[Bug]: qwen3-vl-235B-bf16 FULL_DECODE_ONLY + VLLM_ASCEND_ENABLE_NZ=1 压测报错问题

[Levi-JQ]

Your current environment

The output of `python collect_env.py`

Your output of above commands here

🐛 Describe the bug

报错信息如下：

由于w8a8权重同配置下没有报错，所以怀疑和NZ格式有关
验证和NZ格式有关：
（1）v0.11.0-dev分支加了这个PR #4495 可以解决
（2）验证使用export VLLM_ASCEND_ENABLE_NZ=0 也可以解决

怀疑是FULL_DECODE_ONLY和VLLM_ASCEND_ENABLE_NZ=1有冲突，报错位置在PA算子处

[ponyioy]

@Levi-JQ 请问w8a8用哪个版本的镜像和什么启动参数可以正常运行，thinking可以吗？

[leo-pony]

Yes your are right. Currently for fp16 and bf 16 has changed to VLLM_ASCEND_ENABLE_NZ disable default. See PR: #4495

[Levi-JQ]

Yes your are right. Currently for fp16 and bf 16 has changed to VLLM_ASCEND_ENABLE_NZ disable default. See PR: #4495

I konw this PR. However, what is the root cause? Enabling NZ improves the performance of bf16. When can this issue be resolved?

[leo-pony]

@Levi-JQ
Some operators still have bugs for the NZ format, which are currently being addressed. We will update to NZ as soon as they are resolved.