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

[Refactor][Kernel] Add global helper to deduplicate vectorized memory ops (#35105)

Browse Source 
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: LopezCastroRoberto <roberto.lopez.castro@udc.es>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
This commit is contained in:
Roberto L. Castro
2026-02-28 01:28:17 +01:00
committed by GitHub
parent e3691988d0
commit a201ad72d8
6 changed files with 474 additions and 372 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
csrc/quantization/fp4/nvfp4_experts_quant.cu
View File

@@ -43,7 +43,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
uint32_t* input_offset_by_experts,
uint32_t* output_scale_offset_by_experts, int n_experts,
bool low_latency) {
using PackedVec = PackedVec<Type>;
using PackedVec = PackedVec<Type, CVT_FP4_PACK16>;
static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
(CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,
@@ -155,7 +155,7 @@ __global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
float const* SFScale, uint32_t* out, uint32_t* SFout,
uint32_t* input_offset_by_experts,
uint32_t* output_scale_offset_by_experts, int n_experts) {
using PackedVec = PackedVec<Type>;
using PackedVec = PackedVec<Type, CVT_FP4_PACK16>;
static constexpr int CVT_FP4_NUM_THREADS_PER_SF =
(CVT_FP4_SF_VEC_SIZE / CVT_FP4_ELTS_PER_THREAD);
static_assert(sizeof(PackedVec) == sizeof(Type) * CVT_FP4_ELTS_PER_THREAD,