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

Apply fixes for CUDA 13 (#24599)

Browse Source 
Signed-off-by: Aidyn-A <aidyn.b.aitzhan@gmail.com>
This commit is contained in:
Aidyn-A
2025-09-17 17:15:42 +04:00
committed by GitHub
parent 9fccd04e30
commit bfe9380161
8 changed files with 47 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
csrc/quantization/compressed_tensors/int8_quant_kernels.cu
View File

@@ -7,17 +7,10 @@
#include <cmath>
#include "../../cub_helpers.h"
#include "../../dispatch_utils.h"
#include "../vectorization_utils.cuh"
#ifndef USE_ROCM
#include <cub/cub.cuh>
#include <cub/util_type.cuh>
#else
#include <hipcub/hipcub.hpp>
#include <hipcub/util_type.hpp>
#endif
static inline __device__ int8_t float_to_int8_rn(float x) {
#ifdef USE_ROCM
static constexpr auto i8_min =
@@ -173,7 +166,7 @@ __global__ void dynamic_scaled_int8_quant_kernel(
});
using BlockReduce = cub::BlockReduce<float, 256>;
__shared__ typename BlockReduce::TempStorage tmp;
float block_max = BlockReduce(tmp).Reduce(thread_max, cub::Max{}, blockDim.x);
float block_max = BlockReduce(tmp).Reduce(thread_max, CubMaxOp{}, blockDim.x);
__shared__ float absmax;
if (tid == 0) {
absmax = block_max;

9
csrc/quantization/fp8/common.cu
View File

@@ -1,15 +1,10 @@
#include "common.cuh"
#include "dispatch_utils.h"
#include "../../cub_helpers.h"
#include "../vectorization_utils.cuh"
#include <c10/cuda/CUDAGuard.h>
#include <ATen/cuda/Exceptions.h>
#ifndef USE_ROCM
#include <cub/cub.cuh>
#else
#include <hipcub/hipcub.hpp>
#endif
namespace vllm {
template <typename scalar_t, typename fp8_type>
@@ -116,7 +111,7 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel_strided(
using BlockReduce = cub::BlockReduce<float, 256>;
__shared__ typename BlockReduce::TempStorage tmp;
const float block_max =
BlockReduce(tmp).Reduce(absmax_val, cub::Max{}, blockDim.x);
BlockReduce(tmp).Reduce(absmax_val, CubMaxOp{}, blockDim.x);
__shared__ float token_scale;
if (tid == 0) {

14
csrc/quantization/fused_kernels/layernorm_utils.cuh
View File

@@ -8,11 +8,7 @@
#include "quantization/utils.cuh"
#include "quant_conversions.cuh"
#ifndef USE_ROCM
#include <cub/cub.cuh>
#else
#include <hipcub/hipcub.hpp>
#endif
#include "../../cub_helpers.h"
namespace vllm {
@@ -36,7 +32,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
ss = BlockReduce(reduceStore).Reduce(ss, cub::Sum{}, blockDim.x);
ss = BlockReduce(reduceStore).Reduce(ss, CubAddOp{}, blockDim.x);
__shared__ float s_rms;
if (threadIdx.x == 0) {
@@ -73,7 +69,7 @@ __device__ void compute_dynamic_per_token_scales(
__shared__ typename BlockReduce::TempStorage reduceStore;
block_absmax_val_maybe =
BlockReduce(reduceStore)
.Reduce(block_absmax_val_maybe, cub::Max{}, blockDim.x);
.Reduce(block_absmax_val_maybe, CubMaxOp{}, blockDim.x);
__shared__ float s_token_scale;
if (threadIdx.x == 0) {
@@ -169,7 +165,7 @@ __device__ void compute_rms(float* rms, scalar_t const* __restrict__ input,
using BlockReduce = cub::BlockReduce<float, 1024>;
__shared__ typename BlockReduce::TempStorage reduceStore;
ss = BlockReduce(reduceStore).Reduce(ss, cub::Sum{}, blockDim.x);
ss = BlockReduce(reduceStore).Reduce(ss, CubAddOp{}, blockDim.x);
__shared__ float s_rms;
if (threadIdx.x == 0) {
@@ -240,7 +236,7 @@ __device__ void compute_dynamic_per_token_scales(
__shared__ typename BlockReduce::TempStorage reduceStore;
block_absmax_val_maybe =
BlockReduce(reduceStore)
.Reduce(block_absmax_val_maybe, cub::Max{}, blockDim.x);
.Reduce(block_absmax_val_maybe, CubMaxOp{}, blockDim.x);
__shared__ float s_token_scale;
if (threadIdx.x == 0) {