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

Fix sgl-kernel benchmark dead code (#11022)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-09-29 15:06:40 +08:00
committed by GitHub
parent 71959545df
commit 11965b0daf
25 changed files with 1019 additions and 260 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
sgl-kernel/benchmark/bench_per_token_group_quant_8bit.py
View File

@@ -1,4 +1,5 @@
import itertools
import os
import time
from functools import partial
from pathlib import Path
@@ -16,15 +17,28 @@ from sglang.srt.layers.quantization.fp8_kernel import (
from sglang.srt.layers.quantization.fp8_kernel import sglang_per_token_group_quant_8bit
from sglang.srt.utils import is_hip
# CI environment detection
IS_CI = (
os.getenv("CI", "false").lower() == "true"
or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
)
_is_hip = is_hip()
fp8_type_ = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
num_tokens_range = [1, 4, 16, 64, 256, 768, 2048, 8192, 16384]
hidden_dim_range = [1536, 7168, 18432] # For DeepSeek V3/R1
group_size_range = [128] # For DeepSeek V3/R1
# TODO test int8
dst_dtype_range = [fp8_type_]
# CI environment uses simplified parameters
if IS_CI:
num_tokens_range = [64] # Single value for CI
hidden_dim_range = [1536] # Single value for CI
group_size_range = [128] # Keep as is
dst_dtype_range = [fp8_type_] # Keep as is
else:
num_tokens_range = [1, 4, 16, 64, 256, 768, 2048, 8192, 16384]
hidden_dim_range = [1536, 7168, 18432] # For DeepSeek V3/R1
group_size_range = [128] # For DeepSeek V3/R1
# TODO test int8
dst_dtype_range = [fp8_type_]
flags_range = [
dict(
column_major_scales=False,
@@ -82,7 +96,7 @@ def benchmark(num_tokens, hidden_dim, group_size, dst_dtype, flags, provider):
x = torch.randn(num_tokens, hidden_dim, device=device, dtype=torch.bfloat16)
fn, kernel_names = {
"triton": (triton_per_token_group_quant_8bit, "_per_token_group_quant_fp8"),
"triton": (triton_per_token_group_quant_8bit, "_per_token_group_quant_8bit"),
"sglang": (
sglang_per_token_group_quant_8bit,
"per_token_group_quant_8bit_kernel",