[MTP] optimize mtp infer speed (#2840)

2026-04-24 01:29:57 +08:00 · 2025-07-14 19:50:22 +08:00
parent 4c7b8bc458
commit 7cdd8d290d
6 changed files with 253 additions and 24 deletions
@@ -497,6 +497,8 @@ class MTPProposer(Proposer):
            self.main_model_inputs["seq_lens_encoder"],
            self.max_draft_token_num,
        )
+        if isinstance(target_hidden_states, list):
+            target_hidden_states = target_hidden_states[0]

        return target_hidden_states