[ModelRunner V2] Fix Mamba2 crash on non-spec-decode (#47428)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
2026-07-02 07:05:16 -07:00
parent a2f713002d
commit 3e158ae62d
1 changed files with 1 additions and 1 deletions
@@ -246,7 +246,7 @@ class MambaHybridModelState(DefaultModelState):
        # compute them during actual (non-capture) forward execution.
        num_accepted_tokens = None
        num_decode_draft_tokens_cpu = None
-        if not for_capture:
+        if not for_capture and self.vllm_config.num_speculative_tokens > 0:
            num_accepted_tokens = self.num_accepted_tokens_gpu.new_ones(num_reqs)
            num_accepted_tokens[: input_batch.num_reqs] = self.num_accepted_tokens_gpu[
                input_batch.idx_mapping