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

[BugFix] Fix mlapo accuracy problem related with weight processing. (#3850)

Browse Source 
This PR fixes a mlapo accuracy problem related with weight processing.
Furthermore, add back mlapo related e2e test with quantized deepseek
model.


- vLLM version: v0.11.0rc3
- vLLM main:
83f478bb19

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-10-30 00:34:55 +08:00
committed by GitHub
parent adadd50613
commit dc960e798e
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/attention/mla_v1.py
View File

@@ -826,9 +826,9 @@ class AscendMLAImpl(MLAAttentionImpl):
..., self.q_lora_rank:].contiguous()
q_a_proj_wt = self.fused_qkv_a_proj.weight.data[
..., :self.q_lora_rank].contiguous()
kv_a_proj_wt = kv_a_proj_wt.contiguous()
kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
kv_a_proj_wt = trans_rope_weight(kv_a_proj_wt, self.qk_rope_head_dim)
kv_a_proj_wt = kv_a_proj_wt.contiguous()
kv_a_proj_wt = kv_a_proj_wt.t().contiguous()
wd_qkv = torch.cat((kv_a_proj_wt, q_a_proj_wt), dim=-1)
wd_qkv = wd_qkv.t().contiguous()
wd_qkv = transdata(wd_qkv,