[BugFix] Fixes Qwen3-Next enable nz accuracy problem (#4058)

### What this PR does / why we need it?
- Fixes Qwen3-Next enable nz accuracy problem

### Does this PR introduce _any_ user-facing change?
N/A


- vLLM version: v0.11.0
- vLLM main:
83f478bb19

---------

Signed-off-by: Icey <1790571317@qq.com>
Signed-off-by: wxsIcey <1790571317@qq.com>

vllm_ascend/utils.py

@@ -59,6 +59,7 @@ _MIN_DP_BUFFER_SIZE = 50
 _IS_MOE_MODEL = None
 _ENABLE_SP = None
 _HAS_LAYER_IDX = None
+_ENABLE_NZ = None
 
 
 def is_310p():
@@ -69,8 +70,14 @@ def is_310p():
     return _IS_310P
 
 
-def is_enable_nz():
-    return envs_ascend.VLLM_ASCEND_ENABLE_NZ
+def is_enable_nz(vllm_config: Optional[VllmConfig] = None) -> bool:
+    global _ENABLE_NZ
+    if _ENABLE_NZ is None:
+        if not vllm_config:
+            raise ValueError(
+                "vllm_config must be provided when _ENABLE_NZ is None")
+        _ENABLE_NZ = envs_ascend.VLLM_ASCEND_ENABLE_NZ and vllm_config.model_config.hf_config.model_type != "qwen3_next"
+    return _ENABLE_NZ
 
 
 def sleep_mode_enabled():

vllm_ascend/worker/worker_v1.py

@@ -87,6 +87,7 @@ class NPUWorker(WorkerBase):
         # register patch for vllm
         from vllm_ascend.utils import adapt_patch
         adapt_patch()
+        is_enable_nz(vllm_config)
         # Register ops when worker init.
         from vllm_ascend import ops
         ops.register_dummy_fusion_op()