[Multimodal] Type dummy options per modality (#57576)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-18 17:29:11 +01:00
committed by GitHub
co-authored by Claude Opus 5
parent 56cac80b14
commit 8a5cf54387
93 changed files with 461 additions and 720 deletions
@@ -25,6 +25,7 @@ import torch
from transformers import PretrainedConfig
from tests.models.registry import HF_EXAMPLE_MODELS
from vllm.config.multimodal import MultiModalDummyOptions
class MockAudioFlamingo3Config(PretrainedConfig):
@@ -127,7 +128,7 @@ def test_dummy_data_generation(mock_ctx):
builder = AudioFlamingo3DummyInputsBuilder(info)
mm_counts = {"audio": 2}
dummy_data = builder.get_dummy_mm_data(100, mm_counts, {})
dummy_data = builder.get_dummy_mm_data(100, mm_counts, MultiModalDummyOptions())
assert "audio" in dummy_data
assert len(dummy_data["audio"]) == 2
@@ -13,6 +13,7 @@ from vllm.config.multimodal import (
AudioDummyOptions,
BaseDummyOptions,
ImageDummyOptions,
MultiModalDummyOptions,
VideoDummyOptions,
)
from vllm.inputs import MultiModalDataDict, MultiModalInput
@@ -174,7 +175,7 @@ def get_token_prompt(
inputs = dummy_inputs.get_dummy_processor_inputs(
model_config.max_model_len,
mm_counts,
mm_options={},
mm_options=MultiModalDummyOptions(),
# Assume all Mistral models define this extra argument
mm_data=mm_data, # type: ignore[call-arg]
)
@@ -182,7 +183,7 @@ def get_token_prompt(
inputs = dummy_inputs.get_dummy_processor_inputs(
model_config.max_model_len,
mm_counts,
mm_options={},
mm_options=MultiModalDummyOptions(),
)
if not isinstance(inputs.prompt, list):
@@ -278,10 +279,12 @@ def _test_processing_correctness(
return BaseDummyOptions(count=count)
# Assign normalized DummyOptions to the model config
model_config.get_multimodal_config().limit_per_prompt = {
modality: _to_dummy_options(modality, count)
for modality, count in limit_mm_per_prompt_ints.items()
}
model_config.get_multimodal_config().limit_per_prompt = MultiModalDummyOptions(
{
modality: _to_dummy_options(modality, count)
for modality, count in limit_mm_per_prompt_ints.items()
}
)
baseline_processor = factories.build_processor(ctx, cache=None)
cached_processor = factories.build_processor(ctx, cache=cache)
@@ -6,7 +6,7 @@ from types import SimpleNamespace
import pytest
import torch
from vllm.config.multimodal import VideoDummyOptions
from vllm.config.multimodal import MultiModalDummyOptions, VideoDummyOptions
from vllm.model_executor.models.molmo2 import (
Molmo2DummyInputsBuilder,
build_flat_image_bool_length,
@@ -83,7 +83,9 @@ def test_dummy_video_num_frames_override_honors_min_of_two(
data = builder.get_dummy_mm_data(
seq_len=128,
mm_counts={"image": 0, "video": 1},
mm_options={"video": VideoDummyOptions(num_frames=num_frames_override)},
mm_options=MultiModalDummyOptions(
video=VideoDummyOptions(num_frames=num_frames_override)
),
)
video, _metadata = data["video"][0]
@@ -12,6 +12,7 @@ import numpy as np
import pytest
from vllm.config import ModelConfig
from vllm.config.multimodal import MultiModalDummyOptions
from vllm.multimodal import MULTIMODAL_REGISTRY
from ...registry import HF_EXAMPLE_MODELS
@@ -339,7 +340,9 @@ def test_dummy_video_spreads_budget_when_frame_cap_enabled(model_id: str) -> Non
limit_mm_per_prompt={"image": 0, "video": 1},
)
capped = MULTIMODAL_REGISTRY.create_processor(capped_ctx.model_config)
capped_dummy = capped.dummy_inputs.get_dummy_mm_data(1024, {"video": 1}, {})
capped_dummy = capped.dummy_inputs.get_dummy_mm_data(
1024, {"video": 1}, MultiModalDummyOptions()
)
capped_frames = capped_dummy["video"][0][0].shape[0]
assert capped_frames == 16, (
f"Expected the dummy to spread the budget over 16 frames, got {capped_frames}"
@@ -351,7 +354,9 @@ def test_dummy_video_spreads_budget_when_frame_cap_enabled(model_id: str) -> Non
limit_mm_per_prompt={"image": 0, "video": 1},
)
uncapped = MULTIMODAL_REGISTRY.create_processor(uncapped_ctx.model_config)
uncapped_dummy = uncapped.dummy_inputs.get_dummy_mm_data(1024, {"video": 1}, {})
uncapped_dummy = uncapped.dummy_inputs.get_dummy_mm_data(
1024, {"video": 1}, MultiModalDummyOptions()
)
uncapped_frames = uncapped_dummy["video"][0][0].shape[0]
assert uncapped_frames == 2, (
f"Expected the uncapped dummy to keep 2 frames, got {uncapped_frames}"
@@ -18,6 +18,7 @@ from vllm.config.multimodal import (
AudioDummyOptions,
BaseDummyOptions,
ImageDummyOptions,
MultiModalDummyOptions,
VideoDummyOptions,
)
from vllm.distributed import (
@@ -98,7 +99,7 @@ def create_batched_mm_kwargs(
processor_inputs = dummy_inputs.get_dummy_processor_inputs(
seq_len=model_config.max_model_len,
mm_counts=mm_counts,
mm_options={},
mm_options=MultiModalDummyOptions(),
)
mm_items = processor_inputs.mm_data_items
resized_mm_data = {
@@ -248,10 +249,12 @@ def test_model_tensor_schema(model_id: str):
return AudioDummyOptions(count=count)
return BaseDummyOptions(count=count)
model_config.get_multimodal_config().limit_per_prompt = {
modality: _to_dummy_options(modality, count)
for modality, count in limit_mm_per_prompt.items()
}
model_config.get_multimodal_config().limit_per_prompt = MultiModalDummyOptions(
{
modality: _to_dummy_options(modality, count)
for modality, count in limit_mm_per_prompt.items()
}
)
processor = factories.build_processor(ctx, cache=None)
with initialize_dummy_model(model_cls, model_config) as model: