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

[Doc] Polish sgl-kernel readme for cu126 build error (#8704)

Browse Source
This commit is contained in:
PGFLMG
2025-08-02 17:03:07 +08:00
committed by GitHub
parent 4ca43b061c
commit ac6962ccd6
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
sgl-kernel/README.md
View File

@@ -58,7 +58,15 @@ And if you build the sgl-kernel with cmake, you need to add `CMAKE_BUILD_PARALLE
CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) python -m uv build --wheel -Cbuild-dir=build --color=always . CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) python -m uv build --wheel -Cbuild-dir=build --color=always .
``` ```
### FlashAttention on Hopper ### ⚠️ Compilation Issue with `sgl-kernel` and CUDA 12.6
When compiling `sgl-kernel` with FlashAttention on a Hopper GPU using CUDA 12.6, you may encounter a segmentation fault:
```bash
kernel/build/_deps/repo-flash-attention-src/hopper/instantiations/flash_fwd_hdimall_bf16_paged_softcap_sm90.cu -o CMakeFiles/flash_ops.dir/_deps/repo-flash-attention-src/hopper/instantiations/flash_fwd_hdimall_bf16_paged_softcap_sm90.cu.o
Segmentation fault (core dumped)
```
⚠️ **Note**: To ensure that FlashAttention compiles correctly on Hopper GPU Architecture(sm90), it is strongly [recommended](https://github.com/Dao-AILab/flash-attention/issues/1453) to use: ⚠️ **Note**: To ensure that FlashAttention compiles correctly on Hopper GPU Architecture(sm90), it is strongly [recommended](https://github.com/Dao-AILab/flash-attention/issues/1453) to use:
- nvcc version: 12.6 - nvcc version: 12.6
- ptxas version: 12.8 - ptxas version: 12.8