[Bugfix][Multimodal] Bound renderer warmup to the prefill token budget (#55448)

Signed-off-by: Luca Motz <luca.motz@icloud.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
lucamotz
2026-09-05 16:43:28 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 7985444339
commit f4eccdadef
4 changed files with 59 additions and 8 deletions
+16
View File
@@ -45,6 +45,7 @@ def _make_renderer_mock(mm_limits: dict[str, int]) -> MagicMock:
# MM processor with configurable limits
mm_processor = MagicMock()
mm_processor.info.allowed_mm_limits = mm_limits
mm_processor.apply.return_value = {"prompt_token_ids": [1]}
renderer.mm_processor = mm_processor
renderer._readonly_mm_processor = None
renderer._warmup_mm_processor = BaseRenderer._warmup_mm_processor.__get__(
@@ -65,6 +66,8 @@ def _make_renderer_mock(mm_limits: dict[str, int]) -> MagicMock:
renderer._mm_warmup_done = False
renderer.clear_mm_cache = MagicMock()
renderer.model_config.max_model_len = 128
renderer.config.scheduler_config.max_num_batched_tokens = 8192
renderer.config.scheduler_config.enable_chunked_prefill = True
renderer.model_config.get_multimodal_config.return_value.limit_per_prompt = {}
return renderer
@@ -115,6 +118,19 @@ class TestMmWarmupRunsNormally:
# MM warmup must run when mm_processor is set and limits > 0; the chat
# template warmup must run alongside it.
@pytest.mark.parametrize(
("max_model_len", "expected_seq_len"), [(491520, 8192), (128, 128)]
)
def test_warmup_sequence_length_is_bounded(self, max_model_len, expected_seq_len):
renderer = _make_renderer_mock({"image": 1, "video": 1})
renderer.model_config.max_model_len = max_model_len
with patch("vllm.multimodal.processing.TimingContext", autospec=True):
BaseRenderer.warmup(renderer, ChatParams())
get_inputs = renderer.mm_processor.dummy_inputs.get_dummy_processor_inputs
assert get_inputs.call_args.kwargs["seq_len"] == expected_seq_len
def test_processor_apply_called(self):
renderer = _make_renderer_mock({"image": 1})