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

[bugfix] fix deeepseek accuracy (#1118)

Browse Source 
### What this PR does / why we need it?
fix deeepseek accuracy in mix-parallel case.


Signed-off-by: zzzzwwjj <1183291235@qq.com>
This commit is contained in:
zzzzwwjj
2025-06-07 21:11:36 +08:00
committed by GitHub
parent c8742146d3
commit f1543d5e0d
3 changed files with 23 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm_ascend/platform.py
View File

@@ -142,7 +142,11 @@ class NPUPlatform(Platform):
# NOTE: When enable_expert_parallel is True, we follow vLLM convention:
# ep_size = world_size, which means expert_tensor_parallel_size must be 1
if ascend_config.expert_tensor_parallel_size > 0 and not parallel_config.enable_expert_parallel:
if parallel_config.enable_expert_parallel:
parallel_config.expert_tensor_parallel_size = 1
# NOTE: When enable_expert_parallel is False and param `asceend_config.expert_tensor_parallel_size`
# is configured, use ascend_config
elif ascend_config.expert_tensor_parallel_size > 0:
parallel_config.expert_tensor_parallel_size = ascend_config.expert_tensor_parallel_size
# Calculate expert parallel size based on world size