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, 02:00 UTC
[Misc] Rename `MultiModalInputsV2 -> MultiModalInputs` (#12244)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jan 21, 2025 at 15:31 UTC
96912550c8399af2632f3f6830f7c3fa9e10a75a
12 files changed
+31
-31
docs/source/api/multimodal/inputs.md
+1
-1
@@ -43,7 +43,7 @@
43
```
44
45
```{eval-rst}
46
-.. autoclass:: vllm.multimodal.inputs.MultiModalInputsV2
46
+.. autoclass:: vllm.multimodal.inputs.MultiModalInputs
47
:members:
48
:show-inheritance:
49
```
vllm/inputs/data.py
+6
-6
@@ -9,7 +9,7 @@ from typing_extensions import NotRequired, TypedDict, TypeVar, assert_never
9
if TYPE_CHECKING:
10
from vllm.multimodal import (MultiModalDataDict, MultiModalKwargs,
11
MultiModalPlaceholderDict)
12
- from vllm.multimodal.inputs import MultiModalInputsV2
12
+ from vllm.multimodal.inputs import MultiModalInputs
13
14
15
class TextPrompt(TypedDict):
@@ -207,7 +207,7 @@ def token_inputs(
207
return inputs
208
209
210
-DecoderOnlyInputs = Union[TokenInputs, "MultiModalInputsV2"]
210
+DecoderOnlyInputs = Union[TokenInputs, "MultiModalInputs"]
211
"""
212
The inputs in :class:`~vllm.LLMEngine` before they are
213
passed to the model executor.
@@ -222,14 +222,14 @@ class EncoderDecoderInputs(TypedDict):
222
223
This specifies the required data for encoder-decoder models.
224
"""
225
- encoder: Union[TokenInputs, "MultiModalInputsV2"]
225
+ encoder: Union[TokenInputs, "MultiModalInputs"]
226
"""The inputs for the encoder portion."""
227
228
- decoder: Union[TokenInputs, "MultiModalInputsV2"]
228
+ decoder: Union[TokenInputs, "MultiModalInputs"]
229
"""The inputs for the decoder portion."""
230
231
232
-SingletonInputs = Union[TokenInputs, "MultiModalInputsV2"]
232
+SingletonInputs = Union[TokenInputs, "MultiModalInputs"]
233
"""
234
A processed :class:`SingletonPrompt` which can be passed to
235
:class:`vllm.sequence.Sequence`.
@@ -311,7 +311,7 @@ class SingletonInputsAdapter:
311
return inputs.get("multi_modal_hashes", [])
312
313
if inputs["type"] == "multimodal":
314
- # only the case when we use MultiModalInputsV2
314
+ # only the case when we use MultiModalInputs
315
return inputs.get("mm_hashes", []) # type: ignore[return-value]
316
317
assert_never(inputs) # type: ignore[arg-type]
vllm/inputs/preprocess.py
+3
-3
@@ -7,7 +7,7 @@ from vllm.config import ModelConfig
7
from vllm.logger import init_logger
8
from vllm.lora.request import LoRARequest
9
from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry
10
-from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputsV2
10
+from vllm.multimodal.inputs import MultiModalDataDict, MultiModalInputs
11
from vllm.prompt_adapter.request import PromptAdapterRequest
12
from vllm.transformers_utils.tokenizer_group import BaseTokenizerGroup
13
@@ -247,7 +247,7 @@ class InputPreprocessor:
247
mm_data: MultiModalDataDict,
248
mm_processor_kwargs: Optional[Mapping[str, object]],
249
lora_request: Optional[LoRARequest],
250
- ) -> MultiModalInputsV2:
250
+ ) -> MultiModalInputs:
251
"""
252
Apply the model's multi-modal processor to a multi-modal prompt,
253
returning the corresponding token IDs and metadata.
@@ -271,7 +271,7 @@ class InputPreprocessor:
271
mm_data: MultiModalDataDict,
272
mm_processor_kwargs: Optional[Mapping[str, object]],
273
lora_request: Optional[LoRARequest],
274
- ) -> MultiModalInputsV2:
274
+ ) -> MultiModalInputs:
275
"""Async version of :meth:`_process_multimodal`."""
276
tokenizer_group = self.get_tokenizer_group()
277
tokenizer = await tokenizer_group.get_lora_tokenizer_async(lora_request
vllm/model_executor/models/blip2.py
+2
-2
@@ -15,7 +15,7 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
15
from vllm.model_executor.sampling_metadata import SamplingMetadata
16
from vllm.multimodal import MULTIMODAL_REGISTRY
17
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
18
- MultiModalInputsV2, MultiModalKwargs,
18
+ MultiModalInputs, MultiModalKwargs,
19
NestedTensors, PlaceholderRange)
20
from vllm.multimodal.parse import MultiModalDataItems
21
from vllm.multimodal.processing import (BaseMultiModalProcessor,
@@ -490,7 +490,7 @@ class Blip2MultiModalProcessor(BaseMultiModalProcessor[Blip2ProcessingInfo]):
490
prompt: Union[str, list[int]],
491
mm_data: MultiModalDataDict,
492
hf_processor_mm_kwargs: Mapping[str, object],
493
- ) -> MultiModalInputsV2:
493
+ ) -> MultiModalInputs:
494
result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
495
496
# Only <image> tokens should be considered as placeholders,
vllm/model_executor/models/chameleon.py
+2
-2
@@ -29,7 +29,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
29
from vllm.model_executor.utils import set_weight_attrs
30
from vllm.multimodal import MULTIMODAL_REGISTRY
31
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
32
- MultiModalInputsV2, MultiModalKwargs,
32
+ MultiModalInputs, MultiModalKwargs,
33
NestedTensors, PlaceholderRange)
34
from vllm.multimodal.parse import MultiModalDataItems
35
from vllm.multimodal.processing import (BaseMultiModalProcessor,
@@ -159,7 +159,7 @@ class ChameleonMultiModalProcessor(
159
prompt: Union[str, list[int]],
160
mm_data: MultiModalDataDict,
161
hf_processor_mm_kwargs: Mapping[str, object],
162
- ) -> MultiModalInputsV2:
162
+ ) -> MultiModalInputs:
163
result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
164
165
# Only <image> tokens should be considered as placeholders,
vllm/model_executor/models/fuyu.py
+2
-2
@@ -31,7 +31,7 @@ from vllm.model_executor.models.persimmon import PersimmonForCausalLM
31
from vllm.model_executor.sampling_metadata import SamplingMetadata
32
from vllm.multimodal import MULTIMODAL_REGISTRY
33
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
34
- MultiModalInputsV2, MultiModalKwargs,
34
+ MultiModalInputs, MultiModalKwargs,
35
NestedTensors, PlaceholderRange)
36
from vllm.multimodal.parse import (ImageProcessorItems, ImageSize,
37
MultiModalDataItems)
@@ -232,7 +232,7 @@ class FuyuMultiModalProcessor(BaseMultiModalProcessor[FuyuProcessingInfo]):
232
prompt: Union[str, list[int]],
233
mm_data: MultiModalDataDict,
234
hf_processor_mm_kwargs: Mapping[str, object],
235
- ) -> MultiModalInputsV2:
235
+ ) -> MultiModalInputs:
236
result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
237
238
# Only |SPEAKER| (image) tokens should be considered as placeholders,
vllm/model_executor/models/llava.py
+3
-3
@@ -24,7 +24,7 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
24
from vllm.model_executor.sampling_metadata import SamplingMetadata
25
from vllm.multimodal import MULTIMODAL_REGISTRY
26
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
27
- MultiModalInputsV2, MultiModalKwargs,
27
+ MultiModalInputs, MultiModalKwargs,
28
NestedTensors)
29
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
30
ImageSize, MultiModalDataItems)
@@ -746,7 +746,7 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
746
prompt: Union[str, list[int]],
747
mm_data: MultiModalDataDict,
748
hf_processor_mm_kwargs: Mapping[str, object],
749
- ) -> MultiModalInputsV2:
749
+ ) -> MultiModalInputs:
750
hf_config = self.info.get_hf_config()
751
image_token_id = hf_config.image_token_index
752
@@ -805,7 +805,7 @@ class MantisMultiModalProcessor(LlavaMultiModalProcessor):
805
for modality, placeholders in mm_placeholders.items()
806
}
807
808
- return MultiModalInputsV2(
808
+ return MultiModalInputs(
809
type="multimodal",
810
prompt=prompt,
811
prompt_token_ids=prompt_ids,
vllm/model_executor/models/phi3v.py
+2
-2
@@ -31,7 +31,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
31
from vllm.model_executor.sampling_metadata import SamplingMetadata
32
from vllm.multimodal import MULTIMODAL_REGISTRY
33
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
34
- MultiModalInputsV2, MultiModalKwargs,
34
+ MultiModalInputs, MultiModalKwargs,
35
NestedTensors, PlaceholderRange)
36
from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
37
ImageSize, MultiModalDataItems)
@@ -484,7 +484,7 @@ class Phi3VMultiModalProcessor(BaseMultiModalProcessor[Phi3VProcessingInfo]):
484
prompt: Union[str, list[int]],
485
mm_data: MultiModalDataDict,
486
hf_processor_mm_kwargs: Mapping[str, object],
487
- ) -> MultiModalInputsV2:
487
+ ) -> MultiModalInputs:
488
result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
489
490
# Only <|image|> tokens should be considered as placeholders,
vllm/model_executor/models/qwen2_audio.py
+2
-2
@@ -37,7 +37,7 @@ from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
37
from vllm.model_executor.sampling_metadata import SamplingMetadata
38
from vllm.multimodal import MULTIMODAL_REGISTRY
39
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
40
- MultiModalInputsV2, MultiModalKwargs,
40
+ MultiModalInputs, MultiModalKwargs,
41
NestedTensors, PlaceholderRange)
42
from vllm.multimodal.parse import (AudioProcessorItems, MultiModalDataItems,
43
MultiModalDataParser)
@@ -245,7 +245,7 @@ class Qwen2AudioMultiModalProcessor(
245
prompt: Union[str, list[int]],
246
mm_data: MultiModalDataDict,
247
hf_processor_mm_kwargs: Mapping[str, object],
248
- ) -> MultiModalInputsV2:
248
+ ) -> MultiModalInputs:
249
result = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
250
251
# Only <|AUDIO|> tokens should be considered as placeholders,
vllm/multimodal/inputs.py
+1
-1
@@ -491,7 +491,7 @@ A dictionary containing placeholder ranges for each modality.
491
"""
492
493
494
-class MultiModalInputsV2(TypedDict):
494
+class MultiModalInputs(TypedDict):
495
"""
496
Represents the outputs of
497
:class:`vllm.multimodal.processing.BaseMultiModalProcessor`,
vllm/multimodal/processing.py
+5
-5
@@ -18,8 +18,8 @@ from vllm.utils import LRUCache, flatten_2d_lists, full_groupby
18
19
from .hasher import MultiModalHasher
20
from .inputs import (MultiModalDataDict, MultiModalFieldConfig,
21
- MultiModalInputsV2, MultiModalKwargs,
22
- MultiModalKwargsItem, PlaceholderRange)
21
+ MultiModalInputs, MultiModalKwargs, MultiModalKwargsItem,
22
+ PlaceholderRange)
23
from .parse import MultiModalDataItems, MultiModalDataParser
24
25
if TYPE_CHECKING:
@@ -609,7 +609,7 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
609
prompt: str,
610
mm_data: MultiModalDataDict,
611
hf_processor_mm_kwargs: Mapping[str, object],
612
- ) -> MultiModalInputsV2:
612
+ ) -> MultiModalInputs:
613
return self.apply(prompt, mm_data, hf_processor_mm_kwargs)
614
615
def _get_data_parser(self) -> MultiModalDataParser:
@@ -1067,7 +1067,7 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
1067
prompt: Union[str, list[int]],
1068
mm_data: MultiModalDataDict,
1069
hf_processor_mm_kwargs: Mapping[str, object],
1070
- ) -> MultiModalInputsV2:
1070
+ ) -> MultiModalInputs:
1071
"""
1072
Process multi-modal inputs to be used in vLLM.
1073
@@ -1169,7 +1169,7 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
1169
for modality, placeholders in mm_placeholders.items()
1170
}
1171
1172
- return MultiModalInputsV2(
1172
+ return MultiModalInputs(
1173
type="multimodal",
1174
prompt=prompt,
1175
prompt_token_ids=prompt_ids,
vllm/multimodal/profiling.py
+2
-2
@@ -11,7 +11,7 @@ import vllm.envs as envs
11
from vllm.inputs import DummyData
12
from vllm.logger import init_logger
13
14
-from .inputs import MultiModalDataDict, MultiModalInputsV2
14
+from .inputs import MultiModalDataDict, MultiModalInputs
15
from .processing import BaseMultiModalProcessor, BaseProcessingInfo
16
17
logger = init_logger(__name__)
@@ -131,7 +131,7 @@ class MultiModalProfiler(Generic[_I]):
131
self,
132
seq_len: int,
133
mm_counts: Mapping[str, int],
134
- ) -> MultiModalInputsV2:
134
+ ) -> MultiModalInputs:
135
factory = self.dummy_inputs
136
processor_inputs = factory.get_dummy_processor_inputs(
137
seq_len, mm_counts)