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

[Performance]Optimize DeepSeekOCR2 RelPosAttention and CustomQwen2Decoder (#7737)

Browse Source 
### What this PR does / why we need it?
Optimize DeepSeekOCR2 RelPosAttention and CustomQwen2Decoder and add doc
for DeepSeekOCR2.md

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
- vllm 0.18.0
- vllm-ascend main

1. _create_custom_4d_mask during 141ms49us620ns -->
_create_npu_optimized_mask during 1ms227us780ns
2. convd2d : 27ms --> matmul <1ms
3. relposattention：sdpa->prompt_flash_attention

---------

Signed-off-by: Wangbei25 <wangbei41@huawie.com>
Signed-off-by: Wangbei25 <wangbei41@huawei.com>
Co-authored-by: Wangbei25 <wangbei41@huawie.com>
This commit is contained in:
Wangbei25
2026-03-31 14:49:29 +08:00
committed by GitHub
parent 2a0a588311
commit 4f259d4fd8
4 changed files with 432 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -49,3 +49,4 @@ import vllm_ascend.patch.worker.patch_deepseek_mtp # noqa
import vllm_ascend.patch.worker.patch_v2.patch_input_batch # noqa
import vllm_ascend.patch.worker.patch_v2.patch_model_state # noqa
import vllm_ascend.patch.worker.patch_v2.patch_block_table # noqa
import vllm_ascend.patch.worker.patch_deepencoder2 # noqa