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

Fix bench script making input data on L2 cache (#7739)

Browse Source
This commit is contained in:
fzyzcjy
2025-07-27 15:30:24 +08:00
committed by GitHub
parent 62222bd27e
commit e34cf6ad75
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
View File

@@ -205,9 +205,9 @@ def benchmark(batch_size, seq_len, group_size, dst_dtype, provider):
quantiles = [0.5, 0.2, 0.8] quantiles = [0.5, 0.2, 0.8]
if provider == "triton": if provider == "triton":
fn = lambda: triton_per_token_group_quant_8bit(x.clone(), group_size, dst_dtype) fn = lambda: triton_per_token_group_quant_8bit(x, group_size, dst_dtype)
elif provider == "sglang": elif provider == "sglang":
fn = lambda: sglang_per_token_group_quant_8bit(x.clone(), group_size, dst_dtype) fn = lambda: sglang_per_token_group_quant_8bit(x, group_size, dst_dtype)
ms, min_ms, max_ms = triton.testing.do_bench(fn, quantiles=quantiles) ms, min_ms, max_ms = triton.testing.do_bench(fn, quantiles=quantiles)