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

[sgl-kernel][Feat][B200][1/N] Support MXFP8 Grouped GEMM in Blackwell (#13731)

Browse Source 
Co-authored-by: Yineng Zhang <me@zhyncs.com>
This commit is contained in:
Qi Yuhang
2025-12-04 10:09:09 +08:00
committed by GitHub
co-authored by Yineng Zhang
parent df026bb110
commit 16ff892c18
12 changed files with 1174 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
sgl-kernel/csrc/common_extension.cc
+8
View File
@@ -576,6 +576,14 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
"stride_a, Tensor stride_b, Tensor stride_d, Tensor problem_sizes, Tensor expert_offsets, Tensor workspace) -> "
"()");
m.impl("es_fp8_blockwise_scaled_grouped_mm", &es_fp8_blockwise_scaled_grouped_mm);
m.def(
"es_sm100_mxfp8_blockscaled_grouped_mm(Tensor a, Tensor b, Tensor sfa, Tensor sfb, Tensor d, Tensor "
"problem_sizes, Tensor expert_offsets, Tensor blockscale_offsets) -> ()");
m.impl("es_sm100_mxfp8_blockscaled_grouped_mm", &es_sm100_mxfp8_blockscaled_grouped_mm);
m.def(
"es_sm100_mxfp8_blockscaled_grouped_quant(Tensor input, Tensor problem_sizes, Tensor expert_offsets, Tensor "
"blockscale_offsets, Tensor quant_output, Tensor scale_factor) -> () ");
m.impl("es_sm100_mxfp8_blockscaled_grouped_quant", &es_sm100_mxfp8_blockscaled_grouped_quant);
/*
* From fast-hadamard-transform