Sana:基于扩散模型的高分辨率图像视频生成项目

Accelerated video-diffusion inference

📖 Docs  |  Pipelines  |  Techniques  |  Install

Agent-native workflow · Full-stack acceleration techniques · A wide range of video generation models

---

Sol-Engine is an efficiency-oriented inference codebase for high-resolution video diffusion, built on SGLang's multimodal_gen runtime. It features an agent-native inference workflow and reduces three production models into one unambiguous acceleration line. This is powered by a full-stack solution composed of five reusable acceleration techniques, delivering a 2× to 3× end-to-end speedup across the three models. We are actively continuing development to support a wider range of models.

📰 News

• [2026/09/08] 🔥 Sol-H3 [Code | Project Page] — production MiniMax-H3 inference for T2V, I2V, and Ref2VA with synchronized audio. On 8× NVIDIA B300 at 1344×768, warm BF16-compute T2V latency is 1.653 / 3.732 / 6.612 s for 5 / 10 / 15-second outputs; optional fused MXFP8 compute reaches 1.472 s for the 5-second case.
• [2026/08/22] 🔥 MiniMax-H3 Super Acceleration [Code | Blog] — combines a 4-step MiniMax-H3 draft with a 3-step LTX-2.5 refinement pass, reaching 22.2× end-to-end speedup for 5-second 768p video and 27.7× for 10-second video on one NVIDIA GB200.
• [2026/08/17] 🔥 MiniMax-H3 on GeForce RTX 4090 [Code] — reaches 4.44× end-to-end on RTX 4090. Sol-Attn now includes an optimized SM89 CuTe DSL kernel for RTX 4090.
• [2026/08/13] 🔥 LTX-2.5 across B200, GeForce RTX 5090, and DGX Spark [Code | Blog] — reaches up to 4.68× multi-step pipeline speedup and 1.90× distilled pipeline speedup.
• [2026/08/09] 🔥 MiniMax-H3 on H100 and A100 [Code | Blog] — reaches 3.56× end-to-end on 4×H100 and 3.55× on 4×A100.
• [2026/08/06] 🔥 MiniMax-H3 on GB10 and RTX 5090 [Code | Blog] — reaches 3.92× end-to-end on DGX Spark (GB10) and 4.52× on RTX 5090.
• [2026/08/03] 🔥 MiniMax-H3 on GB200 [Code | Blog] — reaches 3.95× end-to-end on 8×GB200.
• [2026/07/28] 🔥 Sol-Attn [Paper | Code] — sparse video attention lands with released SM89/SM90/SM100/SM120 kernels for HunyuanVideo-13B (~5.03×) and Wan2.1-T2V-14B (~3.48×) end-to-end.
• [2026/07/15] 🔥 Three new models — Wan2.2 TI2V-5B ~2.89×, Wan2.2-A14B ~2.17×, and LingBot-Video ~2.60× end-to-end.
• [2026/07/13] ⚙️ Agent workflow update — refreshed the agent-native optimization workflow (a master orchestrator driving per-technique executor sub-agents with automatic quality gates). See the agent-workflow page.
• [2026/06] 📖 Docs release — full documentation site live: 3 pipeline designs + 5 acceleration techniques.
• [2026/06] 🔥 SANA-Video — EasyCache + kernel fusion + torch.compile → ~2.77× end-to-end.
• [2026/06] 🔥 LTX-2.3 — KWL fusion + cache + PISA + NVFP4 + token-prune → ~2.38× end-to-end.
• [2026/06] 🔥 Cosmos3-Super — TeaCache + step-selective NVFP4 → ~2.27× end-to-end (4×GB200).

⚡ Models & speedups

_{Speedup is warm end-to-end latency versus the unoptimized same-topology baseline, with model, prompt/seed, resolution, step count, batch size, and GPU topology held fixed. SANA 480p (832×480, 81f, 50 steps); Cosmos3 1280×720, 189f, 35 steps; LTX-2.3 1088×1920, 241f. LTX-2.5 uses the 4×B200 BF16 multi-step pipeline at 1536×1024 for 5 s (30-step Stage 1 + 2-step Stage 2), comparing the official sequence-parallel baseline (39.23 s) with full-opt (8.38 s). Wan-5B 704×1280, 121f, 50 steps (1 GPU); Wan-14B 720×1280, 81f, 40 steps (1 GPU); LingBot base 480×832→refiner 1088×1920, 121f (4 GPU CP4); HunyuanVideo 1280×720, 129f, 50 steps (1 GPU); Wan2.1-14B 720×1280, 81f, 50 steps (1 GPU). MiniMax-H3 uses 8×GB200 at 1344×768, 124 frames, 50 steps, and batch 1, comparing the Diffusers baseline (27.205 s) with full-opt (6.883 s).}

🧩 The five acceleration methods

Video diffusion inference exposes redundancy at three complementary levels. At the algorithm level, adjacent denoising steps run structurally similar computations over slowly changing latents, so cache can reuse or skip step outputs. At the model level, long spatiotemporal sequences contain redundant tokens and attention interactions, motivating sparse attention and token pruning. At the kernel level, DiT blocks repeatedly launch memory-bound work around GEMMs, layout movement, normalization, activation, and precision conversion, which quantization and fusion reduce. Sol-Engine composes the five methods across these levels.

🌀 Sol-Attention

Sol-Attn is our sparse attention technique for video DiTs: video-token self-attention runs through a sparse kernel that computes only the most relevant key blocks, while using light weight compensating operation to maintain visual quality. It plugs into a model runtime through env-gated hooks and powers the optimization stacks of HunyuanVideo-13B, Wan2.1-T2V-14B and MiniMax-H3. The HunyuanVideo and Wan2.1 stacks reach ~5.03× and ~3.48× end-to-end speedup, respectively.

Backend selection is automatic: RTX 4090 (SM89), H100 (SM90), B200 (SM100), and RTX 5090 (SM120) use their optimized CuTe DSL kernels when CuTe is installed. Older supported GPUs such as A100 (SM80), or installations without CuTe DSL, use the portable Triton implementation. Apple Silicon uses the tiled Metal backend when torch.mps.compile_shader is available. Model call sites do not need architecture-specific changes. See the Sol-Attn usage guide for requirements and API details.

🚀 Quick start (agent-native)

Sol-Engine is installed and launched the agent-native way. Rather than hand-running the setup steps, you hand a coding agent — OpenAI Codex or Claude Code — a single goal and let it create the environment, fetch the weights, and run all three models in both baseline and fullopt settings, troubleshooting and adapting the scripts to your machine as it goes.

From the repo root, give the agent this goal:

/goal Execute the inference code for the three models using both baseline and full-opt
settings with the following requirements. Refer to AGENTS.md for the environment creation,
model download, and inference guides. For the environment, you need to create a new
environment. For model weights, you are allowed to reuse existing weights if they are
locally available; otherwise, you need to download them. Regarding adaptability, be aware
that the provided guides for environment creation, download scripts, and inference may
contain system incompatibilities, so you are expected to troubleshoot and adapt them to
your specific machine.

📖 Getting started

• 📚 Full documentation — a comprehensive guidebook to the whole project: pipeline designs, acceleration techniques, setup, and model references in one place
• 🛠️ Installation — conda env, editable install, CUDA-JIT fixups, and the HF model repos + download helpers
• 🎬 Optimized pipelines — SANA-Video · Cosmos3-Super · LTX-2.3 · HunyuanVideo-13B · Wan2.1-T2V-14B
• ⚙️ Acceleration techniques — Cache · Quantization · Kernel fusion · Sparse attention · Token pruning

✅ To-do

• SANA-Video acceleration line — EasyCache + fusion + compile
• Cosmos3-Super acceleration line — TeaCache + step-selective NVFP4
• LTX-2.3 acceleration line — KWL fusion + cache + PISA + NVFP4 + token-prune
• Wan2.2 TI2V-5B, Wan2.2-A14B, and LingBot-Video acceleration lines — kernel fusion + cache + PISA/compile
• HunyuanVideo-13B and Wan2.1-T2V-14B acceleration lines — kernel fusion + cache + Sol-Attn
• Sol-Attn sparse-attention release — optimized SM89/SM90/SM100/SM120 kernels + portable Triton backend
• MiniMax-H3 across GB200, H100, A100, GB10, RTX 4090, and RTX 5090 — context parallel + kernel fusion + Sol-Attn + FirstBlockCache/TeaCache
• LTX-2.5 across B200, GeForce RTX 5090, and DGX Spark — parallel optimization + FBCache + kernel fusion + Sol-Attn
• More backends for each acceleration method
• Agent-native workflow without human-in-the-loop

🙏 Acknowledgements

Built on SGLang and 🤗 Diffusers. Pipelines wrap SANA-Video, NVIDIA Cosmos, and Lightricks LTX-Video. Acceleration methods draw on TeaCache, EasyCache, SVDQuant/Nunchaku, FlashAttention, TransformerEngine, and the sparse-attention / token-reduction literature surveyed in the docs.

📌 Citation

@misc{li2026solvideoinferenceengine,
  title         = {Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation},
  author        = {Yitong Li and Junsong Chen and Haopeng Li and Haozhe Liu and Jincheng Yu and Ligeng Zhu and Ping Luo and Song Han and Enze Xie},
  year          = {2026},
  eprint        = {2606.23743},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2606.23743},
  url           = {https://arxiv.org/abs/2606.23743},
}

📄 License

Code in this repository is released under the Apache-2.0 license. The paper is available on arXiv under the arXiv.org perpetual, non-exclusive distribution license. Model weights follow their respective upstream licenses (SANA-Video, NVIDIA Cosmos, Lightricks LTX) — see each model card.