Apply fixes for CUDA 13 (#24599)

Signed-off-by: Aidyn-A <aidyn.b.aitzhan@gmail.com>
This commit is contained in:
Aidyn-A
2025-09-17 09:15:42 -04:00
committed by GitHub
parent 9fccd04e30
commit bfe9380161
8 changed files with 47 additions and 56 deletions
+4 -9
View File
@@ -8,16 +8,11 @@
#include "type_convert.cuh"
#include "quantization/fp8/common.cuh"
#include "dispatch_utils.h"
#include "cub_helpers.h"
#include <torch/cuda.h>
#include <c10/cuda/CUDAGuard.h>
#ifndef USE_ROCM
#include <cub/cub.cuh>
#else
#include <hipcub/hipcub.hpp>
#endif
namespace vllm {
// TODO(woosuk): Further optimize this kernel.
@@ -39,7 +34,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
if (threadIdx.x == 0) {
s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -100,7 +95,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
if (threadIdx.x == 0) {
s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -149,7 +144,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
if (threadIdx.x == 0) {
s_variance = rsqrtf(variance / hidden_size + epsilon);