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

[MTP][Bugfix] Fix GLM5-W8A8 precision issues caused by rotary quant MTP weights (#7139)

Browse Source 
### What this PR does / why we need it?
When GLM5 target model uses rotary quant, the final hidden states passes
to MTP need to do an extra rotary.

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Wangbingjie <wangbj1207@126.com>
Signed-off-by: wangbj127 <256472688+wangbj127@users.noreply.github.com>
This commit is contained in:
wangbj127
2026-03-12 20:01:24 +08:00
committed by GitHub
parent de93790d08
commit 0c659e91ed
3 changed files with 96 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/patch/worker/__init__.py
View File

@@ -44,3 +44,4 @@ import vllm_ascend.patch.worker.patch_npugraph_ex_triton # noqa
import vllm_ascend.patch.worker.patch_kimi_k25 # noqa
import vllm_ascend.patch.worker.patch_draft_quarot # noqa
import vllm_ascend.patch.worker.patch_cudagraph # noqa
import vllm_ascend.patch.worker.patch_deepseek_mtp # noqa