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

6
sgl-kernel/benchmark/bench_moe_align_block_size.py
View File

@@ -324,7 +324,7 @@ def benchmark(num_tokens, num_experts, topk, provider):
quantiles = [0.5, 0.2, 0.8]
if provider == "sgl":
ms, min_ms, max_ms = triton.testing.do_bench(
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
lambda: sgl_moe_align_block_size_with_empty(
topk_ids,
num_experts,
@@ -336,7 +336,7 @@ def benchmark(num_tokens, num_experts, topk, provider):
quantiles=quantiles,
)
elif provider == "sgl_fusion":
ms, min_ms, max_ms = triton.testing.do_bench(
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
lambda: sgl_moe_align_block_size_with_empty(
topk_ids,
num_experts,
@@ -350,7 +350,7 @@ def benchmark(num_tokens, num_experts, topk, provider):
)
elif provider == "triton":
sorted_ids.fill_(topk_ids.numel())
ms, min_ms, max_ms = triton.testing.do_bench(
ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
lambda: moe_align_block_size_triton(
topk_ids,
num_experts,