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

Use cutlass fp4 gemm by default (#11813)

Browse Source
This commit is contained in:
Qiaolin Yu
2025-10-18 14:10:15 -07:00
committed by GitHub
parent f4f8a1b4d8
commit ebda73dc72
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -79,7 +79,7 @@ CUTEDSL_MOE_SCALAR_INPUT_SCALE = get_bool_env_var(
"SGLANG_CUTEDSL_MOE_SCALAR_INPUT_SCALE", "true" "SGLANG_CUTEDSL_MOE_SCALAR_INPUT_SCALE", "true"
) )
USE_CUTLASS_BACKEND_FOR_FP4_GEMM = get_bool_env_var( USE_CUTLASS_BACKEND_FOR_FP4_GEMM = get_bool_env_var(
"SGLANG_USE_CUTLASS_BACKEND_FOR_FP4_GEMM" "SGLANG_USE_CUTLASS_BACKEND_FOR_FP4_GEMM", "true"
) )
# TODO make it true by default when the DeepEP PR is merged # TODO make it true by default when the DeepEP PR is merged
CUTEDSL_MOE_NVFP4_DISPATCH = get_bool_env_var( CUTEDSL_MOE_NVFP4_DISPATCH = get_bool_env_var(