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

[Bugfix]Fix moe error when sp chunked the hidden_states (#4212)

Browse Source 
### What this PR does / why we need it?
Fix moe error when sp chunked the hidden_states by disabling sp by a hacky way

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
This commit is contained in:
weiguihua2
2025-11-17 22:55:17 +08:00
committed by GitHub
parent 3677202594
commit da1cd9c7ca
1 changed files with 4 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm_ascend/platform.py
View File

@@ -310,7 +310,10 @@ class NPUPlatform(Platform):
if parallel_config and parallel_config.worker_cls == "auto": if parallel_config and parallel_config.worker_cls == "auto":
# TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm. # TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm.
if vllm_version_is("0.11.0"):
os.environ["VLLM_ALL2ALL_BACKEND"] = "flashinfer_all2allv" os.environ["VLLM_ALL2ALL_BACKEND"] = "flashinfer_all2allv"
else:
parallel_config.all2all_backend = "flashinfer_all2allv"
if ascend_config.torchair_graph_config.enabled or ascend_config.enable_shared_expert_dp: if ascend_config.torchair_graph_config.enabled or ascend_config.enable_shared_expert_dp:
parallel_config.worker_cls = "vllm_ascend.torchair.torchair_worker.NPUTorchairWorker" parallel_config.worker_cls = "vllm_ascend.torchair.torchair_worker.NPUTorchairWorker"
else: else: