Download 04_c_sdpa_attention.py from ariG23498/profiling-pytorch: direct link, hf CLI and curl.
- Browser
- Download file 3.01 kB
-
https://huggingface.co/datasets/ariG23498/profiling-pytorch/resolve/main/04_c_sdpa_attention.py
- Command line
-
hf download hf://datasets/ariG23498/profiling-pytorch/04_c_sdpa_attention.py
-
curl -L -o 04_c_sdpa_attention.py https://huggingface.co/datasets/ariG23498/profiling-pytorch/resolve/main/04_c_sdpa_attention.py
3.01 kB
| import argparse | |
| import os | |
| import torch | |
| from torch.nn import functional as F | |
| from torch.nn.attention import SDPBackend, sdpa_kernel | |
| # the backends torch.nn.functional.scaled_dot_product_attention can dispatch to | |
| BACKENDS = { | |
| "math": SDPBackend.MATH, | |
| "flash": SDPBackend.FLASH_ATTENTION, | |
| "efficient": SDPBackend.EFFICIENT_ATTENTION, | |
| "cudnn": SDPBackend.CUDNN_ATTENTION, | |
| } | |
| def main(): | |
| p = argparse.ArgumentParser() | |
| p.add_argument("--batch", type=int, default=8) | |
| p.add_argument("--heads", type=int, default=16) | |
| p.add_argument("--seq", type=int, default=1024) | |
| p.add_argument("--head_dim", type=int, default=64) | |
| # "auto" lets SDPA pick the backend; the rest force a single backend | |
| p.add_argument( | |
| "--backend", | |
| choices=["auto", "math", "flash", "efficient", "cudnn"], | |
| default="auto", | |
| ) | |
| p.add_argument("--compile", action="store_true") | |
| p.add_argument("--trace_dir", default="./traces/04_c_sdpa_attention") | |
| args = p.parse_args() | |
| device = "cuda" | |
| dtype = torch.bfloat16 | |
| shape = (args.batch, args.heads, args.seq, args.head_dim) | |
| q = torch.randn(shape, device=device, dtype=dtype) | |
| k = torch.randn(shape, device=device, dtype=dtype) | |
| v = torch.randn(shape, device=device, dtype=dtype) | |
| def attn(q, k, v): | |
| # is_causal=True asks SDPA to apply the causal mask internally, | |
| # so we never build or materialize a [seq, seq] mask ourselves. | |
| return F.scaled_dot_product_attention(q, k, v, is_causal=True) | |
| fwd = torch.compile(attn) if args.compile else attn | |
| def step(): | |
| with torch.profiler.record_function("sdpa_fwd"), torch.no_grad(): | |
| if args.backend == "auto": | |
| return fwd(q, k, v) | |
| with sdpa_kernel(BACKENDS[args.backend]): | |
| return fwd(q, k, v) | |
| for _ in range(3): | |
| step() | |
| torch.cuda.synchronize() | |
| os.makedirs(args.trace_dir, exist_ok=True) | |
| compile_tag = "compile" if args.compile else "eager" | |
| tag = f"{args.batch}_{args.heads}_{args.seq}_{args.head_dim}_{args.backend}_{compile_tag}" | |
| table_path = os.path.join(args.trace_dir, f"{tag}.txt") | |
| trace_path = os.path.join(args.trace_dir, f"{tag}.json") | |
| schedule = torch.profiler.schedule(wait=1, warmup=1, active=3, repeat=1) | |
| with torch.profiler.profile( | |
| activities=[ | |
| torch.profiler.ProfilerActivity.CPU, | |
| torch.profiler.ProfilerActivity.CUDA, | |
| ], | |
| schedule=schedule, | |
| record_shapes=False, # adds CPU overhead | |
| profile_memory=False, # adds CPU overhead | |
| with_stack=False, # adds CPU overhead | |
| ) as prof: | |
| for _ in range(5): | |
| step() | |
| prof.step() | |
| torch.cuda.synchronize() | |
| print(f"saving traces ... {trace_path}") | |
| prof.export_chrome_trace(trace_path) | |
| with open(table_path, "w") as f: | |
| f.write(prof.key_averages().table(sort_by="cuda_time_total", row_limit=15)) | |
| if __name__ == "__main__": | |
| main() | |