biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Support CUTLASS NVFP4 (w4a4) for Blackwell Geforce GPUs (SM120) (#21309)

Browse Source 
Signed-off-by: LopezCastroRoberto <roberto.lopez.castro@udc.es>
This commit is contained in:
Roberto L. Castro
2025-08-03 09:54:22 +02:00
committed by GitHub
parent 3f36c325fa
commit 789562c28c
6 changed files with 329 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
csrc/quantization/fp4/nvfp4_quant_entry.cu
View File

@@ -16,14 +16,15 @@
#include <torch/all.h>
#if defined ENABLE_NVFP4 && ENABLE_NVFP4
void scaled_fp4_quant_sm100a(torch::Tensor const& output,
#if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
(defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
torch::Tensor const& input,
torch::Tensor const& output_sf,
torch::Tensor const& input_sf);
#endif
#if defined ENABLE_NVFP4 && ENABLE_NVFP4
#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
void scaled_fp4_experts_quant_sm100a(
torch::Tensor& output, torch::Tensor& output_scale,
torch::Tensor const& input, torch::Tensor const& input_global_scale,
@@ -33,8 +34,9 @@ void scaled_fp4_experts_quant_sm100a(
void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
torch::Tensor& output_sf, torch::Tensor const& input_sf) {
#if defined ENABLE_NVFP4 && ENABLE_NVFP4
return scaled_fp4_quant_sm100a(output, input, output_sf, input_sf);
#if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
(defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
return scaled_fp4_quant_sm1xxa(output, input, output_sf, input_sf);
#endif
TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 quantization kernel");
}
@@ -44,7 +46,7 @@ void scaled_fp4_experts_quant(
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
#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
return scaled_fp4_experts_quant_sm100a(
output, output_scale, input, input_global_scale, input_offset_by_experts,
output_scale_offset_by_experts);