fix: Defer SOC version initialization to avoid NPU lazy init issues

vllm_npu/utils.py

@@ -67,11 +67,13 @@ _IS_EAGLE_MODE = None
 def is_310p():
     global _IS_310P
     if _IS_310P is None:
-        try:
-            soc_version = torch_npu.npu.get_soc_version()
-            # 310P soc_version range: 200-209
-            _IS_310P = 200 <= soc_version <= 209
-        except Exception:
+        # Check if SOC version is already known from init_ascend_soc_version()
+        if _ascend_soc_version is not None:
+            _IS_310P = False  # 310P is not A2 or A3
+        else:
+            # Avoid calling torch_npu.npu.get_soc_version() here as it
+            # triggers NPU lazy init which breaks forked subprocesses.
+            # Default to False; will be updated after init_device().
             _IS_310P = False
     return _IS_310P