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

3
vllm_ascend/ops/fused_moe.py
View File

@@ -1027,8 +1027,9 @@ class AscendFusedMoE(FusedMoE):
ascend_config = get_ascend_config()
self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
# NOTE: multistream only effective when `VLLM_ENABLE_MC2` is on
self.enable_multistream_shared_expert = \
ascend_config.torchair_graph_config.enable_multistream_shared_expert
ascend_config.torchair_graph_config.enable_multistream_shared_expert and VLLM_ENABLE_MC2
if self.scoring_func != "softmax" and not self.use_grouped_topk:
raise ValueError("Only softmax scoring function is supported for "