THIRD-PARTY SOFTWARE NOTICES AND INFORMATION This product includes third-party software components. The notices below are provided in compliance with the terms of the respective licenses. NVIDIA-authored source code and published documentation in this repository are licensed under Apache-2.0. See the LICENSE file in the repository root. Version numbers below reflect pinned versions where this repository declares them (for example in avlm/data_prep_example/tennis/pyproject.toml). Other packages are supplied by the recommended NGC containers or resolved at pip install time; verify the license and version of your installed build. ================================================================================ Data preparation (avlm/data_prep_example/tennis/) ================================================================================ Used for optional LLM-assisted MCQ generation and the offline data-prep pipeline. Install via uv/pip in that directory (see README). openai 2.45.0 Apache-2.0 python-dotenv 1.2.2 BSD-3-Clause tqdm (unpinned) MPL-2.0 AND MIT (dual-licensed) Sources: https://github.com/openai/openai-python https://github.com/theskumar/python-dotenv https://github.com/tqdm/tqdm ================================================================================ Evaluation (LLM judge and related scripts) ================================================================================ openai (see above) Apache-2.0 PyYAML (container/pip) MIT tqdm (see above) MPL-2.0 AND MIT ================================================================================ Video and audio decoding — VLM gap-fill (runtime pip install) ================================================================================ The git release does not bundle codec binaries. AutoModel and Megatron-Bridge launch scripts may download and install the packages below via pip into a user-configurable cache directory (for example ${CACHE_DIR}/automodel_container_pkgs or ${CACHE_DIR}/megatron_bridge_container_pkgs) when they are missing from the NeMo container. Set SKIP_VLM_TRAINING_DEPS=1 or SKIP_VLM_INFERENCE_DEPS=1 to disable automatic installation. When no system ffmpeg executable is available, scripts may expose the FFmpeg binary shipped inside imageio-ffmpeg on PATH. decord (pip) Apache-2.0 imageio-ffmpeg (pip) BSD-2-Clause librosa (pip) ISC audioread (pip) MIT pooch (pip) BSD-3-Clause lazy-loader (pip) BSD-3-Clause soundfile (optional pip) BSD-3-Clause Sources: https://github.com/dmlc/decord https://github.com/imageio/imageio-ffmpeg https://github.com/librosa/librosa https://github.com/beetbox/audioread https://github.com/fatiando/pooch https://github.com/scientific-python/lazy_loader https://github.com/bastibe/python-soundfile decord depends on FFmpeg/libav at runtime for video decoding. -------------------------------------------------------------------------------- FFmpeg (via imageio-ffmpeg or system install) -------------------------------------------------------------------------------- Purpose: Audio demux/decode from video containers; clip extraction in data prep (build_sample_from_raw.py --extract-clips requires ffmpeg on PATH). License: FFmpeg is typically licensed under LGPL-2.1-or-later; some builds include GPL-2.0-or-later components. License terms depend on the specific binary/build. Source: https://ffmpeg.org/ Note: imageio-ffmpeg downloads and redistributes a prebuilt FFmpeg binary. Review the license terms of the FFmpeg build used by your installed imageio-ffmpeg version before redistribution or commercial use. ================================================================================ Optional components ================================================================================ wandb (optional) MIT nvidia-ml-py (optional) BSD Sources: https://github.com/wandb/wandb https://github.com/gpuopenanalytics/nvidia-ml-py ================================================================================ deep_ep (vendored wheels — wheels/deepep/) ================================================================================ Pre-built deep_ep wheels are checked into wheels/deepep/ and redistributed as compiled binaries (not on PyPI). Installed automatically on supported hardware when recipes use dispatcher: deepep. Build and install details: wheels/deepep/DEEPEP.MD. NVIDIA install/build scripts under wheels/deepep/ are Apache-2.0. Source: https://github.com/deepseek-ai/DeepEP MIT License Copyright (c) 2025 DeepSeek Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. ================================================================================ NGC container baseline (not bundled in this repository) ================================================================================ Training and inference are designed to run in NVIDIA NeMo containers. These images supply the core Python stack and many dependencies below. Container images and their contents are subject to their own license terms on NGC: nvcr.io/nvidia/nemo-automodel:26.06.00 nvcr.io/nvidia/nemo:26.06.00 Representative components preinstalled or pulled in those environments: torch BSD-3-Clause (PyTorch) transformers Apache-2.0 nemo-automodel Apache-2.0 safetensors Apache-2.0 triton MIT megatron-core Apache-2.0 megatron-bridge Apache-2.0 transformer-engine Apache-2.0 mamba-ssm Apache-2.0 causal-conv1d BSD-3-Clause numpy BSD-3-Clause Pillow HPND PyYAML MIT Review NGC image documentation and upstream project licenses for the complete dependency graph of your chosen container version. ================================================================================ FFmpeg / LGPL notice ================================================================================ If you use FFmpeg (whether installed on your system or obtained via imageio-ffmpeg), you are responsible for complying with FFmpeg's license terms. FFmpeg may be built with optional components under LGPL or GPL. Consult https://ffmpeg.org/legal.html and the documentation for your specific FFmpeg build to determine your obligations.