`nemo_rl.models.generation.vllm.video_utils`#

Module Contents#

Functions#

`_round_video_frame_count`
`_timestamp_to_video_frame_index`	Convert a timestamp according to the configured sampling contract.
`_select_video_frame_count`
`_compute_video_timestamps`
`_build_video_metadata`
`_resolve_cached_video_media_path`
`build_cached_video_frame_metadata`	Return the shared synthetic timing contract for cached video frames.
`build_cached_video_frame_data_url`	Build a compact native-video URL backed by lossless cached PNG frames.
`_load_cached_video_frame_manifest`	Load an internal cached-frame manifest passed through vLLM VideoMediaIO.
`_is_torchcodec_end_of_stream_error`
`_torchcodec_sample_indices`
`_find_torchcodec_decodable_frame_count`	Find the decodable tail when container metadata overstates frame count.
`_decode_torchcodec_video`	Decode sampled frames, recovering from overstated container metadata.
`_load_video_frames_torchcodec_with_metadata`	Decode video with the repository’s optional TorchCodec dependency.
`register_torchcodec_vllm_video_loader`	Use TorchCodec for raw Nemotron video bytes parsed by vLLM’s HTTP server.
`load_video_frames_with_metadata`	Load sampled RGB frames through the supported TorchCodec backend.

Data#

`VideoSamplingStyle`
`_TORCHCODEC_END_OF_STREAM_ERROR`
`_CACHED_VIDEO_FRAME_MANIFEST_MAGIC`
`_CACHED_VIDEO_FRAME_MANIFEST_MIME`

API#

nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle#: None

nemo_rl.models.generation.vllm.video_utils._TORCHCODEC_END_OF_STREAM_ERROR#: ‘Requested next frame while there are no more frames left to decode.’

nemo_rl.models.generation.vllm.video_utils._CACHED_VIDEO_FRAME_MANIFEST_MAGIC#: b’NEMO_RL_CACHED_VIDEO_FRAMES_V1\n’

nemo_rl.models.generation.vllm.video_utils._CACHED_VIDEO_FRAME_MANIFEST_MIME#: ‘video/x-nemo-rl-cached-frames’

nemo_rl.models.generation.vllm.video_utils._round_video_frame_count( num_frames: int, *, total_frames_in_file: int, max_frames: int, temporal_patch_size: int, ) → int#

nemo_rl.models.generation.vllm.video_utils._timestamp_to_video_frame_index( timestamp_s: float, fps: float, total_frames: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → int#: Convert a timestamp according to the configured sampling contract.

nemo_rl.models.generation.vllm.video_utils._select_video_frame_count( *, requested_num_frames: int, total_frames_in_file: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → int#

nemo_rl.models.generation.vllm.video_utils._compute_video_timestamps( total_duration: float, num_frames: int, total_frames_in_file: int, original_num_frames: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → tuple[int, list[float]]#

nemo_rl.models.generation.vllm.video_utils._build_video_metadata( *, fps: float, total_frames: int, sampled_indices: list[int], backend: str, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → dict[str, Any]#

nemo_rl.models.generation.vllm.video_utils._resolve_cached_video_media_path(value: str) → pathlib.Path#

nemo_rl.models.generation.vllm.video_utils.build_cached_video_frame_metadata( num_frames: int, ) → dict[str, Any]#: Return the shared synthetic timing contract for cached video frames.

nemo_rl.models.generation.vllm.video_utils.build_cached_video_frame_data_url(frame_paths: list[str]) → str#: Build a compact native-video URL backed by lossless cached PNG frames.

nemo_rl.models.generation.vllm.video_utils._load_cached_video_frame_manifest( data: bytes, *, num_frames: int, ) → tuple[numpy.ndarray, dict[str, Any]] | None#: Load an internal cached-frame manifest passed through vLLM VideoMediaIO.

nemo_rl.models.generation.vllm.video_utils._is_torchcodec_end_of_stream_error(exc: RuntimeError) → bool#

nemo_rl.models.generation.vllm.video_utils._torchcodec_sample_indices( *, total_frames: int, fps: float, requested_num_frames: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → list[int]#

nemo_rl.models.generation.vllm.video_utils._find_torchcodec_decodable_frame_count( decoder_factory: collections.abc.Callable[[], Any], declared_total_frames: int, ) → int#: Find the decodable tail when container metadata overstates frame count.

nemo_rl.models.generation.vllm.video_utils._decode_torchcodec_video( source: Any, *, requested_num_frames: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, source_description: str, initial_decoder: Any | None = None, ) → tuple[numpy.ndarray, float, int, list[int]]#: Decode sampled frames, recovering from overstated container metadata.

nemo_rl.models.generation.vllm.video_utils._load_video_frames_torchcodec_with_metadata( video_path: str, *, num_frames: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → tuple[numpy.ndarray, dict[str, Any]]#: Decode video with the repository’s optional TorchCodec dependency.

nemo_rl.models.generation.vllm.video_utils.register_torchcodec_vllm_video_loader( *, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, temporal_patch_size: int, ) → None#

Use TorchCodec for raw Nemotron video bytes parsed by vLLM’s HTTP server.

vLLM’s nemotron_vl loader defaults to OpenCV, while NeMo-RL deliberately does not ship OpenCV or PyAV. Registering this structural VideoLoader implementation under the same extension name keeps vLLM’s media connector contract and makes rollout decoding match policy-logprob preprocessing.

The caller supplies the same validated values materialized into the policy data configuration. Registration is therefore not conditional on process environment state.

nemo_rl.models.generation.vllm.video_utils.load_video_frames_with_metadata( video_path: str, *, num_frames: int, temporal_patch_size: int, sampling_style: nemo_rl.models.generation.vllm.video_utils.VideoSamplingStyle, ) → tuple[numpy.ndarray, dict[str, Any]]#: Load sampled RGB frames through the supported TorchCodec backend.

nemo_rl.models.generation.vllm.video_utils#

Module Contents#

Functions#

Data#

API#

`nemo_rl.models.generation.vllm.video_utils`#