[Perf] Improve multimodal item handling from O(n) to O(log n) per step (#44212)

Signed-off-by: Andy Lo <andy@mistral.ai>
This commit is contained in:
Andy Lo
2026-06-03 11:00:26 +00:00
committed by GitHub
parent 1fa9ea09f6
commit 95b1615ec9
6 changed files with 92 additions and 44 deletions
+25
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import bisect
import mimetypes
from collections import defaultdict
from collections.abc import Generator, Sequence
@@ -18,6 +19,7 @@ from vllm.utils.import_utils import LazyLoader
from .hasher import MultiModalHasher
from .inputs import (
BatchedTensorInputs,
MultiModalFeatureSpec,
MultiModalFieldElem,
MultiModalKwargsItem,
MultiModalSharedField,
@@ -109,6 +111,29 @@ def encode_video_url(
return f"data:{mimetype};base64,{video_b64}"
def get_mm_features_in_window(
mm_features: list[MultiModalFeatureSpec],
start: int,
end: int,
) -> tuple[int, int]:
"""Return (lo, hi) indices for features overlapping [start, end).
Assumes mm_features are sorted by offset and non-overlapping, so
offset + length is also sorted.
"""
lo = bisect.bisect_left(
mm_features,
start + 1,
key=lambda f: f.mm_position.offset + f.mm_position.length,
)
hi = bisect.bisect_left(
mm_features,
end,
key=lambda f: f.mm_position.offset,
)
return lo, hi
def argsort_mm_positions(
mm_positions: MultiModalPlaceholders,
) -> list[tuple[str, int]]: