[Structured Output] Let terminal grammars stop under min_tokens (II) (#54218)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
This commit is contained in:
Artem Perevedentsev
2026-08-31 20:39:27 +04:00
committed by GitHub
parent e9dd6d4834
commit 39e276eaeb
7 changed files with 501 additions and 60 deletions
+24
View File
@@ -13,11 +13,14 @@ Covers:
5) Multiple stop conditions
"""
import json
import pytest
from vllm import LLM, SamplingParams
from vllm.exceptions import VLLMValidationError
from vllm.outputs import RequestOutput
from vllm.sampling_params import StructuredOutputsParams
# Test configuration
TEST_MODEL = "facebook/opt-125m" # Small model for fast CI execution
@@ -168,6 +171,27 @@ def llm_v1():
return llm
def test_min_tokens_with_terminal_structured_output(llm_v1: LLM):
"""A grammar whose only legal continuation is EOS must stop under
min_tokens instead of sampling outside the grammar."""
params = SamplingParams(
temperature=GREEDY,
min_tokens=20,
max_tokens=40,
structured_outputs=StructuredOutputsParams(
json=json.dumps({"type": "boolean"})
),
)
output = llm_v1.generate("Answer:", params)[0].outputs[0]
eos_token_id = llm_v1.get_tokenizer().eos_token_id
assert output.finish_reason == "stop"
assert isinstance(json.loads(output.text), bool)
assert output.token_ids[-1] == eos_token_id
assert len(output.token_ids) < params.min_tokens
def get_token_count(output: RequestOutput) -> int:
"""Extract token count from LLM output"""
if not output.outputs: