diff --git a/vllm/v1/worker/gpu/model_states/mamba_hybrid.py b/vllm/v1/worker/gpu/model_states/mamba_hybrid.py
index c6a0632c2d..cf6ab0ac06 100644
--- a/vllm/v1/worker/gpu/model_states/mamba_hybrid.py
+++ b/vllm/v1/worker/gpu/model_states/mamba_hybrid.py
@@ -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