reformat

Author: shihaobai

lightllm/distributed/communication_op.py

@@ -38,6 +38,7 @@
 vllm_reduce = None
 logger = init_logger(__name__)
 
+
 @contextmanager
 def lightllm_capture_graph():
     if vllm_reduce is not None:
@@ -47,6 +48,7 @@ def lightllm_capture_graph():
         yield
     pass
 
+
 def _all_reduce(input_, op=ReduceOp.SUM, group=None, async_op=False):
     if op != ReduceOp.SUM or async_op:
         original_all_reduce(input_, op, group, async_op)
@@ -58,6 +60,7 @@ def _all_reduce(input_, op=ReduceOp.SUM, group=None, async_op=False):
                 return
         original_all_reduce(input_, op, group, async_op)
 
+
 def set_custom_reduce():
     global vllm_reduce
     global device_group

lightllm/distributed/custom_all_reduce.py

@@ -31,10 +31,10 @@
 from vllm.platforms import current_platform
 from vllm.utils import cuda_device_count_stateless
 from lightllm.common.basemodel.layer_infer.cache_tensor_manager import g_cache_manager
+
 ops.meta_size()
 custom_ar = True
 
-
 logger = init_logger(__name__)
 
 
@@ -67,7 +67,6 @@ def __init__(self, group: ProcessGroup, device: Union[int, str, torch.device], m
             # disable because of missing custom allreduce library
             # e.g. in a non-cuda environment
             return
-
         self.group = group
         assert dist.get_backend(group) != dist.Backend.NCCL, "CustomAllreduce should be attached to a non-NCCL group."