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

[Feat][BugFix]Support the Qwen3-Next-80B-A3B-Instruct quantization model&Fix the NZ issue (#4245)

Browse Source 
### What this PR does / why we need it?
Support the Qwen3-Next-80B-A3B-Instruct quantization model and Fix the
NZ issue. Triton kernel doesn't support data format nz, thus we skip
converting weight to nz on layer `conv1d`

- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: IncSec <1790766300@qq.com>
This commit is contained in:
InSec
2025-11-21 10:42:56 +08:00
committed by GitHub
parent cbb27feaf2
commit 5a4e8cdeba
10 changed files with 39 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/ut/models/test_qwen2_5_vl.py
View File

@@ -1,5 +1,3 @@
from unittest.mock import patch
import pytest
import torch
import torch.nn.functional as F
@@ -367,7 +365,6 @@ class TestAscendQwen2_5_VisionTransformer(PytestBase):
res = attention.pad_qkv_bias(torch.rand((300)))
assert res.shape[0] == 384
@patch('vllm_ascend.utils._ENABLE_NZ', True)
def test_pad_qkv_weight(self, mocker: MockerFixture):
attention = self.init_vision_transformer(mocker)
mocker.patch("torch.nn.Module.__setattr__")
@@ -380,7 +377,6 @@ class TestAscendQwen2_5_VisionTransformer(PytestBase):
res = attention.pad_qkv_weight(torch.rand((300, 300)))
assert res.shape == (384, 300)
@patch('vllm_ascend.utils._ENABLE_NZ', True)
def test_pad_proj_weight(self, mocker: MockerFixture):
attention = self.init_vision_transformer(mocker)
mocker.patch("torch.nn.Module.__setattr__")