hack nvlink

Signed-off-by: Bill Nell <bnell@redhat.com>

Author: bnellnm

vllm/distributed/device_communicators/all2all.py

@@ -262,7 +262,7 @@ def __init__(self, cpu_group, device):
  self.device = device
  self.handle_cache = Cache()
 
- def get_handle(self, kwargs):
+ def get_handle(self, kwargs, nvlink: bool = True):
  from rose.distributed.torch_group import TorchParallelGroup
  from rose.kernels.efa_all_to_all import EfaAllToAll
 
@@ -285,11 +285,9 @@ def get_handle(self, kwargs):
  ranks=dp_group.ranks,
  )
 
- kwargs["nvlink"] = True
-
  kwargs["nets_per_gpu"] = _nets_per_gpu()
  kwargs["dp_group"] = tp_group
- kwargs["node_group"] = global_group if kwargs.get("nvlink", False) else None
+ kwargs["node_group"] = global_group if nvlink else None
  kwargs["global_group"] = global_group
  kwargs["device"] = self.device