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

[bugfix] layerwise D first plan (#3866)

Browse Source 
### What this PR does / why we need it?
Refactored the layerwise code to send to the D node first, preventing
P-node hangs due to communication timeouts when DP > 1.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
By ci

- vLLM version: v0.11.0
- vLLM main:
83f478bb19

---------

Signed-off-by: wangxiaoteng <wangxiaoteng@huawei.com>
Signed-off-by: liziyu <liziyu16@huawei.com>
Co-authored-by: liziyu <liziyu16@huawei.com>
This commit is contained in:
wangxiaoteng888
2025-10-30 22:20:34 +08:00
committed by GitHub
parent 627f20ce26
commit 2c291bc63f
4 changed files with 963 additions and 1354 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
vllm_ascend/distributed/parallel_state.py
View File

@@ -96,7 +96,8 @@ def init_ascend_model_parallel(parallel_config: ParallelConfig, ):
parallel_config.data_parallel_size, num_head_replica, -1,
alltoall_group_size
) # [DP_size, num_head_replica, num_alltoall_group, alltoall_group_size]
group_ranks = group_ranks.view(-1, alltoall_group_size).unbind(0)
group_ranks = group_ranks.reshape(-1,
alltoall_group_size).unbind(0)
group_ranks = [x.tolist() for x in group_ranks]
local_rank = get_world_group().local_rank
num = next(