chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[fix] fix DeepGEMM blackwell input quant & ut & fix style and log (#7247)

Browse Source
This commit is contained in:
JieXin Liang
2025-06-17 02:45:54 +08:00
committed by GitHub
parent e30ef368ab
commit 5ca07eed90
10 changed files with 285 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
python/sglang/srt/layers/quantization/fp8_utils.py
View File

@@ -241,9 +241,10 @@ def deepgemm_w8a8_block_fp8_linear_with_fallback(
scale_ue8m0=deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0,
)
if get_bool_env_var("SGLANG_W8A8_DEEPGEMM_SANITY_CHECK_UE8M0"):
_check_ue8m0("x_scale", x_scale)
_check_ue8m0("weight_scale", weight_scale)
# NOTE(alcanderian): Useless when scale is packed to int32
# if get_bool_env_var("SGLANG_W8A8_DEEPGEMM_SANITY_CHECK_UE8M0"):
# _check_ue8m0("x_scale", x_scale)
# _check_ue8m0("weight_scale", ws)
output = w8a8_block_fp8_matmul_deepgemm(
q_input, weight, x_scale, weight_scale, block_size, output_dtype=output_dtype