 ms 51.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
addmm 0.0246 ms 49.7%
triton_mm_10882 0.0246 ms 49.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10876 0.0256 ms 47.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10867 0.0266 ms 45.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_10879 0.0277 ms 44.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_10878 0.0308 ms 39.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10868 0.0317 ms 38.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 7.9496 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 13, "num_triton_choices": 12, "best_kernel": "triton_convolution2d_10818", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8", "best_time": 0.37273600697517395, "best_triton_pos": 0}
AUTOTUNE convolution(96x4x64x64, 320x4x3x3)
strides: [16384, 4096, 64, 1], [36, 9, 3, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_10818 0.3727 ms 100.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_10826 0.3983 ms 93.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10817 0.4096 ms 91.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_10819 0.4106 ms 90.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10823 0.4116 ms 90.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_10822 0.4168 ms 89.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10825 0.4219 ms 88.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
convolution 0.4291 ms 86.9%
triton_convolution2d_10827 0.4567 ms 81.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10820 0.4772 ms 78.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 0.6643 seconds and 0.0002 seconds precompiling for 13 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 5.025792121887207, "best_triton_pos": 1, "best_triton_time": 11.517951965332031, "best_triton_kernel": "triton_convolution2d_10829", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x320x64x64, 320x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 5.0258 ms 100.0%
triton_convolution2d_10829 11.5180 ms 43.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_10834 11.6716 ms 43.1% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10838 17.2329 ms 29.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10830 18.7464 ms 26.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_10831 19.1590 ms 26.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10835 20.1697 ms 24.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_10843 22.2659 ms 22.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_10837 24.9713 ms 20.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_10839 30.4497 ms 16.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 4.1639 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:20:17.824000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:17.824000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:17.824000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:19.526000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:19.526000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:19.526000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:22.097000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:22.097000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:22.097000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:22.929000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:22.929000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:22.929000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_10889", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.010143999941647053, "best_triton_pos": 0}
AUTOTUNE addmm(96x320, 96x1280, 1280x320)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_10889 0.0101 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_10886 0.0133 ms 76.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_10892 0.0143 ms 70.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_10888 0.0164 ms 61.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_10896 0.0164 ms 61.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10887 0.0173 ms 58.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
addmm 0.0174 ms 58.3%
triton_mm_10899 0.0174 ms 58.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_10895 0.0203 ms 49.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_10898 0.0215 ms 47.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.1101 seconds and 0.0003 seconds precompiling for 21 choices
E0801 17:20:23.384000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:23.384000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:23.384000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:23.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:23.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:23.530000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:23.773000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:23.773000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:23.773000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:23.774000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:23.774000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:23.774000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_10933", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.7895039916038513, "best_triton_pos": 0}
AUTOTUNE mm(393216x320, 320x320)
strides: [320, 1], [1, 320]
dtypes: torch.float16, torch.float16
triton_mm_10933 0.7895 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10929 0.7924 ms 99.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_10939 0.7936 ms 99.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10932 0.8346 ms 94.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_10936 0.8397 ms 94.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
mm 0.8428 ms 93.7%
triton_mm_10938 0.8694 ms 90.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10931 0.8755 ms 90.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_10928 0.8868 ms 89.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_10937 0.8878 ms 88.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8409 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:24.234000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:24.234000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:24.234000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:24.380000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:24.380000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:24.380000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:24.621000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:24.621000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:24.621000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:24.622000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:24.622000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:24.622000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11013", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.7915520071983337, "best_triton_pos": 0}
AUTOTUNE addmm(393216x320, 393216x320, 320x320)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11013 0.7916 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11009 0.7926 ms 99.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11019 0.7936 ms 99.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11012 0.8325 ms 95.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11016 0.8397 ms 94.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11011 0.8520 ms 92.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11015 0.8663 ms 91.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11018 0.8714 ms 90.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11008 0.8796 ms 90.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11017 0.9851 ms 80.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8415 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:24.949000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:24.949000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:24.949000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:25.050000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:25.050000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:25.050000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:25.220000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:25.220000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:25.220000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:25.221000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:25.221000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:25.221000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11049", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8", "best_time": 0.04505600035190582, "best_triton_pos": 0}
AUTOTUNE mm(7392x1024, 1024x320)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_11049 0.0451 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11055 0.0461 ms 97.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11058 0.0502 ms 89.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11053 0.0512 ms 88.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11051 0.0522 ms 86.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11056 0.0522 ms 86.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
mm 0.0543 ms 83.0%
triton_mm_11059 0.0553 ms 81.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11045 0.0614 ms 73.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_11052 0.0614 ms 73.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.5960 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:26.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:26.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:26.081000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:26.300000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:26.300000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:26.300000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:26.663000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:26.663000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:26.663000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:26.666000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:26.666000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:26.666000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11118", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 3.8830080032348633, "best_triton_pos": 0}
AUTOTUNE addmm(393216x2560, 393216x320, 320x2560)
strides: [0, 1], [320, 1], [1, 320]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11118 3.8830 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11119 4.2875 ms 90.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11115 4.5875 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11111 4.5906 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11113 4.6489 ms 83.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11112 4.8394 ms 80.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11116 4.8753 ms 79.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11108 5.2521 ms 73.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11109 5.5910 ms 69.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11117 5.6648 ms 68.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.4409 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 2.2231040000915527, "best_triton_pos": 1, "best_triton_time": 4.9244160652160645, "best_triton_kernel": "triton_convolution2d_11457", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x320x64x64, 320x320x3x3)
strides: [1310720, 4096, 64, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 2.2231 ms 100.0%
triton_convolution2d_11457 4.9244 ms 45.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_11459 6.3232 ms 35.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11462 6.3519 ms 35.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11466 6.4461 ms 34.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11465 7.0717 ms 31.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_11463 8.2309 ms 27.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_11467 8.7071 ms 25.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11471 8.9446 ms 24.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11468 9.0399 ms 24.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.8174 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 2.0695040225982666, "best_triton_pos": 1, "best_triton_time": 5.607423782348633, "best_triton_kernel": "triton_convolution2d_11479", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x320x32x32, 640x320x3x3)
strides: [327680, 1024, 32, 1], [2880, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 2.0695 ms 100.0%
triton_convolution2d_11479 5.6074 ms 36.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11474 5.6658 ms 36.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_11475 8.5074 ms 24.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_11483 8.6047 ms 24.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11476 9.7321 ms 21.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11480 9.7812 ms 21.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_11488 10.4632 ms 19.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11482 12.7130 ms 16.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_11484 15.3631 ms 13.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 2.3040 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:20:33.920000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:33.920000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:33.920000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:35.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:35.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:35.645000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:38.232000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:38.232000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:38.232000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:39.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:39.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:39.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11494", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4", "best_time": 0.009216000325977802, "best_triton_pos": 0}
AUTOTUNE addmm(96x640, 96x1280, 1280x640)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11494 0.0092 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=4
triton_mm_11491 0.0184 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11497 0.0184 ms 50.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11501 0.0195 ms 47.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.0204 ms 45.1%
triton_mm_11504 0.0205 ms 45.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11492 0.0225 ms 40.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_11493 0.0236 ms 39.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=32, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=5, num_warps=8
triton_mm_11503 0.0236 ms 39.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11507 0.0246 ms 37.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
SingleProcess AUTOTUNE benchmarking takes 8.2380 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:39.800000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:39.800000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:39.800000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:39.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:39.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:39.801000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.2949120104312897, "best_triton_pos": 1, "best_triton_time": 0.4904960095882416, "best_triton_kernel": "triton_convolution2d_11510", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x320x32x32, 640x320x1x1)
strides: [327680, 1024, 32, 1], [320, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.2949 ms 100.0%
triton_convolution2d_11510 0.4905 ms 60.1% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_11521 0.4956 ms 59.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11526 0.5048 ms 58.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11515 0.5110 ms 57.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11513 0.5949 ms 49.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11520 0.6144 ms 48.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11525 0.6257 ms 47.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_11514 0.6287 ms 46.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_11522 0.6298 ms 46.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8367 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 3.6464641094207764, "best_triton_pos": 1, "best_triton_time": 11.2424955368042, "best_triton_kernel": "triton_convolution2d_11533", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x640x32x32, 640x640x3x3)
strides: [655360, 1024, 32, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 3.6465 ms 100.0%
triton_convolution2d_11533 11.2425 ms 32.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11528 11.7883 ms 30.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_11537 17.0230 ms 21.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11530 18.6398 ms 19.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11534 19.1734 ms 19.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_11542 20.2067 ms 18.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11536 25.1249 ms 14.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_11538 30.4835 ms 12.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_11543 31.9580 ms 11.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 4.3257 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:20:44.673000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:44.673000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:44.673000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:44.817000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:44.817000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:44.817000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:45.058000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.058000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.058000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:45.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.059000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11561", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.4505600035190582, "best_triton_pos": 0}
AUTOTUNE mm(98304x640, 640x640)
strides: [640, 1], [1, 640]
dtypes: torch.float16, torch.float16
triton_mm_11561 0.4506 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11560 0.4782 ms 94.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
mm 0.5140 ms 87.6%
triton_mm_11555 0.5243 ms 85.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11559 0.5274 ms 85.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_11553 0.5468 ms 82.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11557 0.5540 ms 81.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11554 0.5622 ms 80.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11558 0.5693 ms 79.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11551 0.6267 ms 71.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8327 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:45.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:45.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.660000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:45.899000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.899000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.899000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:45.900000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:45.900000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:45.900000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11641", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.46694400906562805, "best_triton_pos": 0}
AUTOTUNE addmm(98304x640, 98304x640, 640x640)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11641 0.4669 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11640 0.4823 ms 96.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11635 0.5325 ms 87.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11633 0.5519 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11637 0.5560 ms 84.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11634 0.5755 ms 81.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11638 0.5806 ms 80.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11631 0.6349 ms 73.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11630 0.6492 ms 71.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11639 0.6625 ms 70.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8346 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:46.263000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:46.263000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:46.263000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:46.370000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:46.370000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:46.370000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:46.550000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:46.550000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:46.550000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:46.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:46.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:46.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11681", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.06758400052785873, "best_triton_pos": 0}
AUTOTUNE mm(7392x1024, 1024x640)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_11681 0.0676 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11675 0.0686 ms 98.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11680 0.0707 ms 95.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11671 0.0768 ms 88.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_11674 0.0778 ms 86.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11677 0.0787 ms 85.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11673 0.0799 ms 84.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11678 0.0860 ms 78.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
mm 0.0870 ms 77.6%
triton_mm_11679 0.1044 ms 64.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.6480 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:47.349000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:47.349000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:47.349000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:47.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:47.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:47.551000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:47.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:47.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:47.886000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:47.888000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:47.888000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:47.888000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_11740", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 3.458048105239868, "best_triton_pos": 0}
AUTOTUNE addmm(98304x5120, 98304x640, 640x5120)
strides: [0, 1], [640, 1], [1, 640]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_11740 3.4580 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11741 3.7366 ms 92.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11735 4.0643 ms 85.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11739 4.1103 ms 84.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_11737 4.1349 ms 83.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11733 4.1789 ms 82.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_11734 4.2906 ms 80.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11738 4.4995 ms 76.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_11730 4.7514 ms 72.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=4
triton_mm_11731 4.8712 ms 71.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.3323 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.4766080379486084, "best_triton_pos": 1, "best_triton_time": 5.981184005737305, "best_triton_kernel": "triton_convolution2d_12079", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x640x32x32, 640x640x3x3)
strides: [655360, 1024, 32, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 1.4766 ms 100.0%
triton_convolution2d_12079 5.9812 ms 24.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_12081 5.9945 ms 24.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12084 6.7840 ms 21.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12085 6.9171 ms 21.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12087 7.5612 ms 19.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_12089 8.1398 ms 18.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12091 9.5130 ms 15.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12090 9.5693 ms 15.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12088 9.7270 ms 15.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.7392 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.7018879652023315, "best_triton_pos": 1, "best_triton_time": 5.7139201164245605, "best_triton_kernel": "triton_convolution2d_12101", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x640x16x16, 1280x640x3x3)
strides: [163840, 256, 16, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 1.7019 ms 100.0%
triton_convolution2d_12101 5.7139 ms 29.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12096 5.7149 ms 29.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_12097 7.8838 ms 21.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_12105 8.6917 ms 19.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12098 9.7311 ms 17.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12102 9.7567 ms 17.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12110 10.0147 ms 17.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12104 12.1672 ms 14.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_12106 15.7071 ms 10.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 2.3448 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:20:52.786000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:52.786000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:52.786000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:52.787000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:52.787000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:52.787000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_12143", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8", "best_time": 0.23654399812221527, "best_triton_pos": 0}
AUTOTUNE convolution(96x640x16x16, 1280x640x1x1)
strides: [163840, 256, 16, 1], [640, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_12143 0.2365 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
convolution 0.2427 ms 97.5%
triton_convolution2d_12135 0.2478 ms 95.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_12139 0.2488 ms 95.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_12142 0.2488 ms 95.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_12148 0.2488 ms 95.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_12147 0.2559 ms 92.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_12133 0.2580 ms 91.7% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_12138 0.2601 ms 90.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_12137 0.2724 ms 86.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8826 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 3.1651840209960938, "best_triton_pos": 1, "best_triton_time": 11.4268159866333, "best_triton_kernel": "triton_convolution2d_12155", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1280x16x16, 1280x1280x3x3)
strides: [327680, 256, 16, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 3.1652 ms 100.0%
triton_convolution2d_12155 11.4268 ms 27.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12150 11.5528 ms 27.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_12151 16.3635 ms 19.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_12159 17.2902 ms 18.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12152 18.7791 ms 16.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12156 19.4253 ms 16.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12164 20.0387 ms 15.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12158 24.5596 ms 12.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_12160 31.1685 ms 10.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 4.1753 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:20:57.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:57.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:57.514000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:57.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:57.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:57.659000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:57.903000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:57.903000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:57.903000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:57.904000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:57.904000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:57.904000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.39526399970054626, "best_triton_pos": 1, "best_triton_time": 0.4188160002231598, "best_triton_kernel": "triton_mm_12183", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4"}
AUTOTUNE mm(24576x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
mm 0.3953 ms 100.0%
triton_mm_12183 0.4188 ms 94.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12182 0.4352 ms 90.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12181 0.4833 ms 81.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_12177 0.4844 ms 81.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12175 0.5048 ms 78.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12179 0.5171 ms 76.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12176 0.5274 ms 75.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12180 0.5294 ms 74.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12173 0.5847 ms 67.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8341 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:58.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:58.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:58.358000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:58.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:58.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:58.504000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:58.746000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:58.746000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:58.746000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:58.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:58.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:58.747000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_12263", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.42291200160980225, "best_triton_pos": 0}
AUTOTUNE addmm(24576x1280, 24576x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_12263 0.4229 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12262 0.4393 ms 96.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12261 0.4895 ms 86.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_12257 0.4915 ms 86.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12255 0.5100 ms 82.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12259 0.5192 ms 81.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.5263 ms 80.4%
triton_mm_12256 0.5325 ms 79.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12260 0.5386 ms 78.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12253 0.5847 ms 72.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8327 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:20:59.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:59.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:59.185000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:59.307000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:59.307000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:59.307000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:59.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:59.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:59.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:20:59.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:20:59.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:20:59.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_12303", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.10751999914646149, "best_triton_pos": 0}
AUTOTUNE mm(7392x1024, 1024x1280)
strides: [1024, 1], [1, 1024]
dtypes: torch.float16, torch.float16
triton_mm_12303 0.1075 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12302 0.1157 ms 92.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12297 0.1188 ms 90.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
mm 0.1280 ms 84.0%
triton_mm_12295 0.1341 ms 80.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12296 0.1352 ms 79.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12299 0.1352 ms 79.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12300 0.1382 ms 77.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12301 0.1444 ms 74.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_12293 0.1454 ms 73.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.7579 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:21:00.258000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:00.258000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:00.258000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:00.450000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:00.450000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:00.450000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:00.759000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:00.759000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:00.759000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:00.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:00.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:00.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_12362", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 3.1918718814849854, "best_triton_pos": 0}
AUTOTUNE addmm(24576x10240, 24576x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_12362 3.1919 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12363 3.2082 ms 99.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12361 3.3260 ms 96.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_12357 3.7079 ms 86.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12355 3.8062 ms 83.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12359 3.8431 ms 83.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12360 3.9291 ms 81.2% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12356 3.9414 ms 81.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
addmm 4.3448 ms 73.5%
triton_mm_12353 4.4370 ms 71.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.2471 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.0926079750061035, "best_triton_pos": 1, "best_triton_time": 6.57919979095459, "best_triton_kernel": "triton_convolution2d_12707", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8"}
AUTOTUNE convolution(96x1280x16x16, 1280x1280x3x3)
strides: [327680, 256, 16, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 1.0926 ms 100.0%
triton_convolution2d_12707 6.5792 ms 16.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_12703 6.6406 ms 16.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12706 6.8106 ms 16.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12701 7.0072 ms 15.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_12710 7.5766 ms 14.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12709 7.7230 ms 14.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_12715 7.9278 ms 13.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12711 7.9759 ms 13.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_12716 9.3348 ms 11.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=2, STRIDE_W=2, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.6926 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:21:02.956000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:02.956000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:02.956000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:03.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:03.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:03.078000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:03.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:03.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:03.280000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:03.281000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:03.281000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:03.281000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "mm", "best_time": 0.11468800157308578, "best_triton_pos": 1, "best_triton_time": 0.11878400295972824, "best_triton_kernel": "triton_mm_12895", "best_triton_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4"}
AUTOTUNE mm(6144x1280, 1280x1280)
strides: [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16
mm 0.1147 ms 100.0%
triton_mm_12895 0.1188 ms 96.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12896 0.1188 ms 96.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12890 0.1229 ms 93.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12888 0.1321 ms 86.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12892 0.1321 ms 86.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12889 0.1393 ms 82.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12893 0.1423 ms 80.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12886 0.1464 ms 78.3% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_12894 0.1495 ms 76.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.7629 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:21:03.734000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:03.734000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:03.734000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:03.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:03.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:03.858000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:04.062000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.062000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.062000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:04.063000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.063000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.063000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_12976", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.12070400267839432, "best_triton_pos": 0}
AUTOTUNE addmm(6144x1280, 6144x1280, 1280x1280)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_12976 0.1207 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12970 0.1239 ms 97.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12975 0.1260 ms 95.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12972 0.1341 ms 90.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
addmm 0.1362 ms 88.6%
triton_mm_12969 0.1393 ms 86.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12973 0.1454 ms 83.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_12968 0.1464 ms 82.4% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_12966 0.1526 ms 79.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
triton_mm_12974 0.1577 ms 76.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.7705 seconds and 0.0002 seconds precompiling for 21 choices
E0801 17:21:04.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.552000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:04.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.698000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:04.940000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.940000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.940000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:04.941000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:04.941000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 196608 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:04.941000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 21, "num_triton_choices": 20, "best_kernel": "triton_mm_13076", "best_kernel_desc": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4", "best_time": 0.8028159737586975, "best_triton_pos": 0}
AUTOTUNE addmm(6144x10240, 6144x1280, 1280x10240)
strides: [0, 1], [1280, 1], [1, 1280]
dtypes: torch.float16, torch.float16, torch.float16
triton_mm_13076 0.8028 ms 100.0% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13075 0.8202 ms 97.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13074 0.8765 ms 91.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=2, num_warps=8
triton_mm_13070 0.9247 ms 86.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13068 0.9667 ms 83.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13072 0.9728 ms 82.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=4
triton_mm_13069 0.9851 ms 81.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=128, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
triton_mm_13073 1.0056 ms 79.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=4, num_warps=8
addmm 1.0803 ms 74.3%
triton_mm_13066 1.1203 ms 71.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, OUT_DTYPE='tl.float16', USE_FAST_ACCUM=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8675 seconds and 0.0002 seconds precompiling for 21 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.7674239873886108, "best_triton_pos": 1, "best_triton_time": 6.0200958251953125, "best_triton_kernel": "triton_convolution2d_13174", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x2560x8x8, 1280x2560x3x3)
strides: [163840, 64, 8, 1], [23040, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 1.7674 ms 100.0%
triton_convolution2d_13174 6.0201 ms 29.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_13179 6.2638 ms 28.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13175 9.0778 ms 19.5% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_13183 9.2252 ms 19.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13176 9.4597 ms 18.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13180 10.1632 ms 17.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_13188 10.7622 ms 16.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13182 12.2194 ms 14.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_13184 16.0410 ms 11.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 2.7490 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:21:08.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:08.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:08.761000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:08.762000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:08.762000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:08.762000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "triton_convolution2d_13217", "best_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8", "best_time": 0.2242559939622879, "best_triton_pos": 0}
AUTOTUNE convolution(96x2560x8x8, 1280x2560x1x1)
strides: [163840, 64, 8, 1], [2560, 1, 1, 1]
dtypes: torch.float16, torch.float16
triton_convolution2d_13217 0.2243 ms 100.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_13225 0.2314 ms 96.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13213 0.2314 ms 96.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13226 0.2345 ms 95.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13220 0.2417 ms 92.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13221 0.2427 ms 92.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13216 0.2519 ms 89.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13219 0.2519 ms 89.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_13218 0.2591 ms 86.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=64, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=4, num_warps=4
triton_convolution2d_13222 0.2611 ms 85.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.1565 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 6.185984134674072, "best_triton_pos": 1, "best_triton_time": 25.394208908081055, "best_triton_kernel": "triton_convolution2d_13404", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x2560x16x16, 1280x2560x3x3)
strides: [655360, 256, 16, 1], [23040, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 6.1860 ms 100.0%
triton_convolution2d_13404 25.3942 ms 24.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_13409 26.4970 ms 23.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13405 38.7389 ms 16.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_13406 39.2509 ms 15.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13410 39.7875 ms 15.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_13413 41.3665 ms 15.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13418 43.2742 ms 14.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_13412 44.2982 ms 14.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_13414 62.7528 ms 9.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.2633 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:21:17.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:17.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:17.935000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:17.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:17.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:17.936000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.8826879858970642, "best_triton_pos": 1, "best_triton_time": 0.92876797914505, "best_triton_kernel": "triton_convolution2d_13451", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x2560x16x16, 1280x2560x1x1)
strides: [655360, 256, 16, 1], [2560, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.8827 ms 100.0%
triton_convolution2d_13451 0.9288 ms 95.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13456 0.9667 ms 91.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13445 1.0803 ms 81.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13441 1.1182 ms 78.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_13440 1.3117 ms 67.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_13455 1.3445 ms 65.7% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13450 1.3783 ms 64.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13446 1.3834 ms 63.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_13452 1.3875 ms 63.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8696 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 4.669439792633057, "best_triton_pos": 1, "best_triton_time": 17.729536056518555, "best_triton_kernel": "triton_convolution2d_14026", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x1920x16x16, 1280x1920x3x3)
strides: [491520, 256, 16, 1], [17280, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 4.6694 ms 100.0%
triton_convolution2d_14026 17.7295 ms 26.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_14031 17.8340 ms 26.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14027 25.5478 ms 18.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=1024, BLOCK_N=16, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=1, num_warps=8
triton_convolution2d_14035 27.2456 ms 17.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14032 28.9751 ms 16.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_14028 29.1430 ms 16.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14034 31.9898 ms 14.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_14040 32.4291 ms 14.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14036 46.6104 ms 10.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.1226 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:21:24.967000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:24.967000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:24.967000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:24.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:24.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:24.968000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.6737920045852661, "best_triton_pos": 1, "best_triton_time": 0.6758400201797485, "best_triton_kernel": "triton_convolution2d_14073", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1920x16x16, 1280x1920x1x1)
strides: [491520, 256, 16, 1], [1920, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.6738 ms 100.0%
triton_convolution2d_14073 0.6758 ms 99.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14078 0.7035 ms 95.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14063 0.7875 ms 85.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_14065 0.7967 ms 84.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14067 0.7997 ms 84.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14072 0.8110 ms 83.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14069 0.8182 ms 82.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_14077 0.8182 ms 82.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14068 0.8387 ms 80.3% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8236 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 12.74777603149414, "best_triton_pos": 1, "best_triton_time": 46.912513732910156, "best_triton_kernel": "triton_convolution2d_14342", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1280x32x32, 1280x1280x3x3)
strides: [1310720, 1024, 32, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 12.7478 ms 100.0%
triton_convolution2d_14342 46.9125 ms 27.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14337 63.7532 ms 20.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_14346 68.5794 ms 18.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14339 76.2347 ms 16.7% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14343 76.7293 ms 16.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_14351 83.7489 ms 15.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14345 102.6540 ms 12.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_14347 122.5329 ms 10.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14352 126.3729 ms 10.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 16.6198 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 10.049535751342773, "best_triton_pos": 1, "best_triton_time": 49.38035202026367, "best_triton_kernel": "triton_convolution2d_14354", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4"}
AUTOTUNE convolution(96x1920x32x32, 640x1920x3x3)
strides: [1966080, 1024, 32, 1], [17280, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 10.0495 ms 100.0%
triton_convolution2d_14354 49.3804 ms 20.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_14359 49.7940 ms 20.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14360 57.6707 ms 17.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_14363 58.2543 ms 17.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14356 59.5128 ms 16.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14368 67.9926 ms 14.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14362 70.3713 ms 14.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_14364 92.9280 ms 10.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14369 96.0184 ms 10.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 12.8324 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:21:55.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:55.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:55.508000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:21:55.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:21:55.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:21:55.509000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.3363200426101685, "best_triton_pos": 1, "best_triton_time": 1.8565119504928589, "best_triton_kernel": "triton_convolution2d_14401", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1920x32x32, 640x1920x1x1)
strides: [1966080, 1024, 32, 1], [1920, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 1.3363 ms 100.0%
triton_convolution2d_14401 1.8565 ms 72.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14406 1.8790 ms 71.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14395 1.9814 ms 67.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14390 2.2190 ms 60.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_14393 2.8816 ms 46.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14405 2.9000 ms 46.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14400 2.9020 ms 46.0% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14397 2.9174 ms 45.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_14402 2.9204 ms 45.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.0598 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 7.055359840393066, "best_triton_pos": 1, "best_triton_time": 23.339008331298828, "best_triton_kernel": "triton_convolution2d_14670", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1280x32x32, 640x1280x3x3)
strides: [1310720, 1024, 32, 1], [11520, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 7.0554 ms 100.0%
triton_convolution2d_14670 23.3390 ms 30.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14665 32.1690 ms 21.9% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_14674 34.7802 ms 20.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14671 37.9156 ms 18.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_14667 38.2013 ms 18.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14679 41.9062 ms 16.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14673 50.9737 ms 13.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_14675 61.2577 ms 11.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14680 64.0420 ms 11.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.4295 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:22:04.924000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:04.924000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:04.924000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:22:04.925000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:04.925000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:04.925000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.9154559969902039, "best_triton_pos": 1, "best_triton_time": 1.321984052658081, "best_triton_kernel": "triton_convolution2d_14712", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x1280x32x32, 640x1280x1x1)
strides: [1310720, 1024, 32, 1], [1280, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.9155 ms 100.0%
triton_convolution2d_14712 1.3220 ms 69.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14717 1.3384 ms 68.4% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14706 1.3957 ms 65.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14701 1.5155 ms 60.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_14704 2.0070 ms 45.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14711 2.0163 ms 45.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14708 2.0306 ms 45.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_14713 2.0326 ms 45.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_14716 2.0326 ms 45.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.9439 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 5.268479824066162, "best_triton_pos": 1, "best_triton_time": 17.278976440429688, "best_triton_kernel": "triton_convolution2d_14981", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x960x32x32, 640x960x3x3)
strides: [983040, 1024, 32, 1], [8640, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 5.2685 ms 100.0%
triton_convolution2d_14981 17.2790 ms 30.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14976 23.8940 ms 22.0% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_14985 25.7126 ms 20.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14978 28.4580 ms 18.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14982 28.9382 ms 18.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_14990 31.8413 ms 16.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14984 38.8495 ms 13.6% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_14986 46.3268 ms 11.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_14991 48.0707 ms 11.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 6.4291 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:22:12.293000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:12.293000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:12.293000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:22:12.294000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:12.294000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:12.294000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 0.7004160284996033, "best_triton_pos": 1, "best_triton_time": 1.0588159561157227, "best_triton_kernel": "triton_convolution2d_15023", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x960x32x32, 640x960x1x1)
strides: [983040, 1024, 32, 1], [960, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 0.7004 ms 100.0%
triton_convolution2d_15023 1.0588 ms 66.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15028 1.0680 ms 65.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15017 1.1039 ms 63.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15012 1.1786 ms 59.4% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_15015 1.5739 ms 44.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15022 1.5790 ms 44.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15019 1.5820 ms 44.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_15024 1.5852 ms 44.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15027 1.5913 ms 44.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 0.8990 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 15.544320106506348, "best_triton_pos": 1, "best_triton_time": 48.03993606567383, "best_triton_kernel": "triton_convolution2d_15292", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x640x64x64, 640x640x3x3)
strides: [2621440, 4096, 64, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 15.5443 ms 100.0%
triton_convolution2d_15292 48.0399 ms 32.4% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15287 49.8565 ms 31.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_15296 67.1212 ms 23.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15289 77.0458 ms 20.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15293 78.4189 ms 19.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_15301 80.7803 ms 19.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15295 102.0938 ms 15.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_15297 122.4602 ms 12.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15298 126.6412 ms 12.3% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 16.7493 seconds and 0.0002 seconds precompiling for 18 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 13.86188793182373, "best_triton_pos": 1, "best_triton_time": 34.48729705810547, "best_triton_kernel": "triton_convolution2d_15309", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x960x64x64, 320x960x3x3)
strides: [3932160, 4096, 64, 1], [8640, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 13.8619 ms 100.0%
triton_convolution2d_15309 34.4873 ms 40.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15304 48.1608 ms 28.8% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_15313 52.2701 ms 26.5% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15306 58.1253 ms 23.8% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15310 59.3725 ms 23.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_15318 67.2881 ms 20.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15312 75.3551 ms 18.4% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_15314 91.8313 ms 15.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15316 93.0396 ms 14.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 12.7575 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:22:43.000000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:43.000000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:43.000000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:22:43.002000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:43.002000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:43.002000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.5820800065994263, "best_triton_pos": 1, "best_triton_time": 2.525183916091919, "best_triton_kernel": "triton_convolution2d_15351", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x960x64x64, 320x960x1x1)
strides: [3932160, 4096, 64, 1], [960, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 1.5821 ms 100.0%
triton_convolution2d_15351 2.5252 ms 62.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15356 2.5416 ms 62.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15345 2.6798 ms 59.0% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15340 2.7617 ms 57.3% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_15343 3.5789 ms 44.2% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15352 3.5912 ms 44.1% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15347 3.6004 ms 43.9% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_15350 3.6106 ms 43.8% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15355 3.6250 ms 43.6% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.1783 seconds and 0.0002 seconds precompiling for 19 choices
Autotune Choices Stats:
{"num_choices": 18, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 9.4453763961792, "best_triton_pos": 1, "best_triton_time": 23.829504013061523, "best_triton_kernel": "triton_convolution2d_15620", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x640x64x64, 320x640x3x3)
strides: [2621440, 4096, 64, 1], [5760, 9, 3, 1]
dtypes: torch.float16, torch.float16
convolution 9.4454 ms 100.0%
triton_convolution2d_15620 23.8295 ms 39.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15615 24.7583 ms 38.2% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=256, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=4
triton_convolution2d_15624 34.7996 ms 27.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15617 37.9208 ms 24.9% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15621 39.8203 ms 23.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=3, num_warps=8
triton_convolution2d_15629 41.9860 ms 22.5% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15623 50.3962 ms 18.7% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=64, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=4, num_warps=4
triton_convolution2d_15625 61.5721 ms 15.3% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
triton_convolution2d_15626 63.8188 ms 14.8% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=3, KERNEL_W=3, PADDING_H=1, PADDING_W=1, STRIDE_H=1, STRIDE_W=1, UNROLL=False, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 8.5777 seconds and 0.0002 seconds precompiling for 18 choices
E0801 17:22:52.661000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:52.661000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 131072 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:52.661000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
E0801 17:22:52.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Runtime error during autotuning:
E0801 17:22:52.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] No valid triton configs. OutOfMemoryError: out of resource: triton_convolution2d Required: 147456 Hardware limit:101376 Reducing block sizes or `num_stages` may help..
E0801 17:22:52.662000 2856711 torch/_inductor/select_algorithm.py:4888] [0/2] Ignoring this choice.
Autotune Choices Stats:
{"num_choices": 19, "num_triton_choices": 17, "best_kernel": "convolution", "best_time": 1.124351978302002, "best_triton_pos": 1, "best_triton_time": 1.8688000440597534, "best_triton_kernel": "triton_convolution2d_15662", "best_triton_kernel_desc": "ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8"}
AUTOTUNE convolution(96x640x64x64, 320x640x1x1)
strides: [2621440, 4096, 64, 1], [640, 1, 1, 1]
dtypes: torch.float16, torch.float16
convolution 1.1244 ms 100.0%
triton_convolution2d_15662 1.8688 ms 60.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15667 1.9057 ms 59.0% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15656 1.9191 ms 58.6% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15651 1.9200 ms 58.6% ALLOW_TF32=False, BLOCK_K=16, BLOCK_M=64, BLOCK_N=256, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=4
triton_convolution2d_15654 2.5294 ms 44.5% ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15663 2.5436 ms 44.2% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15658 2.5457 ms 44.2% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=128, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=3, num_warps=8
triton_convolution2d_15661 2.5477 ms 44.1% ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
triton_convolution2d_15666 2.5590 ms 43.9% ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=256, BLOCK_N=128, GROUPS=1, KERNEL_H=1, KERNEL_W=1, PADDING_H=0, PADDING_W=0, STRIDE_H=1, STRIDE_W=1, UNROLL=True, num_stages=2, num_warps=8
SingleProcess AUTOTUNE benchmarking takes 1.0370 seconds and 0.0002 seconds precompiling for 19 choices
Out[20]:
Batch B=32 Summary:
Total Batch Time : 17061.18 ms
Effective Per-Image Time : 533.16 ms
Peak VRAM Memory : 19.89 GB
Effective Speedup Factor : 3.17x Faster!
In [21]:
## Compilation
import torch
print("Compiling pipe.unet with PyTorch 2.0 TorchInductor (mode='reduce-overhead')...")
t_comp_start = time.perf_counter()
pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead")
# pipe.unet = torch.compile(pipe.unet,
# options={
# "triton.cudagraphs": True,
# "max_autotune": True})
if device == "cuda":
torch.cuda.synchronize()
print(f"UNet marked for compilation. (Setup time: {time.perf_counter() - t_comp_start:.2f} s)") Out[21]:
Compiling pipe.unet with PyTorch 2.0 TorchInductor (mode='reduce-overhead')... UNet marked for compilation. (Setup time: 0.00 s)
In [22]:
batch_paths = [test_img_path] * 32
batch_id_prompts = [id_prompt] * 32
batch_ood_prompts = [ood_prompt] * 32
for B_size in [1, 2, 4, 8, 16, 32]:
print(f"\n Testing Micro-Batch Size B={B_size}...")
sub_paths = batch_paths[:B_size]
sub_id = batch_id_prompts[:B_size]
sub_ood = batch_ood_prompts[:B_size]
# Reset peak memory tracker
torch.cuda.reset_peak_memory_stats()
# 1. Warmup run specifically for this batch size shape
_, _, _ = profiled_generate_sona_outlier_optimized5(
pipe, sub_paths, sub_id, sub_ood, stop_timestep=23
)
# 2. Benchmark runs
r_total_ms, r_per_img_ms = [], []
for run in range(3):
_, total_ms, per_img_ms = profiled_generate_sona_outlier_optimized5(
pipe, sub_paths, sub_id, sub_ood, stop_timestep=23
)
r_total_ms.append(total_ms)
r_per_img_ms.append(per_img_ms)
avg_total = np.mean(r_total_ms)
avg_per_img = np.mean(r_per_img_ms)
eff_speedup = baseline_total / avg_per_img
peak_vram = torch.cuda.max_memory_allocated() / 1e9
print(f" Batch B={B_size} Summary:")
print(f" Total Batch Time : {avg_total:8.2f} ms")
print(f" Effective Per-Image Time : {avg_per_img:8.2f} ms")
print(f" Peak VRAM Memory : {peak_vram:8.2f} GB")
print(f" Effective Speedup Factor : {eff_speedup:8.2f}x Faster!")
Out[22]:
Testing Micro-Batch Size B=1...
Batch B=1 Summary:
Total Batch Time : 601.80 ms
Effective Per-Image Time : 601.80 ms
Peak VRAM Memory : 4.63 GB
Effective Speedup Factor : 2.81x Faster!
Testing Micro-Batch Size B=2...
Batch B=2 Summary:
Total Batch Time : 1070.28 ms
Effective Per-Image Time : 535.14 ms
Peak VRAM Memory : 4.63 GB
Effective Speedup Factor : 3.16x Faster!
Testing Micro-Batch Size B=4...
Batch B=4 Summary:
Total Batch Time : 2100.27 ms
Effective Per-Image Time : 525.07 ms
Peak VRAM Memory : 4.77 GB
Effective Speedup Factor : 3.22x Faster!
Testing Micro-Batch Size B=8...
Batch B=8 Summary:
Total Batch Time : 4244.80 ms
Effective Per-Image Time : 530.60 ms
Peak VRAM Memory : 6.93 GB
Effective Speedup Factor : 3.19x Faster!
Testing Micro-Batch Size B=16...
Batch B=16 Summary:
Total Batch Time : 8680.53 ms
Effective Per-Image Time : 542.53 ms
Peak VRAM Memory : 11.25 GB
Effective Speedup Factor : 3.12x Faster!
Testing Micro-Batch Size B=32...
Batch B=32 Summary:
Total Batch Time : 17137.06 ms
Effective Per-Image Time : 535.53 ms
Peak VRAM Memory : 19.89 GB
Effective Speedup Factor : 3.16x Faster!
In [23]:
print(test_img_path)
print(id_prompt)
print(ood_prompt) Out[23]:
/storage/ice-shared/cs8903onl/kernel-datasets/data/images_largescale/imagenet_1k/train/n01484850/n01484850_22902.JPEG a photo of a great white shark a photo of a harvester
In [78]:
img_check, _ = profiled_generate_sona_outlier_optimized3(
pipe,
input_image_path=test_img_path,
id_class_prompt=id_prompt,
ood_target_prompt="a photo of a harvester",
stop_timestep=23,
guidance_scale=13.0,
threshold_percentile=0.2
)
plt.imshow(img_check) Out[78]:
Out[78]:
In [82]:
import matplotlib.pyplot as plt
import numpy as np
stages = [
'Baseline\n(Serial UNet)',
'Opt 1\n(+ Batched UNet)',
'Opt 2\n(+ Text Cache)',
'Opt 3\n(+ Inference Mode)',
'Opt 4\n(+ TorchCompile)',
'Opt 5\n(+ Micro-Batch B=4)'
]
latencies_ms = [1691.39, 685.75, 683.82, 669.58, 601.80, 525.07]
baseline_total = 1691.39
speedups = [baseline_total / l for l in latencies_ms]
# Academic style setup
plt.rcParams['font.family'] = 'DejaVu Sans'
fig, ax = plt.subplots(figsize=(9, 5.5), dpi=300)
# Muted academic color palette (Slate Blue to Deep Navy)
colors = ['#4a6572', '#34495e', '#2c3e50', '#2c3e50', '#1a252f', '#0f171e']
bars = ax.bar(stages, latencies_ms, color=colors, width=0.52, edgecolor='black', linewidth=0.8)
# Minimalist academic grid & styling
ax.set_ylabel('Per-Image Latency (ms)', fontsize=11, fontweight='bold')
ax.set_title('SONA Diffusion Optimization Progression (NVIDIA L40S GPU)', fontsize=12, fontweight='bold', pad=12)
ax.set_ylim(0, 1950)
# Light horizontal grid lines only
ax.yaxis.grid(True, linestyle='--', linewidth=0.5, alpha=0.7, color='#cccccc')
ax.set_axisbelow(True)
# Remove top and right spines (IEEE/ACM style)
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.spines['left'].set_linewidth(0.8)
ax.spines['bottom'].set_linewidth(0.8)
# Annotate bars cleanly with exact values and speedup factors
for bar, ms, sp in zip(bars, latencies_ms, speedups):
yval = bar.get_height()
label_text = f'{ms:.1f} ms\n(1.00x)' if sp == 1.0 else f'{ms:.1f} ms\n({sp:.2f}x)'
ax.text(
bar.get_x() + bar.get_width()/2.0,
yval + 25,
label_text,
ha='center',
va='bottom',
fontsize=9,
fontweight='medium',
color='#111111'
)
plt.xticks(fontsize=9.5)
plt.tight_layout()
plt.savefig('sona_optimization_benchmark.png', dpi=300, bbox_inches='tight')
plt.show()
Out[82]:
