Custom All Reduce for Piecewise Cuda Graph (#15356)

Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>
2025-12-25 23:55:15 +08:00
parent b6702d72cf
commit 5c243ba588
5 changed files with 146 additions and 145 deletions
@@ -18,6 +18,7 @@ class CompilationConfig:
        self.split_ops = [
            "sglang.unified_attention_with_output",
            "sglang.gdn_with_output",
+            "sglang.inplace_all_reduce",
        ]

    def add_split_op(self, op: str):