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

[NVIDIA] Fix CUDA arch requirement in nvfp4 cast (#12581)

Browse Source 
Signed-off-by: Serge Panev <spanev@nvidia.com>
Co-authored-by: Fan Yin <1106310035@qq.com>
This commit is contained in:
Serge Panev
2026-01-22 05:21:11 +01:00
committed by GitHub
parent 3373545b9f
commit e95668abc7
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/csrc/gemm/nvfp4_quant.cuh
View File

@@ -52,7 +52,7 @@ constexpr int CVT_FP4_SF_VEC_SIZE = 16;
inline __device__ uint32_t fp32_vec_to_e2m1(float (&array)[8]) {
// PTX instructions used here requires >= sm100f.
#if CUTLASS_ARCH_MMA_SM100A_ENABLED || CUTLASS_ARCH_MMA_SM103A_ENABLED || CUTLASS_ARCH_MMA_SM120A_ENABLED || \
(defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ > 1000))
(defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ >= 1000))
uint32_t val;
asm volatile(
"{\n"
@@ -87,7 +87,7 @@ inline __device__ uint32_t fp32_vec_to_e2m1(float (&array)[8]) {
inline __device__ uint32_t fp32_vec_to_e2m1(float2 (&array)[4]) {
// PTX instructions used here requires >= sm100f.
#if CUTLASS_ARCH_MMA_SM100A_ENABLED || CUTLASS_ARCH_MMA_SM103A_ENABLED || CUTLASS_ARCH_MMA_SM120A_ENABLED || \
(defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ > 1000))
(defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ >= 1000))
uint32_t val;
asm volatile(
"{\n"