[platform] support pytorch custom op pluggable (vllm-project#11328)

Signed-off-by: wangxiyuan <[email protected]>
Signed-off-by: Bowen Wang <[email protected]>

Author: wangxiyuan

vllm/model_executor/custom_op.py

@@ -57,6 +57,11 @@ def forward_hpu(self, *args, **kwargs):
         # PyTorch-native implementation.
         return self.forward_native(*args, **kwargs)
 
+    def forward_oot(self, *args, **kwargs):
+        # By default, we assume that OOT ops are compatible with the
+        # PyTorch-native implementation.
+        return self.forward_native(*args, **kwargs)
+
     def dispatch_forward(self):
         # NOTE(woosuk): Here we assume that vLLM was built for only one
         # specific backend. Currently, we do not support dynamic dispatching.
@@ -81,6 +86,8 @@ def dispatch_forward(self):
             return self.forward_tpu
         elif current_platform.is_xpu():
             return self.forward_xpu
+        elif current_platform.is_out_of_tree():
+            return self.forward_oot
         else:
             return self.forward_cuda
 

vllm/platforms/interface.py

@@ -45,6 +45,7 @@ class PlatformEnum(enum.Enum):
     CPU = enum.auto()
     NEURON = enum.auto()
     OPENVINO = enum.auto()
+    OOT = enum.auto()
     UNSPECIFIED = enum.auto()
 
 
@@ -107,6 +108,9 @@ def is_neuron(self) -> bool:
     def is_openvino(self) -> bool:
         return self._enum == PlatformEnum.OPENVINO
 
+    def is_out_of_tree(self) -> bool:
+        return self._enum == PlatformEnum.OOT
+
     def is_cuda_alike(self) -> bool:
         """Stateless version of :func:`torch.cuda.is_available`."""
         return self._enum in (PlatformEnum.CUDA, PlatformEnum.ROCM)