[Feature] support custom all-reduce (#2758)

* [Feature] support custom all-reduce

* add vllm adapted

fastdeploy/config.py

@@ -207,6 +207,8 @@ class ParallelConfig:
     guided_decoding_backend: str = None
     # disable any whitespace for guided decoding
     disable_any_whitespace: bool = True
+    # enable the custom all-reduce kernel and fall back to NCCL(dist.all_reduce).
+    enable_custom_all_reduce: str = "store_true"
 
 
 @dataclass