Add pyhccl (#503)

This is the first step to support trl vllm serve on Ascend NPU
https://github.com/vllm-project/vllm-ascend/issues/459.
This PR can work properly only when
https://github.com/vllm-project/vllm/pull/16464 is merged into vLLM.

---------

Signed-off-by: hzji210@gmail.com <hzji210@gmail.com>

vllm_ascend/utils.py

@@ -17,8 +17,11 @@
 # limitations under the License.
 #
 import torch
+import torch_npu  # noqa: F401
 from vllm.logger import logger
 
+import vllm_ascend.envs as envs
+
 
 def try_register_lib(lib_name: str, lib_info: str = ""):
     import importlib
@@ -33,6 +36,28 @@ def try_register_lib(lib_name: str, lib_info: str = ""):
         pass
 
 
+def find_hccl_library() -> str:
+    """
+    We either use the library file specified by the `HCCL_SO_PATH`
+    environment variable, or we find the library file brought by PyTorch.
+    After importing `torch`, `libhccl.so` can be
+    found by `ctypes` automatically.
+    """
+    so_file = envs.HCCL_SO_PATH
+
+    # manually load the hccl library
+    if so_file:
+        logger.info("Found hccl from environment variable HCCL_SO_PATH=%s",
+                    so_file)
+    else:
+        if torch.version.cann is not None:
+            so_file = "libhccl.so"
+        else:
+            raise ValueError("HCCL only supports Ascend NPU backends.")
+        logger.info("Found hccl from library %s", so_file)
+    return so_file
+
+
 _current_stream = None