Support torchrun and SPMD-style offline inference (#12071)

Signed-off-by: youkaichao <youkaichao@gmail.com>
2025-01-16 19:58:53 +08:00
parent dd7c9ad870
commit bf53e0c70b
14 changed files with 248 additions and 30 deletions
--- a/vllm/v1/executor/multiproc_executor.py
+++ b/vllm/v1/executor/multiproc_executor.py
@@ -246,7 +246,7 @@ class WorkerProc:
        ready_path: str,
    ):
        self.rank = rank
-        wrapper = WorkerWrapperBase(vllm_config=vllm_config, rank=rank)
+        wrapper = WorkerWrapperBase(vllm_config=vllm_config, rpc_rank=rank)
        # TODO: move `init_worker` to executor level as a collective rpc call
        all_kwargs: List[Dict] = [
            {} for _ in range(vllm_config.parallel_config.world_size)