[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:
@@ -6,9 +6,11 @@ Qwen2.5-VL visual component loading behavior.
|
||||
"""
|
||||
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm.config import SchedulerConfig
|
||||
from vllm.multimodal import MULTIMODAL_REGISTRY
|
||||
|
||||
from ..models.utils import build_model_context
|
||||
@@ -16,6 +18,34 @@ from ..models.utils import build_model_context
|
||||
pytestmark = pytest.mark.cpu_test
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("chunked_prefill", "max_model_len", "expected_seq_len"),
|
||||
[(None, 491520, 491520), (True, 491520, 8192), (True, 128, 128), (False, 128, 128)],
|
||||
)
|
||||
def test_dummy_inputs_scheduler_budget(
|
||||
chunked_prefill, max_model_len, expected_seq_len
|
||||
):
|
||||
model_config = MagicMock()
|
||||
model_config.max_model_len = max_model_len
|
||||
processor = MagicMock()
|
||||
processor.apply.return_value = {"prompt_token_ids": [7]}
|
||||
kwargs = {}
|
||||
if chunked_prefill is not None:
|
||||
kwargs["scheduler_config"] = SchedulerConfig(
|
||||
max_model_len=max_model_len,
|
||||
is_encoder_decoder=False,
|
||||
max_num_batched_tokens=8192,
|
||||
max_num_seqs=1,
|
||||
enable_chunked_prefill=chunked_prefill,
|
||||
)
|
||||
result = MULTIMODAL_REGISTRY.get_dummy_mm_inputs(
|
||||
model_config, {"image": 1}, processor=processor, **kwargs
|
||||
)
|
||||
get_inputs = processor.dummy_inputs.get_dummy_processor_inputs
|
||||
assert get_inputs.call_args.kwargs["seq_len"] == expected_seq_len
|
||||
assert len(result["prompt_token_ids"]) == expected_seq_len
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"model_id,limit_mm_per_prompt,expected",
|
||||
[
|
||||
|
||||
@@ -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})
|
||||
|
||||
|
||||
Reference in New Issue
Block a user