feat: support compatibility between MTP and two-batch-overlap (#7225)

Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>
2025-06-27 01:10:27 -07:00
parent 1b95162008
commit 41650b0d70
7 changed files with 325 additions and 38 deletions
--- a/python/sglang/srt/model_executor/model_runner.py
+++ b/python/sglang/srt/model_executor/model_runner.py
@@ -1039,7 +1039,7 @@ class ModelRunner:

    def init_attention_backend(self):
        """Init attention kernel backend."""
-        if self.server_args.enable_two_batch_overlap:
+        if self.server_args.enable_two_batch_overlap and not self.is_draft_worker:
            self.attn_backend = TboAttnBackend.init_new(self._get_attention_backend)
        else:
            self.attn_backend = self._get_attention_backend()