[mypy] Further improve MM type annotations (#25654)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-09-25 10:57:36 +00:00
committed by GitHub
parent 17b4c6685c
commit 12c1287d64
6 changed files with 90 additions and 48 deletions
+15 -8
View File
@@ -14,7 +14,7 @@ import numpy as np
from typing_extensions import NotRequired, TypeAlias, TypeVar, deprecated
from vllm.utils import LazyLoader, full_groupby, is_list_of
from vllm.utils.jsontree import JSONTree, json_map_leaves
from vllm.utils.jsontree import json_map_leaves
if TYPE_CHECKING:
import torch
@@ -203,7 +203,7 @@ def nested_tensors_equal(a: NestedTensors, b: NestedTensors) -> bool:
return a == b
BatchedTensorInputs: TypeAlias = Mapping[str, NestedTensors]
BatchedTensorInputs: TypeAlias = dict[str, NestedTensors]
"""
A dictionary containing nested tensors which have been batched via
[`MultiModalKwargs.batch`][vllm.multimodal.inputs.MultiModalKwargs.batch].
@@ -377,6 +377,7 @@ class MultiModalBatchedField(BaseMultiModalField):
pin_memory: bool,
) -> NestedTensors:
if len(batch) > 0 and is_list_of(batch, torch.Tensor, check="all"):
batch = cast(list[torch.Tensor], batch)
if len(batch) == 1:
# An optimization when `batch` contains only one tensor:
# - produce exactly same result as `torch.stack(batch)`
@@ -422,6 +423,7 @@ class MultiModalFlatField(BaseMultiModalField):
pin_memory: bool,
) -> NestedTensors:
if len(batch) > 0 and is_list_of(batch, torch.Tensor, check="all"):
batch = cast(list[torch.Tensor], batch)
if len(batch) == 1:
# An optimization when `batch` contains only one tensor:
# - produce exactly same result as `torch.concat(batch)`
@@ -764,6 +766,15 @@ class MultiModalKwargsItems(UserDict[str, Sequence[_I]]):
return super().__getitem__(modality) # type: ignore[return-value]
def require_data(self) -> "MultiModalKwargsItems[MultiModalKwargsItem]":
for modality, items in self.items():
for i, item in enumerate(items):
if item is None:
raise RuntimeError(
f"Found empty mm_items[{modality}][{i}]")
return self # type: ignore[return-value]
def get_data(self, *, pin_memory: bool = False) -> "MultiModalKwargs":
elems_by_key = defaultdict[str, list[MultiModalFieldElem]](list)
for modality, items in self.items():
@@ -897,15 +908,11 @@ class MultiModalKwargs(UserDict[str, NestedTensors]):
*,
device: torch.types.Device,
) -> BatchedTensorInputs:
json_inputs = cast(JSONTree[torch.Tensor], batched_inputs)
json_mapped = json_map_leaves(
return json_map_leaves(
lambda x: x.to(device=device, non_blocking=True),
json_inputs,
batched_inputs,
)
return cast(BatchedTensorInputs, json_mapped)
def __getitem__(self, key: str):
if key not in self:
raise KeyError(f"Keyword argument {key!r} not found. "