Allow NVLink traffic for low-latency kernels by default

deep_ep/buffer.py

@@ -32,7 +32,7 @@ class Buffer:
     def __init__(self, group: dist.ProcessGroup,
                  num_nvl_bytes: int = 0, num_rdma_bytes: int = 0,
                  low_latency_mode: bool = False, num_qps_per_rank: int = 12,
-                 allow_nvlink_for_low_latency_mode: bool = False) -> None:
+                 allow_nvlink_for_low_latency_mode: bool = True) -> None:
         """
         Initialize the communication buffer.