EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Feat] Flash comm allgher ep (#3334)

Browse Source 
Support flash comm v1(Sequence Parallelism) for Allgather EP.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: realliujiaxu <realliujiaxu@163.com>
Co-authored-by: zhaozx-cn <zhaozx2116@163.com>
This commit is contained in:
realliujiaxu
2025-10-15 19:36:32 +08:00
committed by GitHub
parent 8abe517870
commit f69a83b7ba
15 changed files with 283 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/envs.py
View File

@@ -133,8 +133,8 @@ env_variables: Dict[str, Callable[[], Any]] = {
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE", '0'))),
# Whether to enable FlashComm optimization when tensor parallel is enabled.
# This feature will get better performance when concurrency is large.
"VLLM_ASCEND_ENABLE_FLASHCOMM":
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", '0'))),
"VLLM_ASCEND_ENABLE_FLASHCOMM1":
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM1", '0'))),
# Whether to enable MLP weight prefetch, only used in small concurrency.
"VLLM_ASCEND_ENABLE_PREFETCH_MLP":
lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_PREFETCH_MLP", '0'))),