vllm-ascend support Ascend950 with Qwen dense model. (#4228)

### What this PR does / why we need it?
vllm-ascend support Ascend950 with Qwen dense model
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?


- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

---------

Signed-off-by: wangyao <iwangyao@outlook.com>
Co-authored-by: weijinqian0 <1184188277@qq.com>
This commit is contained in:
wangyao-i
2025-12-12 15:50:57 +08:00
committed by GitHub
parent 716c4dacfe
commit 0983c5510a
4 changed files with 59 additions and 12 deletions

View File

@@ -410,7 +410,8 @@ class AscendMRotaryEmbedding(MRotaryEmbedding):
query: torch.Tensor,
key: torch.Tensor,
):
if self.mrope_section != [16, 24, 24]:
if self.mrope_section != [16, 24, 24] or \
get_ascend_device_type() == AscendDeviceType._910_95:
return super().forward_oot(positions, query, key)
import torch_npu