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

[Compile] Fix Compile Warning for Ignoring MIN_BLOCK_PER_SM (#25193)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-09-19 18:23:19 -04:00
committed by GitHub
parent e69e0b8b5f
commit 711e912946
4 changed files with 53 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
csrc/quantization/fp4/nvfp4_quant_kernels.cu
View File

@@ -26,13 +26,14 @@
#include "dispatch_utils.h"
#include "cuda_utils.h"
#include "launch_bounds_utils.h"
#include "nvfp4_utils.cuh"
namespace vllm {
// Use UE4M3 by default.
template <class Type, bool UE8M0_SF = false>
__global__ void __launch_bounds__(512, 4)
__global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
float const* SFScale, uint32_t* out, uint32_t* SFout) {
using PackedVec = PackedVec<Type>;
@@ -75,8 +76,9 @@ void invokeFP4Quantization(int m, int n, T const* input, float const* SFScale,
// Grid, Block size.
// Each thread converts 8 values.
dim3 block(std::min(int(n / ELTS_PER_THREAD), 512));
// Get number of blocks per SM (assume we can fully utilize the SM).
int const numBlocksPerSM = 2048 / block.x;
// Get number of blocks per SM
int const numBlocksPerSM =
vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
// Launch the cvt kernel.