EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Restruct sgl-kernel benchmark (#10861)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-09-25 07:45:25 +08:00
committed by GitHub
parent 7a06ef984d
commit c4e314f986
27 changed files with 425 additions and 319 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
sgl-kernel/benchmark/bench_int8_gemm.py
View File

@@ -86,12 +86,12 @@ def benchmark(batch_size, provider, N, K):
quantiles = [0.5, 0.2, 0.8]
if provider == "sgl-kernel":
ms, min_ms, max_ms = triton.testing.do_bench(
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
lambda: int8_scaled_mm(a, b, scale_a, scale_b, torch.float16, bias),
quantiles=quantiles,
)
if provider == "vllm":
ms, min_ms, max_ms = triton.testing.do_bench(
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
lambda: vllm_scaled_mm(a, b, scale_a, scale_b, torch.float16, bias),
quantiles=quantiles,
)
@@ -139,8 +139,6 @@ if __name__ == "__main__":
KN_model_names = prepare_shapes(args)
for K, N, model_name in KN_model_names:
print(f"{model_name} N={N} K={K}: ")
benchmark.run(
print_data=True, show_plots=True, save_path="bench_int8_res", N=N, K=K
)
benchmark.run(print_data=True, N=N, K=K)
print("Benchmark finished!")