๐ Highlights:
- Whole-Graph & FSDP-Aware Compilation: Breaks local boundaries. Full-graph capture for inference & layer-wise compilation for FSDP training to maximize aggressive cross-op fusion.
- Heuristic Recompute: No more manual checkpointing. Auto-preserves compute-bound ops (e.g., Attention) and recomputes memory-bound ones to slash peak VRAM seamlessly.
- JIT Extreme Offloading: Smart Compute + H2D dual-stream overlap. Keeps high "cost-performance" weights in VRAM and prefetches the rest at the last millisecond to eliminate pipeline bubbles.
- Plug-and-Play: Just 2 decorators (@magi_compile & @magi_register_custom_op) deliver system-level optimizations without complex model refactoring.
๐ช Introducing MagiCompiler: Break the Boundaries of Local Compilation for Large Models.
MagiCompiler is a unified compilation framework advancing torch.compile into a global "Compiler as Manager", conquering the VRAM and compute walls for training & multi-modality inference.
- Repo: https://t.co/efDcyqktf7
๐ Highlights:
- Blackwell Support: Early Blackwell support via FFA_FA4 backend, leveraging HSTU function & R2P optimizations for next-gen hardware.
- Native Group Collective: DeepEP-inspired fused kernels (GroupCast/GroupReduce) to break RDMA bottlenecks and achieve zero-redundancy communication.
- Mask-Agnostic SOTA: Constant high-throughput performance even for irregular, complex masks. No more OOM or performance drops in "hard mode."
- System-Level Synergy: Dispatch Solver + Adaptive Overlap ensures near-linear scalability on H100 & B200 clusters.
๐ช Introducing MagiAttention v1.1.0: Defining the Performance Ceiling for Hopper & Blackwell Architectures.
MagiAttention is a distributed attention engine sharing the bottlenecks of ultra-long context and heterogeneous masks.
- Repo: https://t.co/s8d2X6b0aY
- Blog: https://t.co/FPkNm0yR5X
Thanks for reading! ๐
Come visit us at https://t.co/yn4NJpoMrD to see what else we're building. We are incredibly proud to contribute to the open-source community and will keep pushing the boundaries of AI!
โณ 6/6
๐ช Introducing daVinci-MagiHuman: The Performance-Level Audio-Video Generative Foundation Model
Proudly open-sourced and jointly developed by SII GAIR Lab & https://t.co/yn4NJpoMrD, it sets a new standard for multimodal AI.
โณ 1/6
๐ Performance & Benchmarks: Beating Open-Source SOTA
Great architecture is nothing without results. We benchmarked daVinci-MagiHuman against leading models like LTX-2.3, Ovi 1.1, and MoVA. The results speak for themselves:
๐ฅ Human Blind Evals: Tested on a 100-sample dataset across multiple dimensions, achieving a massive 70.5% overall win rate!
๐ Objective Metrics:
๐ฌ Video (VideoScore2): Outperforms LTX-2.3 in Visual Quality & Text Alignment; superior to Ovi 1.1 across the board.
๐๏ธ Audio (TalkVid-Bench): Crushes the competition with vastly lower Word Error Rates (WER) for crystal-clear, accurate speech.
โณ 5/6
GAGA-1 is coming. The AI actor with voice and visuals as one.
๐ญ Hollywood-level emotion in every frame
๐ฌ One-click audio and video sync
๐ Perfect performance in multiple languages
๐ Here are 9 wild examples:
๐ช Magi-1.1: The Most Powerful Video Extension
- Now available at https://t.co/NRrS8ZxkTw
โพ๏ธ Make your video as long as you want
โญ๏ธ Every move flows perfectly into the next
๐ Everything in your shot stays exactly the same
โณ 1/3
We understand your concern. We cannot confirm if unofficial sites use our API. Importantly, these sites existed before we offered any public API access. We are actively working to take them down through complaints and legal action. For the official, secure service, always use https://t.co/WyfGCGWtFd.
๐ฌ Gaga AI: A tool that animates any photo into a lifelike, talking Character.
- Try it for free at https://t.co/YBiu4T0OY0
Key Features:
๐ฃ๏ธ Pro-Level Realism: Get perfect lip-sync, natural facial expressions, and even hand gestures for your Characters.
๐ญ Emotion Control: Let your script breathe life into Characters, turning a simple โHappyโ into a genuine smile.
๐ช Simple & Intuitive: A clean UI designed for a smooth, easy, and fun creation process.
Whether you're a content creator, educator, or marketerโฆ, Gaga AI helps you produce amazing content fast.
@giffmana You can take a look at Magi-Attention, which supports arbitrary attention masks and delivers performance comparable to Flash Attention 3, along with built-in distributed capabilities.
https://t.co/UCTs07ekNS
๐ช Introducing MagiAttention: A Distributed Attention Towards Linear Scalability for Ultra-Long Context, Heterogeneous Mask Training
- Code: https://t.co/s8d2X6aslq
- Blog: https://t.co/SfF49D1QJd
๐ Highlights:
- Designed natively for distributed training and inference
- Supports flexible, heterogeneous attention masks with Flash-Attention-3 level performance
- Scales linearly with CP size, up to 4M tokens
- Plug-and-play with PyTorch FSDP and Megatron-LM
๐ข Introducing the https://t.co/yn4NJpoeC5 Open Platform: access Magi's top models via Open API.
๐ฌ Get started: https://t.co/JVQ95HoC32
๐ฎ API Platform Console: https://t.co/75ZqoCQol0
๐ฐ Price: 0.5$/5s only.
โณ 1/2