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

Fix CUDA kernel index data type in vllm/csrc/quantization/fused_kernels/layernorm_utils.cuh +10 (#15159)

Browse Source 
Signed-off-by: Lu Fang <lufang@fb.com>
Co-authored-by: Richard Barnes <rbarnes@meta.com>
This commit is contained in:
Lu Fang
2025-03-20 19:01:11 -07:00
committed by GitHub
parent 0cfe7d386d
commit d3ccbd6350
10 changed files with 124 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
csrc/quantization/gguf/moe.cuh
View File

@@ -19,10 +19,10 @@ static __device__ __forceinline__ void moe_q(
const int ncols_dst = ncols_y * top_k;
const int row_dst_0 = blockIdx.x * mmq_y;
const auto row_dst_0 = blockIdx.x * mmq_y;
const int& row_x_0 = row_dst_0;
const int col_dst_0 = blockIdx.y * mmq_x;
const auto col_dst_0 = blockIdx.y * mmq_x;
int token_offs[mmq_x / nwarps];
for (int i = 0; i < mmq_x; i += nwarps) {
@@ -56,7 +56,7 @@ static __device__ __forceinline__ void moe_q(
const int n_per_r = ((qk * blocks_per_warp) / qr);
#pragma unroll
for (int ir = 0; ir < qr && ib0 * qk + ir * n_per_r < ncols_x; ++ir) {
const int kqs = ir * WARP_SIZE_GGUF + threadIdx.x;
const auto kqs = ir * WARP_SIZE_GGUF + threadIdx.x;
const int kbxd = kqs / QI8_1;
#pragma unroll
@@ -73,7 +73,7 @@ static __device__ __forceinline__ void moe_q(
}
if (threadIdx.x < n_per_r / QK8_1) {
const int kby = threadIdx.x % (WARP_SIZE_GGUF / QI8_1);
const auto kby = threadIdx.x % (WARP_SIZE_GGUF / QI8_1);
const int col_y_eff = token_offs[threadIdx.y] / top_k;
const int block_x =
ib0 * (qk / QK8_1) + ir * (WARP_SIZE_GGUF / QI8_1) + kby;
@@ -119,7 +119,7 @@ static __device__ __forceinline__ void moe_q(
#pragma unroll
for (int i = 0; i < mmq_y; i += WARP_SIZE_GGUF) {
const int row_dst = row_dst_0 + threadIdx.x + i;
const auto row_dst = row_dst_0 + threadIdx.x + i;
if (row_dst >= nrows_dst) {
continue;
}