Long-context capability is critical for multi-modal foundation models. We introduce LongVILA, a full-stack solution for long-context vision-language models, including system, model training, and dataset development. On the system side, we introduce the first long-context Multi-Modal Sequence Parallelism (MM-SP) system that enables long training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. MM-SP is 2.1x - 5.7x faster than ring sequence parallelism and 1.1x - 1.4x faster than Megatron context parallelism + tensor parallelism in text-only settings. Moreover, it seamlessly integrates with Hugging Face Transformers. For model training, we propose a five-stage pipeline comprising alignment, pre-training, short supervised fine-tuning, context extension, and long supervised fine-tuning. On datasets, we construct large-scale visual language pre-training datasets and long video instruction-following datasets to support our multi-stage training process. LongVILA extends the number of frames of VILA from 8 to 1024, and improves the long video captioning score from 2.00 to 3.26 (1.6x), achieving 99.5% accuracy in 1400-frames video (274k context length) needle-in-a-haystack. LongVILA-8B demonstrates consistent accuracy improvements on long videos in the VideoMME benchmark as the number of frames increases.
翻译:暂无翻译