[Core][Engine] only materialize tokens when thinking budget is in req (#47053)

Signed-off-by: walterbm <walter.beller.morales@gmail.com>

tests/v1/worker/test_gpu_model_runner.py

@@ -18,6 +18,7 @@ from vllm.config import (
     VllmConfig,
     set_current_vllm_config,
 )
+from vllm.config.reasoning import ReasoningConfig
 from vllm.distributed.parallel_state import (
     init_distributed_environment,
     initialize_model_parallel,
@@ -255,6 +256,31 @@ def test_select_common_block_size_uses_largest_shared_int():
     assert selected_size == 64
 
 
+def test_reasoning_config_without_custom_logitsprocs_does_not_need_output_token_ids(
+    dist_init,
+):
+    vllm_config = get_vllm_config()
+    assert vllm_config.model_config.logits_processors is None
+    reasoning_config = ReasoningConfig(
+        reasoning_start_str="<think>", reasoning_end_str="</think>"
+    )
+    reasoning_config._reasoning_start_token_ids = [1]
+    reasoning_config._reasoning_end_token_ids = [2]
+    vllm_config.reasoning_config = reasoning_config
+
+    with set_current_vllm_config(vllm_config):
+        model_config = vllm_config.model_config
+        num_heads = model_config.get_num_kv_heads(vllm_config.parallel_config)
+        head_size = model_config.get_head_size()
+        vllm_config.compilation_config.static_forward_context["layer.0"] = Attention(
+            num_heads, head_size, 0.1
+        )
+        runner = GPUModelRunner(vllm_config, torch.device("cpu"))
+
+    assert runner.input_batch.thinking_budget_state_holder is not None
+    assert runner.input_batch.logitsprocs_need_output_token_ids is False
+
+
 @pytest.mark.skip_global_cleanup
 @pytest.mark.parametrize(
     ("world_size", "is_last_rank", "expected_calls"),

vllm/v1/worker/gpu_model_runner.py

@@ -702,11 +702,9 @@ class GPUModelRunner(
                 custom_logitsprocs,
             ),
             # We currently don't know whether a particular custom logits processor
-            # uses output token ids so we set this conservatively.
-            # ThinkingTokenBudgetLogitsProcessor also needs output token ids to
-            # correctly track think start/end token sequences in async scheduling.
-            logitsprocs_need_output_token_ids=bool(custom_logitsprocs)
-            or self.vllm_config.reasoning_config is not None,
+            # uses output token ids so we set this conservatively. Thinking-budget
+            # tracking is requested dynamically when a budgeted request is in the batch.
+            logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
             is_pooling_model=self.is_pooling_model,
             cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
             reasoning_config=self.vllm_config.reasoning_config,