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

Revert "[refactor]support gatingtopk operator generalization (#4356)" (#4873)

Browse Source 
This reverts commit c4a11a745a.

ops npu_gating_top_k caused Qwen3-30B precision problem, so revert it.

Signed-off-by: 1092626063 <1092626063@qq.com>
This commit is contained in:
1092626063
2025-12-10 15:45:20 +08:00
committed by GitHub
parent 9a144bc7be
commit ceadc2788d
4 changed files with 56 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/ascend_forward_context.py
View File

@@ -96,7 +96,6 @@ def set_ascend_forward_context(
ep_size = (get_ep_group().world_size if
vllm_config.parallel_config.enable_expert_parallel else 1)
# fused_moe_state is used in torchair, it will be deleted along with torchair
is_deepseek_v3_r1 = hasattr(
vllm_config.model_config.hf_config, 'n_routed_experts'
) and vllm_config.model_config.hf_config.n_routed_experts == 256