biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Doc] Convert Sphinx directives ( {class}, {meth}, {attr}, ...) to MkDocs format for better documentation linking (#18663)

Browse Source 
Signed-off-by: Zerohertz <ohg3417@gmail.com>
This commit is contained in:
Hyogeun Oh (오효근)
2025-05-27 17:44:20 +09:00
committed by GitHub
parent 6881107948
commit a68e293cb9
37 changed files with 360 additions and 247 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
vllm/multimodal/processing.py
View File

@@ -114,13 +114,14 @@ class PromptUpdateDetails(Generic[_S]):
is_embed: Optional[Callable[["_BoundPromptSequence"], torch.Tensor]] = None
"""
Given {attr}`full`, return a boolean mask of shape `(len(full),)`
indicating which positions of `full` to assign embeddings to.
Given [`full`][vllm.multimodal.processing.PromptUpdateDetails.full],
return a boolean mask of shape `(len(full),)` indicating which positions
of `full` to assign embeddings to.
`None` (default) means to assign embeddings to all positions of `full`.
The embeddings are obtained by calling
{class}`SupportsMultiModal.get_multimodal_embeddings`.
[`SupportsMultiModal.get_multimodal_embeddings`][vllm.model_executor.models.interfaces.SupportsMultiModal.get_multimodal_embeddings].
"""
@staticmethod
@@ -159,13 +160,15 @@ PromptUpdateInfo = Union[PromptSeq, PromptUpdateDetails]
The token sequence or text that are part of the update.
If only part of the content corresponds to feature placeholders, you can
use {class}`PromptUpdateDetails` to specify which part.
use [`PromptUpdateDetails`][vllm.multimodal.processing.PromptUpdateDetails] to
specify which part.
"""
PromptUpdateContent = Union[Callable[[int], PromptUpdateInfo],
PromptUpdateInfo]
"""
Given the index of the processed item within {attr}`modality`,
Given the index of the processed item within
[`modality`][vllm.multimodal.processing.PromptUpdate.modality],
output the corresponding token sequence (or text).
For convenience, you can directly pass in the token sequence (or text)
@@ -260,8 +263,10 @@ class PromptInsertion(PromptUpdate):
insertion: PromptUpdateContent = field(repr=False)
"""
Given the index of the processed item within {attr}`modality`,
output the token sequence (or text) to insert right after {attr}`target`.
Given the index of the processed item within
[`modality`][vllm.multimodal.processing.PromptUpdate.modality],
output the token sequence (or text) to insert right after
[`target`][vllm.multimodal.processing.PromptUpdate.target].
For convenience, you can directly pass in the token sequence (or text)
instead of a function if it does not depend on the input.
@@ -332,8 +337,10 @@ class PromptReplacement(PromptUpdate):
replacement: PromptUpdateContent = field(repr=False)
"""
Given the index of the processed item within {attr}`modality`,
output the token sequence (or text) to replace {attr}`target`.
Given the index of the processed item within
[`modality`][vllm.multimodal.processing.PromptUpdate.modality],
output the token sequence (or text) to replace
[`target`][vllm.multimodal.processing.PromptUpdate.target].
For convenience, you can directly pass in the token sequence (or text)
instead of a function if it does not depend on the input.
@@ -387,14 +394,16 @@ _M = TypeVar("_M", bound=Union[_HasModalityAttr, _HasModalityProp])
def full_groupby_modality(values: Iterable[_M]) -> ItemsView[str, list[_M]]:
"""Convenience function to apply [full_groupby][] based on modality."""
"""Convenience function to apply [`full_groupby`][vllm.utils.full_groupby]
based on modality."""
return full_groupby(values, key=lambda x: x.modality)
@dataclass
class _BoundPromptSequence:
"""
A {data}`_PromptSeq` bound to a tokenizer to automatically
A [`_PromptSeq`][vllm.multimodal.processing.PromptSeq] bound
to a tokenizer to automatically
convert between token sequence and text representations.
"""
tokenizer: AnyTokenizer = field(repr=False)
@@ -446,9 +455,11 @@ class _BoundPromptContent:
@dataclass
class BoundPromptUpdate:
"""
A {class}`PromptUpdate` bound to a tokenizer to automatically convert
{attr}`target` and the result of {meth}`get_content` between
token sequence and text representations.
A [`PromptUpdate`][vllm.multimodal.processing.PromptUpdate] bound
to a tokenizer to automatically convert
[`target`][vllm.multimodal.processing.PromptUpdate.target] and the result of
[`get_content`][vllm.multimodal.processing.BoundPromptUpdate.get_content]
between token sequence and text representations.
"""
_origin: PromptUpdate
tokenizer: AnyTokenizer = field(repr=False)
@@ -482,7 +493,8 @@ class BoundPromptUpdate:
def get_content(self, item_idx: int) -> _BoundPromptContent:
"""
Given the index of the processed item within {attr}`modality`,
Given the index of the processed item within
[`modality`][vllm.multimodal.processing.PromptUpdate.modality],
output the token sequence (or text) to update.
"""
content = self.content
@@ -1019,7 +1031,8 @@ class ProcessingCache:
) -> None:
"""
Put a processed multi-modal item into the cache
according to its dependencies (see {meth}`get`).
according to its dependencies
(see [`get`][vllm.multimodal.processing.ProcessingCache.get]).
"""
cache_key = MultiModalHasher.hash_kwargs(model_id=model_id,
**{modality: input_item},
@@ -1091,7 +1104,8 @@ _I = TypeVar("_I", bound=BaseProcessingInfo)
MultiModalHashes = dict[str, list[str]]
"""
A collection of hashes with a similar structure as {class}`MultiModalKwargs`.
A collection of hashes with a similar structure as
[`MultiModalKwargs`][vllm.multimodal.inputs.MultiModalKwargs].
"""
@@ -1099,7 +1113,7 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
"""
Abstract base class to process multi-modal inputs to be used in vLLM.
Not to be confused with {class}`transformers.ProcessorMixin`.
Not to be confused with `transformers.ProcessorMixin`.
"""
def __init__(self,
@@ -1126,10 +1140,12 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
def _get_data_parser(self) -> MultiModalDataParser:
"""
Construct a parser to preprocess multi-modal data items
before passing them to {meth}`_get_hf_mm_data`.
before passing them to
[`_get_hf_mm_data`][vllm.multimodal.processing.BaseMultiModalProcessor._get_hf_mm_data].
You can support additional modalities by creating a subclass
of {class}`MultiModalDataParser` that has additional subparsers.
of [`MultiModalDataParser`][vllm.multimodal.parse.MultiModalDataParser]
that has additional subparsers.
"""
return MultiModalDataParser()
@@ -1138,8 +1154,11 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
mm_data: MultiModalDataDict,
) -> MultiModalDataItems:
"""
Normalize {class}`MultiModalDataDict` to {class}`MultiModalDataItems`
before passing them to {meth}`_get_hf_mm_data`.
Normalize
[`MultiModalDataDict`][vllm.multimodal.inputs.MultiModalDataDict]
to [`MultiModalDataItems`][vllm.multimodal.parse.MultiModalDataItems]
before passing them to
[`_get_hf_mm_data`][vllm.multimodal.processing.BaseMultiModalProcessor._get_hf_mm_data].
"""
mm_items = self.data_parser.parse_mm_data(mm_data)
supported_mm_limits = self.info.get_supported_mm_limits()
@@ -1191,7 +1210,8 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
inputs.
Moreover, this information is critical to determine the token positions
in order to construct {class}`~vllm-multimodal.input.PlaceholderRange`
in order to construct
[`PlaceholderRange`][vllm.multimodal.inputs.PlaceholderRange]
for each multi-modal item.
"""
raise NotImplementedError
@@ -1315,7 +1335,9 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
Most HF processors accept prompt text but not prompt tokens.
If the HF processor adds or removes tokens that are not related to
multi-modal data, you should override this method so it is consistent
with the output of {meth}`_apply_hf_processor_text_only` on the
with the output of
[`_apply_hf_processor_text_only`][vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_text_only]
on the
corresponding text.
"""
return prompt_tokens
@@ -1330,7 +1352,8 @@ class BaseMultiModalProcessor(ABC, Generic[_I]):
Since HF processor requires that text and multi-modal items
correspond to each other, we generate dummy text using
{class}`DummyInputsBuilder` to go along with the multi-modal data.
[`DummyInputsBuilder`][vllm.multimodal.profiling.BaseDummyInputsBuilder]
to go along with the multi-modal data.
"""
mm_counts = mm_items.get_all_counts()