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

Signed-off-by: walterbm <walter.beller.morales@gmail.com>
This commit is contained in:
Walter Beller-Morales
2026-07-07 21:02:38 -06:00
committed by GitHub
parent 0ca6eee743
commit dd127d82ed
2 changed files with 29 additions and 5 deletions
+26
View File
@@ -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"),