Merge branch 'v0.8.4-dev-wm' into 'v0.8.4-dev'

[fix]修复多卡eager模式精度问题

See merge request dcutoolkit/deeplearing/vllm!102

vllm/utils.py

@@ -1028,8 +1028,11 @@ def current_stream() -> torch.cuda.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()
+
+        # fix computational precision issue in eager mode 
+        # _current_stream = torch.cuda.Stream() if current_platform.is_rocm(
+        # ) else torch.cuda.current_stream()
+        _current_stream = torch.cuda.current_stream()
     return _current_stream