[Core] manage nccl via a pypi package & upgrade to pt 2.2.1 (#3805)

vllm/model_executor/parallel_utils/pynccl.py

@@ -21,6 +21,7 @@
 
 import ctypes
 import datetime
+import glob
 import os
 
 # ===================== import region =====================
@@ -34,18 +35,27 @@ logger = init_logger(__name__)
 
 so_file = os.environ.get("VLLM_NCCL_SO_PATH", "")
 
+# check if we have vllm-managed nccl
+vllm_nccl_path = None
+if torch.version.cuda is not None:
+    cuda_major = torch.version.cuda.split(".")[0]
+    path = os.path.expanduser(
+        f"~/.config/vllm/nccl/cu{cuda_major}/libnccl.so.*")
+    files = glob.glob(path)
+    vllm_nccl_path = files[0] if files else None
+
 # manually load the nccl library
 if so_file:
     logger.info(
         f"Loading nccl from environment variable VLLM_NCCL_SO_PATH={so_file}")
 else:
     if torch.version.cuda is not None:
-        so_file = "libnccl.so.2"
+        so_file = vllm_nccl_path or "libnccl.so.2"
     elif torch.version.hip is not None:
         so_file = "librccl.so.1"
     else:
         raise ValueError("NCCL only supports CUDA and ROCm backends.")
-    logger.debug(f"Loading nccl from library {so_file}")
+    logger.info(f"Loading nccl from library {so_file}")
 
 try:
     nccl = ctypes.CDLL(so_file)