EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[bugifx] QWen-1M context support[2/3] using current cuda stream in the DCA's kernel for bugfix. (#8611)

Browse Source 
Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
Co-authored-by: sa-buc <linzhu.ht@w32d09270.cloud.sqa.na131>
This commit is contained in:
Tao He
2025-07-31 22:41:39 +08:00
committed by GitHub
parent 016fd25127
commit 5d15fb8c9d
1 changed files with 5 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
sgl-kernel/csrc/attention/vertical_slash_index.cu
View File

@@ -3,6 +3,7 @@
// This file is for blocksparse attention utils cuda kernel.
#include <assert.h>
#include <c10/cuda/CUDAStream.h>
#include <cuda.h>
#include <torch/all.h>
@@ -176,7 +177,8 @@ void convert_vertical_slash_indexes_64x64(
const dim3 dimBlock((int32_t)N_THREADS);
const dim3 dimGrid(
(int32_t)N_HEADS, (int32_t)BATCH_SIZE, ((int32_t)N_ROWS + (int32_t)N_THREADS - 1) / (int32_t)N_THREADS);
convert_vertical_slash_indexes_kernel<<<dimGrid, dimBlock>>>(
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
convert_vertical_slash_indexes_kernel<<<dimGrid, dimBlock, 0, stream>>>(
q_seqlens,
kv_seqlens,
vertical_indexes,
@@ -393,7 +395,8 @@ void convert_vertical_slash_indexes_64x64_mergehead(
const int N_THREADS = 64;
const dim3 dimBlock(N_THREADS);
const dim3 dimGrid(N_HEADS, BATCH_SIZE, (N_ROWS + N_THREADS - 1) / N_THREADS);
convert_vertical_slash_indexes_kernel_mergehead<<<dimGrid, dimBlock>>>(
cudaStream_t stream = at::cuda::getCurrentCUDAStream();
convert_vertical_slash_indexes_kernel_mergehead<<<dimGrid, dimBlock, 0, stream>>>(
q_seqlens,
kv_seqlens,
vertical_indexes,