bridge.data.vlm_datasets.step37_flickr8k.packing

Greedy non-truncation packing.

Walks the sample sizes in order and greedily fills each pack up to max_len without ever truncating a sample. Do NOT modify the loop arithmetic or the flush() semantics — the exact sequence of “drop” vs “extend” decisions determines the contents of every pack, and any reordering would shift the entire downstream packing layout.

Module Contents

Classes

PackingResult

Result metadata from greedy sample packing.

Functions

pack	Pack ordered sample lengths into contiguous groups without truncation.

API

class bridge.data.vlm_datasets.step37_flickr8k.packing.PackingResult

Result metadata from greedy sample packing.

num_packed_samples: int

None

num_droped: int

None

packed_sample_ranges: list[tuple[int, int]]

None

Offset & Num of samples packed for each packed-sample, e.g. [(0, 2), (3, 2), …]

bridge.data.vlm_datasets.step37_flickr8k.packing.pack(

sizes: list[int],

max_len: int,

oversize_policy: Literal[drop, extend],

) → bridge.data.vlm_datasets.step37_flickr8k.packing.PackingResult

Pack ordered sample lengths into contiguous groups without truncation.

Parameters:

• sizes – Token lengths for the samples to pack.

• max_len – Maximum packed sequence length.

• oversize_policy – Whether to drop oversize samples or keep them in extended packs.

Returns:

Metadata describing the packed sample ranges and dropped sample count.