[ROCm] Avoid using the default stream on ROCm (vllm-project#13238)

Signed-off-by: Gregory Shtrasberg <[email protected]>

Author: gshtras

vllm/utils.py

@@ -942,11 +942,16 @@ def current_stream() -> torch.cuda.Stream:
     the underlying hypothesis is that we do not call `torch._C._cuda_setStream`
     from C/C++ code.
     """
+    from vllm.platforms import current_platform
     global _current_stream
     if _current_stream is None:
         # when this function is called before any stream is set,
         # we return the default stream.
-        _current_stream = torch.cuda.current_stream()
+        # On ROCm using the default 0 stream in combination with RCCL
+        # is hurting performance. Therefore creating a dedicated stream
+        # per process
+        _current_stream = torch.cuda.Stream() if current_platform.is_rocm(
+        ) else torch.cuda.current_stream()
     return _current_stream