[ModelRunner V2] Fix Mamba2 crash on non-spec-decode (#47428)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Jul 2, 2026 at 15:05 UTC
3e158ae62d1c227004fa9f702a51126e58ebbcb2
1 file changed
+1
-1
vllm/v1/worker/gpu/model_states/mamba_hybrid.py
+1
-1
@@ -246,7 +246,7 @@ class MambaHybridModelState(DefaultModelState):
246
# compute them during actual (non-capture) forward execution.
247
num_accepted_tokens = None
248
num_decode_draft_tokens_cpu = None
249
- if not for_capture:
249
+ if not for_capture and self.vllm_config.num_speculative_tokens > 0:
250
num_accepted_tokens = self.num_accepted_tokens_gpu.new_ones(num_reqs)
251
num_accepted_tokens[: input_batch.num_reqs] = self.num_accepted_tokens_gpu[
252
input_batch.idx_mapping