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

75
sgl-kernel/benchmark/bench_fp8_blockwise_group_gemm.py
View File

@@ -1,4 +1,11 @@
import argparse
import os
# CI environment detection
IS_CI = (
os.getenv("CI", "false").lower() == "true"
or os.getenv("GITHUB_ACTIONS", "false").lower() == "true"
)
import random
from dataclasses import dataclass
from typing import List, Tuple
@@ -290,36 +297,44 @@ def main():
parser = argparse.ArgumentParser()
parser.add_argument("--num-warmup", type=int, default=3)
parser.add_argument("--num-run", type=int, default=10)
shape_args = [
# Prefill, DeepSeek-R1, gateup, chunk_size = 4096, TP = 8
ShapeArg(expected_m_per_group=128, n=512, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, TP = 8
ShapeArg(expected_m_per_group=256, n=512, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, TP = 16
ShapeArg(expected_m_per_group=256, n=256, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 16384, TP = 16
ShapeArg(expected_m_per_group=512, n=256, k=7168, num_groups=256),
# Decode, DeepSeek-R1, gateup, bs = 32, TP = 8
ShapeArg(expected_m_per_group=1, n=512, k=7168, num_groups=256),
# Decode, DeepSeek-R1, gateup, bs = 64, TP = 16
ShapeArg(expected_m_per_group=2, n=256, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, EP = 8
ShapeArg(expected_m_per_group=256, n=4096, k=7168, num_groups=32),
# Prefill, DeepSeek-R1, gateup, chunk_size = 16384, EP = 16
ShapeArg(expected_m_per_group=512, n=4096, k=7168, num_groups=16),
# Decode, DeepSeek-R1, gateup, bs = 128, EP = 8
ShapeArg(expected_m_per_group=4, n=4096, k=7168, num_groups=32),
# Decode, DeepSeek-R1, gateup, bs = 256, EP = 16
ShapeArg(expected_m_per_group=8, n=4096, k=7168, num_groups=16),
# Prefill, Qwen3-235B-A22B-FP8, gateup, chunk_size = 16384, TP = 4
ShapeArg(expected_m_per_group=1024, n=768, k=4096, num_groups=128),
# Prefill, Qwen3-235B-A22B-FP8, down, chunk_size = 16384, TP = 4
ShapeArg(expected_m_per_group=1024, n=4096, k=384, num_groups=128),
# Decode, Qwen3-235B-A22B-FP8, gateup, bs = 256, TP = 4
ShapeArg(expected_m_per_group=16, n=768, k=4096, num_groups=128),
# Decode, Qwen3-235B-A22B-FP8, down, bs = 256, TP = 4
ShapeArg(expected_m_per_group=16, n=4096, k=384, num_groups=128),
]
# CI environment uses simplified parameters
if IS_CI:
shape_args = [
# Only test one simple shape in CI
ShapeArg(expected_m_per_group=128, n=512, k=7168, num_groups=256),
]
else:
shape_args = [
# Prefill, DeepSeek-R1, gateup, chunk_size = 4096, TP = 8
ShapeArg(expected_m_per_group=128, n=512, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, TP = 8
ShapeArg(expected_m_per_group=256, n=512, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, TP = 16
ShapeArg(expected_m_per_group=256, n=256, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 16384, TP = 16
ShapeArg(expected_m_per_group=512, n=256, k=7168, num_groups=256),
# Decode, DeepSeek-R1, gateup, bs = 32, TP = 8
ShapeArg(expected_m_per_group=1, n=512, k=7168, num_groups=256),
# Decode, DeepSeek-R1, gateup, bs = 64, TP = 16
ShapeArg(expected_m_per_group=2, n=256, k=7168, num_groups=256),
# Prefill, DeepSeek-R1, gateup, chunk_size = 8192, EP = 8
ShapeArg(expected_m_per_group=256, n=4096, k=7168, num_groups=32),
# Prefill, DeepSeek-R1, gateup, chunk_size = 16384, EP = 16
ShapeArg(expected_m_per_group=512, n=4096, k=7168, num_groups=16),
# Decode, DeepSeek-R1, gateup, bs = 128, EP = 8
ShapeArg(expected_m_per_group=4, n=4096, k=7168, num_groups=32),
# Decode, DeepSeek-R1, gateup, bs = 256, EP = 16
ShapeArg(expected_m_per_group=8, n=4096, k=7168, num_groups=16),
# Prefill, Qwen3-235B-A22B-FP8, gateup, chunk_size = 16384, TP = 4
ShapeArg(expected_m_per_group=1024, n=768, k=4096, num_groups=128),
# Prefill, Qwen3-235B-A22B-FP8, down, chunk_size = 16384, TP = 4
ShapeArg(expected_m_per_group=1024, n=4096, k=384, num_groups=128),
# Decode, Qwen3-235B-A22B-FP8, gateup, bs = 256, TP = 4
ShapeArg(expected_m_per_group=16, n=768, k=4096, num_groups=128),
# Decode, Qwen3-235B-A22B-FP8, down, bs = 256, TP = 4
ShapeArg(expected_m_per_group=16, n=4096, k=384, num_groups=128),
]
args = parser.parse_args()
benchmark_one_shape(shape_args, args.num_warmup, args.num_run)