[None][fix] disable async pp send for ray cases. (#9959)

Signed-off-by: Yuxian Qiu <[email protected]>

Author: yuxianq

jenkins/L0_MergeRequest.groovy

@@ -712,6 +712,7 @@ def getMultiGpuFileChanged(pipeline, testFilter, globalVars)
         "tensorrt_llm/_torch/compilation/patterns/ub_allreduce.py",
         "tensorrt_llm/_torch/custom_ops/torch_custom_ops.py",
         "tensorrt_llm/_torch/custom_ops/userbuffers_custom_ops.py",
+        "tensorrt_llm/_torch/distributed/",
         "tensorrt_llm/_torch/models/modeling_llama.py",
         "tensorrt_llm/_torch/models/modeling_qwen3_next.py",
         "tensorrt_llm/_torch/modules/fused_moe/",

tensorrt_llm/_torch/distributed/communicator.py

@@ -856,6 +856,13 @@ def _global_to_local_rank(self, global_rank: int):
     def direct_send(self, tensor: torch.Tensor, dest: int):
         self.pg.send([tensor], self._global_to_local_rank(dest), tag=0).wait()
 
+    # TODO: support async pp send for PPCommTorch
+    def send(self, tensor: torch.Tensor, dest: Optional[int] = None):
+        if dest is None:
+            dest = self.mapping.next_pp_rank()
+
+        self.pg.send([tensor], self._global_to_local_rank(dest), tag=0).wait()
+
     def recv(self, tensor: torch.Tensor, src: Optional[int] = None):
         if src is None:
             src = self.mapping.prev_pp_rank()