lmms-eval uses the MVBench's template, but llava-video uses the different one, so just follow the PerceptionTest's template

Normalize the timestamps and points in the triplets to a range of [0, 100].

Pack examples and compute image shard boundaries for context parallelism.

Changed from MMMU-style complex parsing into simple parsing. Fixed to avoid 'D. A book' be parsed as A. Same as original Long

Parse model prediction text to standardized trajectory format. Supported prediction types: - video_point_track_all_frames

Parse timestamp string like '00:00:01.089' to seconds.

Format question text with options.

Reassign token locations using first-fit descending greedy algorithm. The kept tokens will be rearranged into N slots per frame as packe

Reindexes numbered lines in a string to start from 1. Args: text (str): Multiline string with lines starting with a number and a per

This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, he

torch.repeat_interleave(x, dim=2, repeats=n_rep)

Re-initialize the weights from scratch

Rotates half the hidden dims of the input.

Randomly choose a clip [clip_start, clip_end] such that: - 0 <= clip_start <= start_time <= end_time <= clip_end <= video_duration -

Seed all rng objects.

Divide in image of size [w, h] in up to max_num_patches of size patch_size

This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths, not start and end indices.

This function adds the correct offsets to the `q_idx` and `kv_idx` as the torch API can only accept lengths, not start and end indices.

Remove deprecated keys from old checkpoints.

Scan CLIP_DIR to build extensions.json mapping video IDs to file extensions. Clip filenames follow {vid}_{start}_{end}.{ext}. We parse from t

Walk VIDEO_DIR once and build a mapping from filename stem to full path.

Override to add native_fps for frame subsampling on full video encoding.

Build work items for _create_videos from the fps map. Returns list of dicts: [{frames_dir, output_path, fps, ...}]. Override for data

Compute similarity between predictions and GT for one frame. Similarity = 1 if point in mask, 0 otherwise. Returns:

Copies the `state` . This method copies the data in `state` using `copy.deepcopy()`. If this is not appropriate for your constraint,

Create clips from full videos (symlink if no trimming, ffmpeg trim otherwise).

Create message list with points organized by frame.

Create an MP4 from numbered frames (e.g., 00001.jpg to 00051.jpg) in a directory.

Encode frame dirs to full videos using multiprocessing.

Decode input videos and sample frames if needed.

Decode an RLE-encoded mask into a 2D numpy array. Handles both compressed (string counts via pycocotools) and uncompressed (list-of-int count

Download and extract koala_meta.zip from HF if not already present.

Auto-download vimeo videos from HF; warn about manual downloads for others.

Extract videos to frames, optionally in parallel.

Extract videos to frames, optionally in parallel.

Find the original human-readable label for a normalized label.

Find the original human-readable label for a normalized label.

Find video path for a Molmo2-Cap video_id. Searches VIDEO_CAP_SUBDIRS under video_dir. Handles three layouts: - Folder-based (youtube-cc-*)

Look up video path from the pre-built index.

Format as: id: ([x1,y1,t1], [x2,y2,t2])

Format as time {t}\n{id: [x, y, occluded], ...}

Format as: time {t}\n{id: [x, y], ...}

Format as: [x1,y1,t1], [x2,y2,t2], ...

Update kwargs that need further processing before being validated Can be overridden by subclasses to customize the processing of kwar