A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 08:15 UTC
[0/N] Rename `MultiModalInputs` to `MultiModalKwargs` (#10040)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Nov 9, 2024 at 11:31 UTC
e0191a95d88c454dbb989b7457a41c93cb7f7051
32 files changed
+151
-121
docs/source/design/multimodal/multimodal_index.rst
+1
-1
@@ -53,7 +53,7 @@ Base Classes
53
54
.. autodata:: vllm.multimodal.MultiModalDataDict
55
56
-.. autoclass:: vllm.multimodal.MultiModalInputs
56
+.. autoclass:: vllm.multimodal.MultiModalKwargs
57
:members:
58
:show-inheritance:
59
tests/models/decoder_only/vision_language/mm_processor_kwargs/test_qwen.py
+2
-2
@@ -6,7 +6,7 @@ import torch
6
from PIL.Image import Image
7
8
from vllm.inputs import InputContext, token_inputs
9
-from vllm.multimodal.base import MultiModalInputs
9
+from vllm.multimodal.base import MultiModalKwargs
10
from vllm.multimodal.utils import cached_get_tokenizer
11
12
from .....conftest import IMAGE_ASSETS
@@ -96,7 +96,7 @@ def test_input_mapper_valid_mm_data(input_mapper_for_qwen,
96
mapped_img_data = input_mapper_for_qwen(qwen_vl_context, img_data)
97
# Ensure that we get the appropriately shaped pixel_values
98
# for images and image embeddings, respectively.
99
- assert isinstance(mapped_img_data, MultiModalInputs)
99
+ assert isinstance(mapped_img_data, MultiModalKwargs)
100
assert "pixel_values" in mapped_img_data
101
assert mapped_img_data["pixel_values"].shape == expected_shape
102
tests/multimodal/test_base.py
+11
-11
@@ -1,6 +1,6 @@
1
import torch
2
3
-from vllm.multimodal.base import MultiModalInputs, NestedTensors
3
+from vllm.multimodal.base import MultiModalKwargs, NestedTensors
4
5
6
def assert_nested_tensors_equal(expected: NestedTensors,
@@ -13,8 +13,8 @@ def assert_nested_tensors_equal(expected: NestedTensors,
13
assert_nested_tensors_equal(expected_item, actual_item)
14
15
16
-def assert_multimodal_inputs_equal(expected: MultiModalInputs,
17
- actual: MultiModalInputs):
16
+def assert_multimodal_inputs_equal(expected: MultiModalKwargs,
17
+ actual: MultiModalKwargs):
18
assert set(expected.keys()) == set(actual.keys())
19
for key in expected:
20
assert_nested_tensors_equal(expected[key], actual[key])
@@ -22,7 +22,7 @@ def assert_multimodal_inputs_equal(expected: MultiModalInputs,
22
23
def test_multimodal_input_batch_single_tensor():
24
t = torch.rand([1, 2])
25
- result = MultiModalInputs.batch([{"image": t}])
25
+ result = MultiModalKwargs.batch([{"image": t}])
26
assert_multimodal_inputs_equal(result, {"image": t.unsqueeze(0)})
27
28
@@ -30,7 +30,7 @@ def test_multimodal_input_batch_multiple_tensors():
30
a = torch.rand([1, 1, 2])
31
b = torch.rand([1, 1, 2])
32
c = torch.rand([1, 1, 2])
33
- result = MultiModalInputs.batch([{"image": a}, {"image": b}, {"image": c}])
33
+ result = MultiModalKwargs.batch([{"image": a}, {"image": b}, {"image": c}])
34
assert_multimodal_inputs_equal(result, {"image": torch.stack([a, b, c])})
35
36
@@ -38,7 +38,7 @@ def test_multimodal_input_batch_multiple_heterogeneous_tensors():
38
a = torch.rand([1, 2, 2])
39
b = torch.rand([1, 3, 2])
40
c = torch.rand([1, 4, 2])
41
- result = MultiModalInputs.batch([{"image": a}, {"image": b}, {"image": c}])
41
+ result = MultiModalKwargs.batch([{"image": a}, {"image": b}, {"image": c}])
42
assert_multimodal_inputs_equal(result, {"image": [a, b, c]})
43
44
@@ -46,7 +46,7 @@ def test_multimodal_input_batch_nested_tensors():
46
a = torch.rand([2, 3])
47
b = torch.rand([2, 3])
48
c = torch.rand([2, 3])
49
- result = MultiModalInputs.batch([{
49
+ result = MultiModalKwargs.batch([{
50
"image": [a]
51
}, {
52
"image": [b]
@@ -65,7 +65,7 @@ def test_multimodal_input_batch_heterogeneous_lists():
65
a = torch.rand([1, 2, 3])
66
b = torch.rand([1, 2, 3])
67
c = torch.rand([1, 2, 3])
68
- result = MultiModalInputs.batch([{"image": [a, b]}, {"image": [c]}])
68
+ result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c]}])
69
assert_multimodal_inputs_equal(
70
result,
71
{"image": [torch.stack([a, b]), c.unsqueeze(0)]})
@@ -76,7 +76,7 @@ def test_multimodal_input_batch_multiple_batchable_lists():
76
b = torch.rand([1, 2, 3])
77
c = torch.rand([1, 2, 3])
78
d = torch.rand([1, 2, 3])
79
- result = MultiModalInputs.batch([{"image": [a, b]}, {"image": [c, d]}])
79
+ result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c, d]}])
80
assert_multimodal_inputs_equal(
81
result,
82
{"image": torch.stack([torch.stack([a, b]),
@@ -88,8 +88,8 @@ def test_multimodal_input_batch_mixed_stacking_depths():
88
b = torch.rand([1, 3, 3])
89
c = torch.rand([1, 4, 3])
90
91
- result = MultiModalInputs.batch([{"image": [a, b]}, {"image": [c]}])
91
+ result = MultiModalKwargs.batch([{"image": [a, b]}, {"image": [c]}])
92
assert_multimodal_inputs_equal(result, {"image": [[a, b], c.unsqueeze(0)]})
93
94
- result = MultiModalInputs.batch([{"image": [a]}, {"image": [b, c]}])
94
+ result = MultiModalKwargs.batch([{"image": [a]}, {"image": [b, c]}])
95
assert_multimodal_inputs_equal(result, {"image": [a.unsqueeze(0), [b, c]]})
vllm/model_executor/models/chatglm.py
+2
-2
@@ -30,7 +30,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
30
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
31
from vllm.model_executor.models.glm4_vision_encoder import EVA2CLIPModel
32
from vllm.model_executor.sampling_metadata import SamplingMetadata
33
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalInputs
33
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
34
from vllm.multimodal.base import MultiModalData
35
from vllm.multimodal.utils import cached_get_tokenizer
36
from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE, IntermediateTensors,
@@ -74,7 +74,7 @@ def mm_input_mapper_for_glmv(
74
raise
75
pixel_values = raw_batch_data['images']
76
77
- return MultiModalInputs({'pixel_values': pixel_values})
77
+ return MultiModalKwargs({'pixel_values': pixel_values})
78
79
80
def merge_glm_vision_embeddings(
vllm/model_executor/models/fuyu.py
+2
-2
@@ -34,7 +34,7 @@ from vllm.model_executor.layers.sampler import SamplerOutput
34
from vllm.model_executor.models.persimmon import PersimmonForCausalLM
35
from vllm.model_executor.sampling_metadata import SamplingMetadata
36
from vllm.multimodal import MULTIMODAL_REGISTRY
37
-from vllm.multimodal.base import MultiModalInputs
37
+from vllm.multimodal.base import MultiModalKwargs
38
from vllm.multimodal.image import cached_get_image_processor
39
from vllm.multimodal.utils import (cached_get_tokenizer,
40
consecutive_placeholder_ranges)
@@ -218,7 +218,7 @@ def input_mapper_for_fuyu(ctx: InputContext, data: object):
218
])
219
220
# image has been processed with prompt in input processor
221
- return MultiModalInputs({"pixel_values": data})
221
+ return MultiModalKwargs({"pixel_values": data})
222
223
224
@MULTIMODAL_REGISTRY.register_image_input_mapper(input_mapper_for_fuyu)
vllm/model_executor/models/h2ovl.py
+5
-5
@@ -16,7 +16,7 @@ from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, InputContext,
16
token_inputs)
17
from vllm.model_executor.layers.quantization import QuantizationConfig
18
from vllm.multimodal import MULTIMODAL_REGISTRY
19
-from vllm.multimodal.base import MultiModalInputs
19
+from vllm.multimodal.base import MultiModalKwargs
20
from vllm.multimodal.utils import cached_get_tokenizer
21
from vllm.utils import is_list_of
22
@@ -324,12 +324,12 @@ class H2OVLInputPipeline(InternVLInputPipeline):
324
data: object,
325
*,
326
max_dynamic_patch: Optional[int] = None,
327
- ) -> MultiModalInputs:
327
+ ) -> MultiModalKwargs:
328
329
# NOTE: Preprocessing for the image data is done in the
330
# 'input_processor' function during actual inference.
331
if isinstance(data, dict):
332
- return MultiModalInputs(data)
332
+ return MultiModalKwargs(data)
333
334
# The section below is only used with dummy data during
335
# memory profiling.
@@ -347,7 +347,7 @@ class H2OVLInputPipeline(InternVLInputPipeline):
347
pixel_values = [image_pixel_values_mapper(img) for img in data]
348
349
else:
350
- return MultiModalInputs({"image_embeds": data})
350
+ return MultiModalKwargs({"image_embeds": data})
351
model_config = ctx.model_config
352
tokenizer = cached_get_tokenizer(
353
model_config.tokenizer,
@@ -359,7 +359,7 @@ class H2OVLInputPipeline(InternVLInputPipeline):
359
return_tensors="pt",
360
)[0]
361
362
- return MultiModalInputs({
362
+ return MultiModalKwargs({
363
"pixel_values": pixel_values,
364
"image_token_id": image_token_id
365
})
vllm/model_executor/models/idefics3.py
+2
-2
@@ -36,7 +36,7 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
36
from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
37
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
38
from vllm.model_executor.sampling_metadata import SamplingMetadata
39
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalInputs
39
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
40
from vllm.multimodal.image import cached_get_image_processor
41
from vllm.sequence import IntermediateTensors, SequenceData
42
from vllm.transformers_utils.processor import cached_get_processor
@@ -127,7 +127,7 @@ def input_mapper_for_idefics3(
127
logger.error("Failed to process image (%s)", data)
128
raise
129
130
- return MultiModalInputs(batch_data)
130
+ return MultiModalKwargs(batch_data)
131
132
133
def _resize_output_size(height: int,
vllm/model_executor/models/internvl.py
+3
-3
@@ -26,7 +26,7 @@ from vllm.model_executor.models.intern_vit import (InternVisionModel,
26
InternVisionPatchModel)
27
from vllm.model_executor.sampling_metadata import SamplingMetadata
28
from vllm.multimodal import MULTIMODAL_REGISTRY
29
-from vllm.multimodal.base import MultiModalInputs
29
+from vllm.multimodal.base import MultiModalKwargs
30
from vllm.multimodal.utils import cached_get_tokenizer
31
from vllm.sequence import IntermediateTensors
32
from vllm.utils import is_list_of
@@ -346,7 +346,7 @@ class InternVLInputPipeline:
346
# we can't stack here because images may have different num_patches
347
data = [image_pixel_values_mapper(img) for img in data]
348
else:
349
- return MultiModalInputs({"image_embeds": data})
349
+ return MultiModalKwargs({"image_embeds": data})
350
model_config = ctx.model_config
351
tokenizer = cached_get_tokenizer(
352
model_config.tokenizer,
@@ -355,7 +355,7 @@ class InternVLInputPipeline:
355
add_special_tokens=False,
356
return_tensors="pt")[0]
357
358
- return MultiModalInputs({
358
+ return MultiModalKwargs({
359
"pixel_values": data,
360
"image_token_id": image_token_id
361
})
vllm/model_executor/models/minicpmv.py
+2
-2
@@ -52,7 +52,7 @@ from vllm.model_executor.models.qwen2 import Qwen2Model
52
from vllm.model_executor.models.utils import LLMWrapper
53
from vllm.model_executor.sampling_metadata import SamplingMetadata
54
from vllm.multimodal import MULTIMODAL_REGISTRY
55
-from vllm.multimodal.base import MultiModalInputs
55
+from vllm.multimodal.base import MultiModalKwargs
56
from vllm.multimodal.image import cached_get_image_processor
57
from vllm.multimodal.utils import cached_get_tokenizer
58
from vllm.sequence import IntermediateTensors, SequenceData
@@ -374,7 +374,7 @@ def input_mapper_for_minicpmv(ctx: InputContext, data: object):
374
batch_data["slice_start_id"] = data[0]["slice_start_id"]
375
batch_data["slice_end_id"] = data[0]["slice_end_id"]
376
377
- return MultiModalInputs(batch_data)
377
+ return MultiModalKwargs(batch_data)
378
379
380
class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
vllm/model_executor/models/mllama.py
+1
-1
@@ -1162,7 +1162,7 @@ class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
1162
1163
def _parse_and_validate_image_input(self, **kwargs: object):
1164
# tensor with the same shape will be batched together by
1165
- # MultiModalInputs.batch, so pixel_values here can be:
1165
+ # MultiModalKwargs.batch, so pixel_values here can be:
1166
# - List[List[torch.Tensor]]:
1167
# with shape (num_tiles, 3, image_res, image_res)
1168
# - List[torch.Tensor]:
vllm/model_executor/models/molmo.py
+2
-2
@@ -37,7 +37,7 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
37
from vllm.model_executor.layers.vocab_parallel_embedding import (
38
ParallelLMHead, VocabParallelEmbedding)
39
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
40
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalInputs
40
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
41
from vllm.multimodal.utils import cached_get_tokenizer
42
from vllm.sequence import (VLLM_TOKEN_ID_ARRAY_TYPE, IntermediateTensors,
43
SequenceData)
@@ -866,7 +866,7 @@ def image_input_mapper_for_molmo(
866
ctx: InputContext,
867
data: object,
868
):
869
- return MultiModalInputs(data)
869
+ return MultiModalKwargs(data)
870
871
872
def dummy_data_for_molmo(ctx: InputContext, seq_len: int,
vllm/model_executor/models/pixtral.py
+5
-5
@@ -30,7 +30,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
30
from vllm.model_executor.models.utils import merge_multimodal_embeddings
31
from vllm.model_executor.sampling_metadata import SamplingMetadata
32
from vllm.multimodal import MULTIMODAL_REGISTRY
33
-from vllm.multimodal.base import MultiModalInputs
33
+from vllm.multimodal.base import MultiModalKwargs
34
from vllm.multimodal.utils import (cached_get_tokenizer,
35
consecutive_placeholder_ranges)
36
from vllm.sequence import IntermediateTensors, SequenceData
@@ -94,8 +94,8 @@ def dummy_data_for_pixtral(ctx: InputContext, seq_len: int,
94
95
96
def input_mapper_for_pixtral(ctx: InputContext,
97
- data: object) -> MultiModalInputs:
98
- """Maps the input data to its MultiModalInputs (if any).
97
+ data: object) -> MultiModalKwargs:
98
+ """Maps the input data to its MultiModalKwargs (if any).
99
100
Args:
101
ctx: Context of the loaded model.
@@ -103,7 +103,7 @@ def input_mapper_for_pixtral(ctx: InputContext,
103
to pixel_values in .forward() for a visual QWenLMHeadModel model.
104
105
Returns:
106
- MultiModalInputs containing the stacked normalized images tensor or
106
+ MultiModalKwargs containing the stacked normalized images tensor or
107
image embeddings.
108
"""
109
# Early exit if we have provided an image to a language only Qwen model
@@ -121,7 +121,7 @@ def input_mapper_for_pixtral(ctx: InputContext,
121
dtype=torch.float16)
122
images.append(image)
123
124
- return MultiModalInputs({"images": images})
124
+ return MultiModalKwargs({"images": images})
125
126
127
def input_processor_for_pixtral(ctx: InputContext, inputs: DecoderOnlyInputs):
vllm/model_executor/models/qwen.py
+6
-6
@@ -43,7 +43,7 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader
43
from vllm.model_executor.models.module_mapping import MultiModelKeys
44
from vllm.model_executor.sampling_metadata import SamplingMetadata
45
from vllm.multimodal import MULTIMODAL_REGISTRY
46
-from vllm.multimodal.base import MultiModalInputs
46
+from vllm.multimodal.base import MultiModalKwargs
47
from vllm.multimodal.utils import cached_get_tokenizer
48
from vllm.sequence import IntermediateTensors, SequenceData
49
from vllm.utils import is_list_of
@@ -722,8 +722,8 @@ def input_processor_for_qwen(ctx: InputContext,
722
multi_modal_data=multi_modal_data)
723
724
725
-def input_mapper_for_qwen(ctx: InputContext, data: object) -> MultiModalInputs:
726
- """Maps the input data to its MultiModalInputs (if any).
725
+def input_mapper_for_qwen(ctx: InputContext, data: object) -> MultiModalKwargs:
726
+ """Maps the input data to its MultiModalKwargs (if any).
727
728
Args:
729
ctx: Context of the loaded model.
@@ -731,7 +731,7 @@ def input_mapper_for_qwen(ctx: InputContext, data: object) -> MultiModalInputs:
731
to pixel_values in .forward() for a visual QWenLMHeadModel model.
732
733
Returns:
734
- MultiModalInputs containing the stacked normalized images tensor or
734
+ MultiModalKwargs containing the stacked normalized images tensor or
735
image embeddings.
736
"""
737
# Early exit if we have provided an image to a language only Qwen model
@@ -740,7 +740,7 @@ def input_mapper_for_qwen(ctx: InputContext, data: object) -> MultiModalInputs:
740
logger.warning(
741
"Images were provided but this model has no visual config; "
742
"multimodal inputs will not be forwarded to the model.")
743
- return MultiModalInputs()
743
+ return MultiModalKwargs()
744
745
model_config = ctx.model_config
746
tokenizer = cached_get_tokenizer(
@@ -784,7 +784,7 @@ def input_mapper_for_qwen(ctx: InputContext, data: object) -> MultiModalInputs:
784
data = [data]
785
transformed_images = [transform(datum) for datum in data]
786
pixel_values = torch.stack(transformed_images, dim=0)
787
- return MultiModalInputs({"pixel_values": pixel_values})
787
+ return MultiModalKwargs({"pixel_values": pixel_values})
788
789
790
def build_normalization_transform(image_size: int) -> transforms.Compose:
vllm/model_executor/models/qwen2_audio.py
+4
-4
@@ -42,7 +42,7 @@ from vllm.model_executor.model_loader.weight_utils import (
42
default_weight_loader, maybe_remap_kv_scale_name)
43
from vllm.model_executor.models.qwen2 import Qwen2Model
44
from vllm.model_executor.sampling_metadata import SamplingMetadata
45
-from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalInputs
45
+from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
46
from vllm.multimodal.utils import consecutive_placeholder_ranges
47
from vllm.sequence import IntermediateTensors, SequenceData
48
@@ -221,13 +221,13 @@ def input_processor_for_qwen2_audio(
221
def input_mapper_for_qwen2_audio(
222
ctx: InputContext,
223
multi_modal_data: Union[np.ndarray, List[np.ndarray]],
224
-) -> MultiModalInputs:
224
+) -> MultiModalKwargs:
225
"""Input mapper for Qwen2-Audio."""
226
if not isinstance(multi_modal_data, list):
227
multi_modal_data = [multi_modal_data]
228
229
if len(multi_modal_data) == 0:
230
- return MultiModalInputs()
230
+ return MultiModalKwargs()
231
232
processor = cached_get_processor(ctx.model_config.model)
233
audio_feature_extractor = processor.feature_extractor
@@ -254,7 +254,7 @@ def input_mapper_for_qwen2_audio(
254
logger.error("Failed to process audio (%s)", multi_modal_data)
255
raise
256
257
- return MultiModalInputs(batch_data)
257
+ return MultiModalKwargs(batch_data)
258
259
260
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_qwen2_audio)
vllm/model_executor/models/qwen2_vl.py
+4
-4
@@ -57,7 +57,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
57
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
58
from vllm.model_executor.models.qwen2 import Qwen2Model
59
from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalDataDict,
60
- MultiModalInputs)
60
+ MultiModalKwargs)
61
from vllm.multimodal.base import MultiModalData
62
from vllm.multimodal.image import cached_get_image_processor
63
from vllm.multimodal.utils import cached_get_tokenizer
@@ -576,10 +576,10 @@ def mm_input_mapper_for_qwen2_vl(
576
*,
577
min_pixels: Optional[int] = None,
578
max_pixels: Optional[int] = None,
579
-) -> MultiModalInputs:
579
+) -> MultiModalKwargs:
580
"""Input mapper for Qwen2-VL."""
581
if data_type_key == "image" and isinstance(data, dict):
582
- return MultiModalInputs({
582
+ return MultiModalKwargs({
583
"image_embeds": data.get("image_embeds"),
584
"image_grid_thw": data.get("image_grid_thw"),
585
})
@@ -613,7 +613,7 @@ def mm_input_mapper_for_qwen2_vl(
613
logger.error("Failed to process image (%s)", data)
614
raise
615
616
- return MultiModalInputs(batch_data)
616
+ return MultiModalKwargs(batch_data)
617
618
619
image_input_mapper_for_qwen2_vl = partial(mm_input_mapper_for_qwen2_vl,
vllm/model_executor/models/ultravox.py
+4
-4
@@ -24,7 +24,7 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
24
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
25
from vllm.model_executor.model_loader.loader import DefaultModelLoader
26
from vllm.model_executor.sampling_metadata import SamplingMetadata
27
-from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalInputs,
27
+from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalKwargs,
28
NestedTensors)
29
from vllm.multimodal.utils import (cached_get_tokenizer,
30
consecutive_placeholder_ranges,
@@ -116,11 +116,11 @@ def input_mapper_for_ultravox(ctx: InputContext, data: object):
116
data = [data]
117
118
if len(data) == 0:
119
- return MultiModalInputs()
119
+ return MultiModalKwargs()
120
121
# If the audio inputs are embeddings, no need for preprocessing
122
if is_list_of(data, torch.Tensor, check="all"):
123
- return MultiModalInputs({"audio_embeds": data})
123
+ return MultiModalKwargs({"audio_embeds": data})
124
125
audio_features = []
126
for audio_input in data:
@@ -154,7 +154,7 @@ def input_mapper_for_ultravox(ctx: InputContext, data: object):
154
# Remove the batch dimension because we're wrapping it in a list.
155
audio_features.append(single_audio_features.squeeze(0))
156
157
- return MultiModalInputs({"audio_features": audio_features})
157
+ return MultiModalKwargs({"audio_features": audio_features})
158
159
160
def input_processor_for_ultravox(ctx: InputContext, inputs: DecoderOnlyInputs):
vllm/multimodal/__init__.py
+17
-2
@@ -1,5 +1,5 @@
1
from .base import (BatchedTensorInputs, MultiModalDataBuiltins,
2
- MultiModalDataDict, MultiModalInputs,
2
+ MultiModalDataDict, MultiModalKwargs,
3
MultiModalPlaceholderDict, MultiModalPlaceholderMap,
4
MultiModalPlugin, NestedTensors)
5
from .registry import MultiModalRegistry
@@ -17,7 +17,7 @@ __all__ = [
17
"BatchedTensorInputs",
18
"MultiModalDataBuiltins",
19
"MultiModalDataDict",
20
- "MultiModalInputs",
20
+ "MultiModalKwargs",
21
"MultiModalPlaceholderDict",
22
"MultiModalPlaceholderMap",
23
"MultiModalPlugin",
@@ -25,3 +25,18 @@ __all__ = [
25
"MULTIMODAL_REGISTRY",
26
"MultiModalRegistry",
27
]
28
+
29
+
30
+def __getattr__(name: str):
31
+ import warnings
32
+
33
+ if name == "MultiModalInputs":
34
+ msg = ("MultiModalInputs has been renamed to MultiModalKwargs. "
35
+ "The original name will take another meaning in an upcoming "
36
+ "version.")
37
+
38
+ warnings.warn(DeprecationWarning(msg), stacklevel=2)
39
+
40
+ return MultiModalKwargs
41
+
42
+ raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
vllm/multimodal/audio.py
+2
-2
@@ -1,5 +1,5 @@
1
from vllm.inputs.registry import InputContext
2
-from vllm.multimodal.base import MultiModalInputs, MultiModalPlugin
2
+from vllm.multimodal.base import MultiModalKwargs, MultiModalPlugin
3
4
5
class AudioPlugin(MultiModalPlugin):
@@ -9,7 +9,7 @@ class AudioPlugin(MultiModalPlugin):
9
return "audio"
10
11
def _default_input_mapper(self, ctx: InputContext, data: object,
12
- **mm_processor_kwargs) -> MultiModalInputs:
12
+ **mm_processor_kwargs) -> MultiModalKwargs:
13
raise NotImplementedError("There is no default audio input mapper")
14
15
def _default_max_multimodal_tokens(self, ctx: InputContext) -> int:
vllm/multimodal/base.py
+24
-9
@@ -30,15 +30,15 @@ Uses a list instead of a tensor if the dimensions of each element do not match.
30
BatchedTensorInputs: TypeAlias = Dict[str, NestedTensors]
31
"""
32
A dictionary containing nested tensors which have been batched via
33
-:meth:`MultiModalInputs.batch`.
33
+:meth:`MultiModalKwargs.batch`.
34
"""
35
36
37
-class _MultiModalInputsBase(UserDict[str, NestedTensors]):
37
+class _MultiModalKwargsBase(UserDict[str, NestedTensors]):
38
pass
39
40
41
-class MultiModalInputs(_MultiModalInputsBase):
41
+class MultiModalKwargs(_MultiModalKwargsBase):
42
"""
43
A dictionary that represents the keyword arguments to
44
:meth:`~torch.nn.Module.forward`.
@@ -58,7 +58,7 @@ class MultiModalInputs(_MultiModalInputsBase):
58
if isinstance(nested_tensors, (int, float)):
59
return torch.tensor(nested_tensors)
60
61
- stacked = [MultiModalInputs._try_stack(t) for t in nested_tensors]
61
+ stacked = [MultiModalKwargs._try_stack(t) for t in nested_tensors]
62
if not is_list_of(stacked, torch.Tensor, check="all"):
63
# Only tensors (not lists) can be stacked.
64
return stacked
@@ -71,7 +71,7 @@ class MultiModalInputs(_MultiModalInputsBase):
71
return torch.stack(tensors_)
72
73
@staticmethod
74
- def batch(inputs_list: List["MultiModalInputs"]) -> BatchedTensorInputs:
74
+ def batch(inputs_list: List["MultiModalKwargs"]) -> BatchedTensorInputs:
75
"""
76
Batch multiple inputs together into a dictionary.
77
@@ -95,7 +95,7 @@ class MultiModalInputs(_MultiModalInputsBase):
95
item_lists[k].append(v)
96
97
return {
98
- k: MultiModalInputs._try_stack(item_list)
98
+ k: MultiModalKwargs._try_stack(item_list)
99
for k, item_list in item_lists.items()
100
}
101
@@ -177,7 +177,7 @@ A dictionary containing placeholder ranges.
177
"""
178
179
MultiModalInputMapper = Callable[[InputContext, MultiModalData[object]],
180
- MultiModalInputs]
180
+ MultiModalKwargs]
181
"""
182
Return a dictionary to be passed as keyword arguments to
183
:meth:`~torch.nn.Module.forward`. This is similar in concept to tokenizers
@@ -226,7 +226,7 @@ class MultiModalPlugin(ABC):
226
ctx: InputContext,
227
data: MultiModalData[object],
228
**mm_processor_kwargs,
229
- ) -> MultiModalInputs:
229
+ ) -> MultiModalKwargs:
230
"""
231
Return a dictionary to be passed as keyword arguments to
232
:meth:`~torch.nn.Module.forward`. This is similar in concept to
@@ -275,7 +275,7 @@ class MultiModalPlugin(ABC):
275
model_config: "ModelConfig",
276
data: MultiModalData[object],
277
mm_processor_kwargs: Dict[str, Any],
278
- ) -> MultiModalInputs:
278
+ ) -> MultiModalKwargs:
279
"""
280
Transform the data into a dictionary of model inputs using the
281
input mapper registered for that model.
@@ -585,3 +585,18 @@ class MultiModalPlaceholderMap:
585
586
return MultiModalPlaceholderMap.IndexMap(src=src_indices,
587
dest=dest_indices)
588
+
589
+
590
+def __getattr__(name: str):
591
+ import warnings
592
+
593
+ if name == "MultiModalInputs":
594
+ msg = ("MultiModalInputs has been renamed to MultiModalKwargs. "
595
+ "The original name will take another meaning in an upcoming "
596
+ "version.")
597
+
598
+ warnings.warn(DeprecationWarning(msg), stacklevel=2)
599
+
600
+ return MultiModalKwargs
601
+
602
+ raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
vllm/multimodal/image.py
+5
-5
@@ -10,7 +10,7 @@ from vllm.logger import init_logger
10
from vllm.transformers_utils.processor import get_image_processor
11
from vllm.utils import is_list_of
12
13
-from .base import MultiModalData, MultiModalInputs, MultiModalPlugin
13
+from .base import MultiModalData, MultiModalKwargs, MultiModalPlugin
14
15
if TYPE_CHECKING:
16
from vllm.config import ModelConfig
@@ -43,12 +43,12 @@ class ImagePlugin(MultiModalPlugin):
43
ctx: InputContext,
44
data: MultiModalData[object],
45
**mm_processor_kwargs,
46
- ) -> MultiModalInputs:
46
+ ) -> MultiModalKwargs:
47
model_config = ctx.model_config
48
49
# Processed by input processor
50
if isinstance(data, BatchFeature):
51
- return MultiModalInputs(data.data)
51
+ return MultiModalKwargs(data.data)
52
53
# PIL image
54
if isinstance(data, Image.Image) or is_list_of(data, Image.Image):
@@ -78,11 +78,11 @@ class ImagePlugin(MultiModalPlugin):
78
type(image_processor).__name__)
79
raise
80
81
- return MultiModalInputs(batch_data)
81
+ return MultiModalKwargs(batch_data)
82
83
# Image embedding
84
elif isinstance(data, torch.Tensor) or is_list_of(data, torch.Tensor):
85
- return MultiModalInputs({"image_embeds": data})
85
+ return MultiModalKwargs({"image_embeds": data})
86
87
raise TypeError(f"Invalid image type: {type(data)}")
88
vllm/multimodal/registry.py
+3
-3
@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING, Any, Dict, Mapping, Optional, Sequence
5
from vllm.logger import init_logger
6
7
from .audio import AudioPlugin
8
-from .base import (MultiModalDataDict, MultiModalInputMapper, MultiModalInputs,
8
+from .base import (MultiModalDataDict, MultiModalInputMapper, MultiModalKwargs,
9
MultiModalPlugin, MultiModalTokensCalc, NestedTensors)
10
from .image import ImagePlugin
11
from .video import VideoPlugin
@@ -103,7 +103,7 @@ class MultiModalRegistry:
103
model_config: "ModelConfig",
104
data: MultiModalDataDict,
105
mm_processor_kwargs: Optional[Dict[str, Any]] = None,
106
- ) -> MultiModalInputs:
106
+ ) -> MultiModalKwargs:
107
"""
108
Apply an input mapper to the data passed to the model.
109
@@ -139,7 +139,7 @@ class MultiModalRegistry:
139
140
merged_dict[input_key] = input_tensor
141
142
- return MultiModalInputs(merged_dict)
142
+ return MultiModalKwargs(merged_dict)
143
144
def create_input_mapper(self, model_config: "ModelConfig"):
145
"""
vllm/multimodal/video.py
+3
-3
@@ -9,7 +9,7 @@ from vllm.transformers_utils.processor import get_video_processor
9
from vllm.transformers_utils.tokenizer import get_tokenizer
10
from vllm.utils import is_list_of
11
12
-from .base import MultiModalData, MultiModalInputs
12
+from .base import MultiModalData, MultiModalKwargs
13
from .image import ImagePlugin
14
15
if TYPE_CHECKING:
@@ -55,7 +55,7 @@ class VideoPlugin(ImagePlugin):
55
ctx: InputContext,
56
data: MultiModalData[object],
57
**mm_processor_kwargs,
58
- ) -> MultiModalInputs:
58
+ ) -> MultiModalKwargs:
59
model_config = ctx.model_config
60
61
if isinstance(data, list) and len(data) == 1:
@@ -79,7 +79,7 @@ class VideoPlugin(ImagePlugin):
79
logger.error("Failed to process video (%s)", data)
80
raise
81
82
- return MultiModalInputs(batch_data)
82
+ return MultiModalKwargs(batch_data)
83
84
raise TypeError(f"Invalid video type: {type(data)}")
85
vllm/spec_decode/draft_model_runner.py
+2
-2
@@ -18,7 +18,7 @@ except (ModuleNotFoundError, ImportError) as err:
18
"CUDA and ROCm flash attention backend.") from err
19
20
from vllm.logger import init_logger
21
-from vllm.multimodal import MultiModalInputs
21
+from vllm.multimodal import MultiModalKwargs
22
from vllm.sequence import ExecuteModelRequest, IntermediateTensors
23
from vllm.worker.model_runner import (ModelInputForGPUWithSamplingMetadata,
24
ModelRunner)
@@ -280,7 +280,7 @@ class TP1DraftModelRunner(ModelRunner):
280
kv_caches=kv_caches,
281
attn_metadata=model_input.attn_metadata,
282
intermediate_tensors=intermediate_tensors,
283
- **MultiModalInputs.as_kwargs(multi_modal_kwargs,
283
+ **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
284
device=self.device),
285
**kwargs,
286
)
vllm/worker/cpu_enc_dec_model_runner.py
+2
-2
@@ -5,7 +5,7 @@ import torch
5
6
from vllm.attention import AttentionMetadata
7
from vllm.model_executor.layers.sampler import SamplerOutput
8
-from vllm.multimodal import MultiModalInputs
8
+from vllm.multimodal import MultiModalKwargs
9
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
10
from vllm.utils import make_tensor_with_pad
11
from vllm.worker.cpu_model_runner import (CPUModelRunner,
@@ -287,7 +287,7 @@ class CPUEncoderDecoderModelRunner(CPUModelRunner):
287
kv_caches,
288
"attn_metadata":
289
model_input.attn_metadata,
290
- **MultiModalInputs.as_kwargs(model_input.multi_modal_kwargs or {},
290
+ **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
291
device=self.device),
292
"intermediate_tensors":
293
intermediate_tensors,
vllm/worker/cpu_model_runner.py
+5
-5
@@ -15,7 +15,7 @@ from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
15
from vllm.model_executor.layers.sampler import SamplerOutput
16
from vllm.model_executor.model_loader import get_model
17
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
18
- MultiModalInputs, MultiModalPlaceholderMap)
18
+ MultiModalKwargs, MultiModalPlaceholderMap)
19
from vllm.sequence import (IntermediateTensors, SequenceData,
20
SequenceGroupMetadata)
21
from vllm.utils import make_tensor_with_pad
@@ -200,7 +200,7 @@ class ModelInputForCPUBuilder(ModelRunnerInputBuilderBase[ModelInputForCPU]):
200
201
slot_mapping: List[int] = []
202
seq_lens: List[int] = []
203
- multi_modal_inputs_list: List[MultiModalInputs] = []
203
+ multi_model_kwargs_list: List[MultiModalKwargs] = []
204
multi_modal_placeholder_maps: Dict[
205
str,
206
MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
@@ -225,7 +225,7 @@ class ModelInputForCPUBuilder(ModelRunnerInputBuilderBase[ModelInputForCPU]):
225
._compute_multi_modal_input(
226
seq_group_metadata, seq_data, computed_len,
227
seq_group_metadata.mm_processor_kwargs)
228
- multi_modal_inputs_list.append(mm_kwargs)
228
+ multi_model_kwargs_list.append(mm_kwargs)
229
for modality, placeholder_map in placeholder_maps.items():
230
multi_modal_placeholder_maps[modality].extend(
231
placeholder_map)
@@ -297,7 +297,7 @@ class ModelInputForCPUBuilder(ModelRunnerInputBuilderBase[ModelInputForCPU]):
297
multi_modal_placeholder_index_maps=placeholder_index_maps,
298
)
299
300
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
300
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
301
302
return (input_tokens, input_positions, attn_metadata, seq_lens,
303
multi_modal_kwargs)
@@ -520,7 +520,7 @@ class CPUModelRunner(ModelRunnerBase[ModelInputForCPU]):
520
kv_caches,
521
"attn_metadata":
522
model_input.attn_metadata,
523
- **MultiModalInputs.as_kwargs(model_input.multi_modal_kwargs or {},
523
+ **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
524
device=self.device),
525
"intermediate_tensors":
526
intermediate_tensors,
vllm/worker/embedding_model_runner.py
+2
-2
@@ -8,7 +8,7 @@ from vllm.distributed import get_pp_group
8
from vllm.forward_context import set_forward_context
9
from vllm.logger import init_logger
10
from vllm.model_executor.pooling_metadata import PoolingMetadata
11
-from vllm.multimodal import MultiModalInputs
11
+from vllm.multimodal import MultiModalKwargs
12
from vllm.pooling_params import PoolingParams
13
from vllm.sequence import (IntermediateTensors, PoolerOutput, SequenceData,
14
SequenceGroupMetadata)
@@ -104,7 +104,7 @@ class EmbeddingModelRunner(
104
kv_caches=kv_caches,
105
attn_metadata=model_input.attn_metadata,
106
intermediate_tensors=intermediate_tensors,
107
- **MultiModalInputs.as_kwargs(multi_modal_kwargs,
107
+ **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
108
device=self.device))
109
110
if (self.observability_config is not None
vllm/worker/enc_dec_model_runner.py
+2
-2
@@ -18,7 +18,7 @@ from vllm.logger import init_logger
18
from vllm.model_executor import SamplingMetadata
19
from vllm.model_executor.layers.sampler import SamplerOutput
20
from vllm.model_executor.model_loader.utils import get_architecture_class_name
21
-from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalInputs,
21
+from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalKwargs,
22
MultiModalRegistry)
23
from vllm.sampling_params import SamplingParams
24
from vllm.sequence import (IntermediateTensors, PoolerOutput,
@@ -206,7 +206,7 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
206
kv_caches=kv_caches,
207
attn_metadata=model_input.attn_metadata,
208
intermediate_tensors=intermediate_tensors,
209
- **MultiModalInputs.as_kwargs(multi_modal_kwargs,
209
+ **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
210
device=self.device),
211
**seqlen_agnostic_kwargs)
212
vllm/worker/hpu_model_runner.py
+4
-4
@@ -36,7 +36,7 @@ from vllm.model_executor import SamplingMetadata
36
from vllm.model_executor.layers.sampler import SamplerOutput
37
from vllm.model_executor.model_loader import get_model
38
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
39
- MultiModalInputs)
39
+ MultiModalKwargs)
40
from vllm.sampling_params import SamplingParams
41
from vllm.sequence import (IntermediateTensors, SequenceData,
42
SequenceGroupMetadata)
@@ -716,7 +716,7 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
716
context_lens: List[int] = []
717
query_lens: List[int] = []
718
prefix_block_tables: List[List[int]] = []
719
- multi_modal_inputs_list: List[MultiModalInputs] = []
719
+ multi_model_kwargs_list: List[MultiModalKwargs] = []
720
721
if len(seq_group_metadata_list) == 0:
722
return PreparePromptMetadata.empty()
@@ -777,7 +777,7 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
777
mm_data = seq_group_metadata.multi_modal_data
778
if mm_data:
779
mm_kwargs = self.multi_modal_input_mapper(mm_data)
780
- multi_modal_inputs_list.append(mm_kwargs)
780
+ multi_model_kwargs_list.append(mm_kwargs)
781
782
if seq_group_metadata.block_tables is None:
783
# During memory profiling, the block tables are not initialized
@@ -876,7 +876,7 @@ class HPUModelRunnerBase(ModelRunnerBase[TModelInputForHPU]):
876
multi_modal_placeholder_index_maps=
877
None # FIXME(kzawora): mutli-modality will not work here
878
)
879
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
879
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
880
881
return PreparePromptMetadata(input_tokens=input_tokens,
882
input_positions=input_positions,
vllm/worker/model_runner.py
+9
-9
@@ -38,7 +38,7 @@ from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
38
from vllm.model_executor.models import supports_lora, supports_multimodal
39
from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
40
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
41
- MultiModalInputs, MultiModalPlaceholderMap,
41
+ MultiModalKwargs, MultiModalPlaceholderMap,
42
MultiModalRegistry)
43
from vllm.platforms import current_platform
44
from vllm.prompt_adapter.layers import PromptAdapterMapping
@@ -252,7 +252,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
252
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
253
254
# Multi-modal inputs.
255
- multi_modal_inputs: Optional[MultiModalInputs] = None,
255
+ multi_model_kwargs: Optional[MultiModalKwargs] = None,
256
multi_modal_placeholder_maps: Optional[Dict[
257
str, MultiModalPlaceholderMap]] = None,
258
@@ -373,7 +373,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
373
prompt_adapter_prompt_mapping or [])
374
375
self.prompt_adapter_request = prompt_adapter_request
376
- self.multi_modal_inputs = multi_modal_inputs
376
+ self.multi_model_kwargs = multi_model_kwargs
377
self.multi_modal_placeholder_maps = multi_modal_placeholder_maps
378
self.prefix_cache_hit = prefix_cache_hit
379
@@ -661,7 +661,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
661
mm_kwargs = self.multi_modal_input_mapper(
662
mm_data,
663
mm_processor_kwargs=seq_group_metadata.mm_processor_kwargs)
664
- inter_data.multi_modal_inputs = mm_kwargs
664
+ inter_data.multi_model_kwargs = mm_kwargs
665
inter_data.multi_modal_placeholder_maps = placeholder_maps
666
667
# special processing for mrope position deltas.
@@ -935,11 +935,11 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
935
)
936
937
# Multi-modal data.
938
- multi_modal_inputs_list = [
939
- data.multi_modal_inputs for data in self.inter_data_list
940
- if data.multi_modal_inputs is not None
938
+ multi_model_kwargs_list = [
939
+ data.multi_model_kwargs for data in self.inter_data_list
940
+ if data.multi_model_kwargs is not None
941
]
942
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
942
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
943
944
return self.model_input_cls(
945
input_tokens=input_tokens_tensor,
@@ -1649,7 +1649,7 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
1649
kv_caches=kv_caches,
1650
attn_metadata=model_input.attn_metadata,
1651
intermediate_tensors=intermediate_tensors,
1652
- **MultiModalInputs.as_kwargs(multi_modal_kwargs,
1652
+ **MultiModalKwargs.as_kwargs(multi_modal_kwargs,
1653
device=self.device),
1654
**seqlen_agnostic_kwargs)
1655
vllm/worker/neuron_model_runner.py
+5
-5
@@ -13,7 +13,7 @@ from vllm.model_executor import SamplingMetadata
13
from vllm.model_executor.layers.sampler import SamplerOutput
14
from vllm.model_executor.model_loader.neuron import get_neuron_model
15
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
16
- MultiModalInputs)
16
+ MultiModalKwargs)
17
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
18
from vllm.utils import is_pin_memory_available, make_tensor_with_pad
19
from vllm.worker.model_runner_base import ModelRunnerBase, ModelRunnerInputBase
@@ -122,7 +122,7 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
122
input_block_ids: List[int] = []
123
124
seq_lens: List[int] = []
125
- multi_modal_inputs_list: List[MultiModalInputs] = []
125
+ multi_model_kwargs_list: List[MultiModalKwargs] = []
126
for seq_group_metadata in seq_group_metadata_list:
127
assert seq_group_metadata.is_prompt
128
seq_ids = list(seq_group_metadata.seq_data.keys())
@@ -149,7 +149,7 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
149
mm_data,
150
mm_processor_kwargs=seq_group_metadata.mm_processor_kwargs,
151
)
152
- multi_modal_inputs_list.append(mm_kwargs)
152
+ multi_model_kwargs_list.append(mm_kwargs)
153
154
max_seq_len = max(seq_lens)
155
assert max_seq_len > 0
@@ -167,7 +167,7 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
167
dtype=torch.long,
168
device=self.device)
169
170
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
170
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
171
172
return (input_tokens, input_positions, input_block_ids, seq_lens,
173
multi_modal_kwargs)
@@ -314,7 +314,7 @@ class NeuronModelRunner(ModelRunnerBase[ModelInputForNeuron]):
314
input_ids=model_input.input_tokens,
315
positions=model_input.input_positions,
316
input_block_ids=model_input.input_block_ids,
317
- **MultiModalInputs.as_kwargs(model_input.multi_modal_kwargs or {},
317
+ **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
318
device=self.device),
319
)
320
vllm/worker/openvino_model_runner.py
+5
-5
@@ -13,7 +13,7 @@ from vllm.model_executor import SamplingMetadata
13
from vllm.model_executor.layers.sampler import SamplerOutput
14
from vllm.model_executor.model_loader.openvino import get_model
15
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
16
- MultiModalInputs, MultiModalPlaceholderMap)
16
+ MultiModalKwargs, MultiModalPlaceholderMap)
17
from vllm.sequence import SequenceGroupMetadata
18
from vllm.worker.model_runner_base import ModelRunnerBase
19
@@ -102,7 +102,7 @@ class OpenVINOModelRunner(ModelRunnerBase):
102
seq_lens: List[int] = []
103
past_lens: List[int] = []
104
query_lens: List[int] = []
105
- multi_modal_inputs_list: List[MultiModalInputs] = []
105
+ multi_model_kwargs_list: List[MultiModalKwargs] = []
106
multi_modal_placeholder_maps: Dict[
107
str,
108
MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
@@ -226,7 +226,7 @@ class OpenVINOModelRunner(ModelRunnerBase):
226
mm_data,
227
mm_processor_kwargs=seq_group_metadata.
228
mm_processor_kwargs)
229
- multi_modal_inputs_list.append(mm_kwargs)
229
+ multi_model_kwargs_list.append(mm_kwargs)
230
231
for modality, placeholder_map in placeholder_maps.items():
232
multi_modal_placeholder_maps[modality].extend(
@@ -275,7 +275,7 @@ class OpenVINOModelRunner(ModelRunnerBase):
275
multi_modal_placeholder_index_maps=placeholder_index_maps,
276
)
277
278
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
278
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
279
280
return ModelInput(
281
input_tokens,
@@ -341,7 +341,7 @@ class OpenVINOModelRunner(ModelRunnerBase):
341
kv_caches,
342
"attn_metadata":
343
attn_metadata,
344
- **MultiModalInputs.as_kwargs(multi_modal_kwargs or {},
344
+ **MultiModalKwargs.as_kwargs(multi_modal_kwargs or {},
345
device=self.device),
346
}
347
vllm/worker/xpu_model_runner.py
+5
-5
@@ -18,7 +18,7 @@ from vllm.model_executor import SamplingMetadataCache
18
from vllm.model_executor.layers.sampler import SamplerOutput
19
from vllm.model_executor.model_loader import get_model
20
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
21
- MultiModalInputs, MultiModalPlaceholderMap,
21
+ MultiModalKwargs, MultiModalPlaceholderMap,
22
MultiModalRegistry)
23
from vllm.sampling_params import SamplingParams
24
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
@@ -160,7 +160,7 @@ class ModelInputForXPUBuilder(ModelRunnerInputBuilderBase[ModelInputForXPU]):
160
input_positions: List[int] = []
161
slot_mapping: List[int] = []
162
seq_lens: List[int] = []
163
- multi_modal_inputs_list: List[MultiModalInputs] = []
163
+ multi_model_kwargs_list: List[MultiModalKwargs] = []
164
multi_modal_placeholder_maps: Dict[
165
str,
166
MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
@@ -192,7 +192,7 @@ class ModelInputForXPUBuilder(ModelRunnerInputBuilderBase[ModelInputForXPU]):
192
.from_seq_group(seq_group_metadata, positions_range)
193
194
mm_kwargs = self.runner.multi_modal_input_mapper(mm_data)
195
- multi_modal_inputs_list.append(mm_kwargs)
195
+ multi_model_kwargs_list.append(mm_kwargs)
196
197
for modality, placeholder_map in placeholder_maps.items():
198
multi_modal_placeholder_maps[modality].extend(
@@ -264,7 +264,7 @@ class ModelInputForXPUBuilder(ModelRunnerInputBuilderBase[ModelInputForXPU]):
264
block_tables=torch.tensor([], device=self.device, dtype=torch.int),
265
)
266
267
- multi_modal_kwargs = MultiModalInputs.batch(multi_modal_inputs_list)
267
+ multi_modal_kwargs = MultiModalKwargs.batch(multi_model_kwargs_list)
268
269
return (input_tokens, input_positions, attn_metadata, seq_lens,
270
multi_modal_kwargs)
@@ -565,7 +565,7 @@ class XPUModelRunner(ModelRunnerBase[ModelInputForXPUWithSamplingMetadata]):
565
kv_caches=kv_caches,
566
attn_metadata=model_input.attn_metadata,
567
intermediate_tensors=intermediate_tensors,
568
- **MultiModalInputs.as_kwargs(model_input.multi_modal_kwargs or {},
568
+ **MultiModalKwargs.as_kwargs(model_input.multi_modal_kwargs or {},
569
device=self.device))
570
# Compute the logits in the last pipeline stage.
571
if not get_pp_group().is_last_rank: