chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Support running FP4 Deepseek on SM120. (#11708)

Browse Source
This commit is contained in:
weiliang
2025-10-28 08:37:49 +08:00
committed by GitHub
parent a6ea3add76
commit 88596739a4
9 changed files with 33 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/quantization/fp8_utils.py
View File

@@ -5,7 +5,7 @@ import torch
from sglang.srt.layers import deep_gemm_wrapper
from sglang.srt.layers.quantization.fp8_kernel import sglang_per_token_group_quant_fp8
from sglang.srt.layers.quantization.mxfp4_tensor import MXFP4QuantizeUtil
from sglang.srt.utils import ceil_div, is_sm100_supported, offloader
from sglang.srt.utils import ceil_div, is_blackwell_supported, offloader
try:
from vllm import _custom_ops as ops
@@ -129,7 +129,7 @@ def cutlass_block_fp8_supported() -> bool:
CUTLASS_BLOCK_FP8_SUPPORTED = cutlass_block_fp8_supported()
ENABLE_FLASHINFER_GEMM = (
get_bool_env_var("SGLANG_ENABLE_FLASHINFER_GEMM")
and is_sm100_supported()
and is_blackwell_supported()
and is_flashinfer_available()
)
if ENABLE_FLASHINFER_GEMM: