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

[1/2] Add Kernel support for Cutlass based Fused FP4 MoE (#6093)

Browse Source 
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Pavani Majety
2025-06-02 13:48:03 -07:00
committed by GitHub
parent df7f61ee7d
commit eb38c7d1ca
12 changed files with 1677 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
sgl-kernel/csrc/gemm/nvfp4_quant_entry.cu
View File

@@ -18,6 +18,15 @@ limitations under the License.
#if defined ENABLE_NVFP4 && ENABLE_NVFP4
void scaled_fp4_quant_sm100a(
torch::Tensor& output, torch::Tensor const& input, torch::Tensor& output_sf, torch::Tensor const& input_sf);
void scaled_fp4_experts_quant_sm100a(
torch::Tensor& output,
torch::Tensor& output_scale,
torch::Tensor const& input,
torch::Tensor const& input_global_scale,
torch::Tensor const& input_offset_by_experts,
torch::Tensor const& output_scale_offset_by_experts);
#endif
void scaled_fp4_quant(
@@ -27,3 +36,17 @@ void scaled_fp4_quant(
#endif
TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 quantization");
}
void scaled_fp4_experts_quant(
torch::Tensor& output,
torch::Tensor& output_scale,
torch::Tensor const& input,
torch::Tensor const& input_global_scale,
torch::Tensor const& input_offset_by_experts,
torch::Tensor const& output_scale_offset_by_experts) {
#if defined ENABLE_NVFP4 && ENABLE_NVFP4
return scaled_fp4_experts_quant_sm100a(
output, output_scale, input, input_global_scale, input_offset_by_experts, output_scale_offset_by_experts);
#endif
TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 experts quantization kernel");
}