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

Reland [1/2] Optimizations and refactors about quant kernel (#10312)

Browse Source 
Co-authored-by: Yineng Zhang <me@zhyncs.com>
This commit is contained in:
fzyzcjy
2025-10-11 15:59:03 +08:00
committed by GitHub
parent 129d299278
commit 21337b22b9
13 changed files with 1065 additions and 178 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
sgl-kernel/csrc/common_extension.cc
View File

@@ -136,14 +136,14 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
m.impl("fp8_blockwise_scaled_mm", torch::kCUDA, &fp8_blockwise_scaled_mm);
m.def(
"sgl_per_token_group_quant_fp8(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
"sgl_per_token_group_quant_8bit(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
" float eps, float fp8_min, float fp8_max, bool scale_ue8m0) -> ()");
m.impl("sgl_per_token_group_quant_fp8", torch::kCUDA, &sgl_per_token_group_quant_fp8);
m.impl("sgl_per_token_group_quant_8bit", torch::kCUDA, &sgl_per_token_group_quant_8bit);
m.def(
"sgl_per_token_group_quant_int8(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
" float eps, float int8_min, float int8_max) -> ()");
m.impl("sgl_per_token_group_quant_int8", torch::kCUDA, &sgl_per_token_group_quant_int8);
"sgl_per_token_group_quant_8bit_v2(Tensor input, Tensor output_q, Tensor output_s, int group_size,"
" float eps, float fp8_min, float fp8_max, bool scale_ue8m0, bool fuse_silu_and_mul, Tensor? masked_m) -> ()");
m.impl("sgl_per_token_group_quant_8bit_v2", torch::kCUDA, &sgl_per_token_group_quant_8bit_v2);
m.def("sgl_per_tensor_quant_fp8(Tensor input, Tensor output_q, Tensor output_s, bool is_static) -> ()");
m.impl("sgl_per_tensor_quant_fp8", torch::kCUDA, &sgl_per_tensor_quant_fp8);