[Chore] Deprecate merge_by_field_config arg (#30035)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-12-04 17:21:24 +00:00
committed by GitHub
parent 990f806473
commit b286a311c2
19 changed files with 90 additions and 302 deletions
@@ -20,7 +20,7 @@ from vllm.config.multimodal import (
)
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalDataDict
from vllm.multimodal.cache import MultiModalProcessorOnlyCache
from vllm.multimodal.inputs import MultiModalInputs
from vllm.multimodal.inputs import MultiModalInputs, batched_tensors_equal
from vllm.multimodal.processing import BaseMultiModalProcessor, InputProcessingContext
from vllm.tokenizers import (
MistralTokenizer,
@@ -418,4 +418,4 @@ def _assert_inputs_equal(
a_data.pop(key, None)
b_data.pop(key, None)
assert a_data == b_data, msg
assert batched_tensors_equal(a_data, b_data), msg
@@ -5,6 +5,7 @@ import pytest
from vllm.assets.video import VideoAsset
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import batched_tensors_equal
from vllm.multimodal.video import OpenCVDynamicVideoBackend, OpenCVVideoBackend
from ...utils import build_model_context
@@ -103,7 +104,7 @@ def test_video_loader_consistency(
dynamic_outputs = processor.apply(prompt, dynamic_mm_data, hf_processor_mm_kwargs)
assert static_outputs["prompt_token_ids"] == dynamic_outputs["prompt_token_ids"]
assert (
static_outputs["mm_kwargs"].get_data()
== dynamic_outputs["mm_kwargs"].get_data()
assert batched_tensors_equal(
static_outputs["mm_kwargs"].get_data(),
dynamic_outputs["mm_kwargs"].get_data(),
)
@@ -130,10 +130,9 @@ def create_batched_mm_kwargs(
hf_processor_mm_kwargs=processor_inputs.hf_processor_mm_kwargs,
tokenization_kwargs=processor_inputs.tokenization_kwargs,
)["mm_kwargs"].require_data()
items = [item for modality in supported_mm_limits for item in mm_kwargs[modality]]
return group_mm_kwargs_by_modality(
items,
merge_by_field_config=model_cls.merge_by_field_config,
[item for modality in supported_mm_limits for item in mm_kwargs[modality]]
)
+3 -6
View File
@@ -85,12 +85,6 @@ def _dummy_items(
(_dummy_item("a", {"a1": 100}), 100),
(_dummy_item("a", {"a1": 100, "a2": 110}), 210),
(_dummy_items({"a": {"a1": 100, "a2": 110}, "b": {"b1": 120, "b2": 130}}), 460), # noqa: E501
(
_dummy_items(
{"a": {"a1": 100, "a2": 110}, "b": {"b1": 120, "b2": 130}}
).get_data(),
460,
), # noqa: E501
],
)
def test_cache_item_size(item, expected_size):
@@ -107,6 +101,9 @@ def test_cache_item_size(item, expected_size):
cache[""] = MultiModalProcessorCacheItemMetadata(item, [prompt_update])
assert cache.currsize == expected_size
cache[""] = item.get_data()
assert cache.currsize == expected_size
def _create_vllm_config(
*,
-91
View File
@@ -1,91 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from vllm.multimodal.inputs import MultiModalKwargs, NestedTensors
pytestmark = pytest.mark.cpu_test
def assert_nested_tensors_equal(expected: NestedTensors, actual: NestedTensors):
assert type(expected) == type(actual) # noqa: E721
if isinstance(expected, torch.Tensor):
assert torch.equal(expected, actual)
else:
for expected_item, actual_item in zip(expected, actual):
assert_nested_tensors_equal(expected_item, actual_item)
def assert_multimodal_inputs_equal(
expected: MultiModalKwargs, actual: MultiModalKwargs
):
assert set(expected.keys()) == set(actual.keys())
for key in expected:
assert_nested_tensors_equal(expected[key], actual[key])
def test_multimodal_input_batch_single_tensor():
t = torch.rand([1, 2])
result = MultiModalKwargs.batch([{"image": t}])
assert_multimodal_inputs_equal(result, {"image": t.unsqueeze(0)})
def test_multimodal_input_batch_multiple_tensors():
a = torch.rand([1, 1, 2])
b = torch.rand([1, 1, 2])
c = torch.rand([1, 1, 2])
result = MultiModalKwargs.batch([{"image": a}, {"image": b}, {"image": c}])
assert_multimodal_inputs_equal(result, {"image": torch.stack([a, b, c])})
def test_multimodal_input_batch_multiple_heterogeneous_tensors():
a = torch.rand([1, 2, 2])
b = torch.rand([1, 3, 2])
c = torch.rand([1, 4, 2])
result = MultiModalKwargs.batch([{"image": a}, {"image": b}, {"image": c}])
assert_multimodal_inputs_equal(result, {"image": [a, b, c]})
def test_multimodal_input_batch_nested_tensors():
a = torch.rand([2, 3])
b = torch.rand([2, 3])
c = torch.rand([2, 3])
result = MultiModalKwargs.batch([{"image": [a]}, {"image": [b]}, {"image": [c]}])
assert_multimodal_inputs_equal(
result, {"image": torch.stack([a.unsqueeze(0), b.unsqueeze(0), c.unsqueeze(0)])}
)
def test_multimodal_input_batch_heterogeneous_lists():
a = torch.rand([1, 2, 3])
b = torch.rand([1, 2, 3])
c = torch.rand([1, 2, 3])
result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c]}])
assert_multimodal_inputs_equal(
result, {"image": [torch.stack([a, b]), c.unsqueeze(0)]}
)
def test_multimodal_input_batch_multiple_batchable_lists():
a = torch.rand([1, 2, 3])
b = torch.rand([1, 2, 3])
c = torch.rand([1, 2, 3])
d = torch.rand([1, 2, 3])
result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c, d]}])
assert_multimodal_inputs_equal(
result, {"image": torch.stack([torch.stack([a, b]), torch.stack([c, d])])}
)
def test_multimodal_input_batch_mixed_stacking_depths():
a = torch.rand([1, 2, 3])
b = torch.rand([1, 3, 3])
c = torch.rand([1, 4, 3])
result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c]}])
assert_multimodal_inputs_equal(result, {"image": [[a, b], c.unsqueeze(0)]})
result = MultiModalKwargs.batch([{"image": [a]}, {"image": [b, c]}])
assert_multimodal_inputs_equal(result, {"image": [a.unsqueeze(0), [b, c]]})