support 1 shot allreduce in 1-node and 2-node using mscclpp (#6277)

2025-06-05 13:11:24 +08:00
parent 4474eaf552
commit 8e3797be1c
20 changed files with 2177 additions and 12 deletions
--- a/python/sglang/srt/layers/dp_attention.py
+++ b/python/sglang/srt/layers/dp_attention.py
@@ -98,11 +98,12 @@ def initialize_dp_attention(
        ],
        local_rank,
        torch.distributed.get_backend(tp_group.device_group),
-        SYNC_TOKEN_IDS_ACROSS_TP,
-        False,
-        False,
-        False,
-        False,
+        use_pynccl=SYNC_TOKEN_IDS_ACROSS_TP,
+        use_pymscclpp=False,
+        use_custom_allreduce=False,
+        use_hpu_communicator=False,
+        use_xpu_communicator=False,
+        use_npu_communicator=False,
        group_name="attention_tp",
    )