A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 22:45 UTC
[BugFix] Fix async scheduling + reasoning with struct output (#31332)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Dec 25, 2025 at 15:01 UTC
81786c87748b0177111dfdc07af5351d8389baa1
3 files changed
+9
-3
tests/v1/entrypoints/llm/test_struct_output_generate.py
+6
-2
@@ -608,7 +608,7 @@ Make the response as short as possible.
608
609
610
@pytest.mark.parametrize(
611
- "model_name, backend, tokenizer_mode, reasoning_parser, speculative_config", # noqa: E501
611
+ "model_name, backend, tokenizer_mode, reasoning_parser, speculative_config, async_scheduling", # noqa: E501
612
[
613
(
614
"deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
@@ -616,8 +616,10 @@ Make the response as short as possible.
616
"auto",
617
"deepseek_r1",
618
NGRAM_SPEC_CONFIG,
619
+ False,
620
),
620
- ("Qwen/Qwen3-1.7B", "xgrammar", "auto", "deepseek_r1", None),
621
+ ("Qwen/Qwen3-1.7B", "xgrammar", "auto", "deepseek_r1", None, False),
622
+ ("Qwen/Qwen3-1.7B", "xgrammar", "auto", "deepseek_r1", None, True),
623
],
624
)
625
def test_structured_output_with_reasoning_matrices(
@@ -626,6 +628,7 @@ def test_structured_output_with_reasoning_matrices(
628
reasoning_parser: str,
629
model_name: str,
630
speculative_config: dict[str, Any] | None,
631
+ async_scheduling: bool,
632
):
633
if current_platform.is_tpu() and speculative_config:
634
pytest.skip("TPU does not support speculative decoding")
@@ -646,6 +649,7 @@ def test_structured_output_with_reasoning_matrices(
649
),
650
tokenizer_mode=tokenizer_mode,
651
speculative_config=speculative_config,
652
+ async_scheduling=async_scheduling,
653
)
654
tokenizer = llm.get_tokenizer()
655
reasoner = ReasoningParserManager.get_reasoning_parser(reasoning_parser)(
tests/v1/structured_output/test_reasoning_structured_output.py
+1
@@ -71,6 +71,7 @@ class TestReasoningStructuredOutput:
71
request.prompt_token_ids = [1, 2, 3, 4, 5]
72
request.all_token_ids = [1, 2, 3, 4, 5, 6, 7, 8]
73
request.num_computed_tokens = 5
74
+ request.num_output_placeholders = 0
75
return request
76
77
def test_should_fill_bitmask_with_enable_in_reasoning(
vllm/v1/structured_output/__init__.py
+2
-1
@@ -339,8 +339,9 @@ class StructuredOutputManager:
339
return True
340
341
# Check if reasoning ends in *this* step
342
+ delta_from = request.num_computed_tokens - request.num_output_placeholders
343
if self.reasoner.is_reasoning_end_streaming(
343
- request.all_token_ids, request.all_token_ids[request.num_computed_tokens :]
344
+ request.all_token_ids, request.all_token_ids[delta_from:]
345
):
346
# Reasoning just ended, so we shouldn't advance til
347
# next pass