feat: Implement the NPU platform plugin for vLLM, including platform registration, device management, custom operations, and configuration adaptation.

vllm_npu/__init__.py

@@ -15,16 +15,15 @@ def register():
     from vllm_npu.cuda_compat import _patch_cuda_to_npu
     _patch_cuda_to_npu()
 
-    # Register NPU custom ops with vLLM's CustomOp dispatch so that
-    # ops like SiluAndMul, RMSNorm, RotaryEmbedding use NPU kernels
-    # instead of falling back to CUDA (which would produce garbage).
-    _register_npu_ops()
-
     return "vllm_npu.platform.NPUPlatform"
 
 
-def _register_npu_ops():
-    """Register Ascend NPU op overrides with vLLM's CustomOp system."""
+def register_npu_ops():
+    """Register Ascend NPU op overrides with vLLM's CustomOp system.
+
+    Must be called AFTER the platform is established (e.g., during
+    worker init or check_and_update_config), NOT during register().
+    """
     from vllm.model_executor.custom_op import CustomOp
 
     from vllm_npu.ops.activation import AscendSiluAndMul