NVlabs / LongLive

2025-09-22 · Source: Github Trending: All languages · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Software Development & Engineering · Depth: Expert, short

Summary

NVlabs released LongLive 2.0 on May 13, 2026, an NVFP4 parallel infrastructure designed for long video generation. This update introduces significant advancements for both training and inference, building upon the real-time interactive capabilities of LongLive 1.0. For training, LongLive 2.0 supports balanced sequence parallel for AR training, multi-shot video training, and utilizes NVFP4 or BF16 precision. Inference features include NVFP4 (W4A4) and NVFP4 KV Cache, multi-shot attention sink, sequence parallel inference, and async decoding, achieving up to 45.7 FPS with the LongLive-2.0-5B-NVFP4-2Step model. The infrastructure also integrates KV cache compression via TriAttention, reducing KV by 50% without quality degradation. Several models are available, including LongLive-2.0-5B and its NVFP4 optimized variants, offering improved performance and VBench scores.

Key takeaway

For machine learning engineers developing or deploying long video generation models, LongLive 2.0 offers a significant performance uplift. You should consider integrating this NVFP4 parallel infrastructure to achieve inference speeds up to 45.7 FPS and reduce KV cache memory by 50% with TriAttention. This enables more efficient training and real-time interactive video generation, directly impacting your project's scalability and responsiveness. Evaluate the NVFP4-optimized models for immediate performance gains.

Key insights

LongLive 2.0 is an NVFP4 parallel infrastructure enabling efficient, high-performance long video generation for both training and inference.

Principles

• Parallelism enhances long video generation.
• Quantization (NVFP4) improves inference speed.
• KV cache compression maintains quality.

Method

LongLive 2.0 employs balanced sequence parallel AR training, multi-shot video processing, and NVFP4/BF16 precision. Inference utilizes NVFP4 (W4A4), NVFP4 KV Cache, multi-shot attention sink, and async decoding.

In practice

• Use "setup_nvfp4_pipeline" for NVFP4 inference.
• Configure "model_quant_use_transformer_engine" for backend.
• Integrate TriAttention for 50% KV reduction.

Topics

• Long Video Generation
• NVFP4 Quantization
• Parallel Inference
• KV Cache Optimization
• Diffusion Models
• Machine Learning Infrastructure

Code references

• NVlabs/LongLive
• WeianMao/triattention
• guandeh17/Self-Forcing
• Wan-Video/Wan2.2

Best for: MLOps Engineer, AI Engineer, Computer Vision Engineer, AI Scientist, Machine Learning Engineer, AI Hardware Engineer

Related on AIssential

• LongLive-2.0: An NVFP4 Parallel Infrastructure for Long Video Generation — LongLive-2.0 leverages NVFP4 and sequence-parallel AR training for efficient long video generation.
• Unlocking High-Performance Inference for DeepSeek with NVFP4 on NVIDIA Blackwell — Blackwell GPUs with NVFP4 quantization and TensorRT LLM dramatically boost MoE model inference performance.
• 3 Ways NVFP4 Accelerates AI Training and Inference — NVFP4 significantly boosts AI training and inference performance and efficiency with minimal accuracy loss.
• KV Caching in LLMs: A Guide for Developers — KV caching reuses attention's key and value projections to accelerate autoregressive LLM inference by avoiding redundant computation.
• Lambda's MLPerf Inference v6.0: hardware leap, software maturity, research breakthrough — Hardware, software, and routing optimizations significantly boost AI inference performance and reduce latency.
• WorldKV: Efficient World Memory with World Retrieval and Compression — WorldKV uses retrieval and compression to maintain persistent world memory in video diffusion models efficiently.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Github Trending: All languages.