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

[1/2] Optimizations and refactors about quant kernel (#9534)

Browse Source
This commit is contained in:
fzyzcjy
2025-09-05 18:45:08 +08:00
committed by GitHub
parent afd9f2f560
commit 339f8eef09
11 changed files with 996 additions and 329 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
sgl-kernel/csrc/common_extension.cc
View File

@@ -121,14 +121,9 @@ 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,"
" 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.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(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", torch::kCUDA, &sgl_per_token_group_quant_8bit);
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);