[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-huggingface-diffusers-h100":3,"news-related-75d2f385-00ef-4074-80bf-ee47ba05a4a4":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"75d2f385-00ef-4074-80bf-ee47ba05a4a4","NVIDIA × HF：Diffusers 微调上 H100，Wan 2.2\u002FFLUX.2 打通","NVIDIA 与 Hugging Face 7 月 17 日联合发布 NeMo Automodel 与 Diffusers 的深度集成,把扩散模型的训练\u002F微调从「每换一个模型就要写一套脚本」拉进「写一次、跑遍整个 Hub」的状态。\n\n核心改动是把 NeMo Automodel 的训练栈从原本的 LLM\u002FMoE 场景扩展到 flow-matching 扩散模型。底座是 DTensor + PyTorch 原生,所有并行策略——FSDP2、Tensor Parallel、Expert Parallel、Context Parallel、Pipeline Parallel——都通过 YAML 切换,不再需要改模型代码。模型类直接复用 Diffusers 的 WanTransformer3DModel、FLUXPipeline,训练完的 checkpoint 能立刻跑回 DiffusionPipeline 推理,完全没有格式转换的中间步骤。\n\n首批官方 recipe 覆盖了开源扩散圈的主力:FLUX.1-dev(12B)与 FLUX.2-dev(32B)的文生图、Wan 2.1 1.3B\u002F14B 与 Wan 2.2 A14B(MoE)的文生视频、HunyuanVideo 1.5(13B)、Qwen-Image(20B MMDiT),全部同时支持 Full FT 与 LoRA。配合 latent 缓存 + 多分辨率 bucketing,真正做到了「数据集预编码一次,后面全是模型与并行策略的旋钮」。\n\n实测数据来自 8×H100 80GB 集群:FLUX.1-dev 全量微调 35.51 imgs\u002Fs、LoRA r64 53.73 imgs\u002Fs;Wan 2.1 14B 全量 2.107 clips\u002Fs;Wan 2.2 A14B 高噪分支 1.73 clips\u002Fs;FLUX.2-dev 32B 也已被列入路线图。单卡显存峰值多压在 60GiB 以内,意味着大部分条目单机 8 卡就能开训。整个栈 Apache 2.0 开源,Pythonic recipe API 也已在路上,下一步是把 YAML 配置和编程式接口并列起来。\n\n这套工具的真正价值不在性能数字,而是把「微调」从一项需要为每个模型写一堆胶水代码的工程活,变成可配置的研究基础设施。对 LoRA 创作者和企业定制模型都是直接利好——新模型一上 Hub,几行 YAML 就能开训。NVIDIA 在 LLM 与 Diffusion 两端同时拿下「训练框架」位置,等于把生态护城河往上游推了一格。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Fnvidia\u002Fscale-diffusers-finetuning-nemo-automodel","474eef8c-e0c3-46cf-adee-c089558220f9",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"74cf9ffe-fccb-4c4a-840c-c48ef2fdd872","en","NVIDIA x HF: Diffusers fine-tuning on H100, Wan 2.2 wired","NVIDIA and Hugging Face jointly released the deep integration of NeMo Automodel and Diffusers on July 17, pushing diffusion-model training \u002F fine-tuning from \"write a new script for each new model\" into a \"write once, run across the whole Hub\" state. The core change is extending NeMo Automodel's training stack from its original LLM\u002FMoE scenarios to flow-matching diffusion models. The base is DTensor + native PyTorch, with all parallel strategies — FSDP2, Tensor Parallel, Expert Parallel, Context Parallel, Pipeline Parallel — switched via YAML, no model code changes required. The model class directly reuses Diffusers' WanTransformer3DModel and FLUXPipeline; the trained checkpoint can immediately run back through DiffusionPipeline for inference, with no format-conversion middle steps. The first batch of official recipes covers the open-source diffusion mainstream: FLUX.1-dev (12B) and FLUX.2-dev (32B) text-to-image, Wan 2.1 1.3B\u002F14B and Wan 2.2 A14B (MoE) text-to-video, HunyuanVideo 1.5 (13B), Qwen-Image (20B MMDiT), all simultaneously supporting Full FT and LoRA. Combined with latent caching + multi-resolution bucketing, the dataset is pre-encoded once and the rest are all model and parallel-strategy knobs. Measurement data comes from an 8×H100 80GB cluster: FLUX.1-dev full fine-tuning 35.51 imgs\u002Fs, LoRA r64 53.73 imgs\u002Fs; Wan 2.1 14B full 2.107 clips\u002Fs; Wan 2.2 A14B high-noise branch 1.73 clips\u002Fs; FLUX.2-dev 32B is also on the roadmap. Single-card VRAM peak is mostly under 60GiB, meaning most entries can train on a single 8-card machine. The entire stack is Apache 2.0 open-source, and the Pythonic recipe API is on its way, with the next step putting YAML configuration and programmatic interface side by side. The real value of this set of tools isn't the performance numbers, but turning \"fine-tuning\" from an engineering task that requires writing a bunch of glue code for each model, into configurable research infrastructure. For LoRA creators and enterprise-customized models it's a direct boon — once a new model hits the Hub, a few lines of YAML lets you start training. NVIDIA takes the \"training framework\" position on both the LLM and Diffusion sides, pushing the ecosystem moat one notch upstream.","nvidia-huggingface-diffusers-h100","2026-07-17T14:00:00Z","2026-07-18T22:11:34.902167Z","2026-08-19T02:08:40.142862Z",true,"agent",80,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"18d2aa73-7244-4b10-b611-46475e17327e","ForgeWM开源:一步去噪72FPS的可玩世界模型,8张卡复现全流程","forgewm-few-step-playable-world-model","2026-08-24T21:10:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"2874a2e5-beae-4627-8f6f-a34cf2cc8d7a","一段随手拍视频直出4D人体:4DAnyone用RCP+TCR破解多视角一致性,代码权重全开源","4danyone-monocular-video-4d-human","2026-08-20T17:59:53+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"5612d186-46ee-4509-9a93-94045ba004ae","LTX-2.5 开放权重视频模型:4K 反而在 Fast 端点,EXR 色彩管线也焊进去了","ltx-2-5-open-weights-video","2026-08-18T15:20:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"6f9e9f94-9dcc-4c6c-b254-6c5d0fe8ed37","京东开源 JoyAI-Video-Edit:16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-realtime-diffusion","2026-08-10T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"bcdc10bc-2f08-4c39-8ffa-e7e34041c112","京东开源 JoyAI-Video-Edit:用 16B 多模态扩散 Transformer 把视频编辑推进「边播边改」实时流时代","jd-joyai-video-edit-real-time-streaming","2026-08-05T03:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"a818c807-2131-4950-8f51-62847a57db41","VideoRAE 把 frozen 视频基础模型改造成生成器 latent:UCF-101 gFVD 40\u002F93,收敛提速 5×","videorae-frozen-video-generator","2026-07-20T04:15:00+00:00"]