[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-orbitquant-rpbh-data-independent":3,"news-related-2b004bb4-8409-40b7-8034-154f608279ec":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},"2b004bb4-8409-40b7-8034-154f608279ec","OrbitQuant：用 RPBH 旋转归一化把 DiT 量化做成「数据无关」，FLUX\u002FWan\u002FCogVideoX 同码本同 SOTA","扩散 Transformer (DiT) 是当下 FLUX、Wan、CogVideoX 这一类图像\u002F视频生成模型的标配架构,跑 SOTA 是真的,推理贵也是真的。多步采样叠上参数规模,DiT 的部署成本始终压在「实时互动」和「端侧」的门外。后训练量化 (PTQ) 被普遍视为最优解——但 DiT 的激活分布在 timestep、prompt、guidance 分支之间到处漂,旧办法每换一个 checkpoint、每跨一个模态,都得重新过一遍校准数据,工程上几乎不可用。\n\nOrbitQuant (arXiv:2607.02461) 直接绕开范围估计:它在归一化、旋转过的基底上做量化。核心是一层 randomized permuted block-Hadamard (RPBH) 旋转,能把任意输入的坐标「压」到同一个固定、已知的边缘分布附近——也就是说,一个 Lloyd-Max 码本就能覆盖同一输入维度下的所有 timestep、prompt、layer;weight 行同样离线吸收旋转,运行时只剩 activation 端一次前向旋转;同一套配方从图像迁移到视频无需重新调参,「一次量化,跨模态部署」由此有了工程基础。\n\n实测跨 FLUX.1、Z-Image-Turbo、Wan 2.1、CogVideoX,在多个低比特档位都刷新 PTQ SOTA,最狠的是把图像 DiT 的 PTQ 推到了 W2A4 并保留可用生成质量。对部署而言,OrbitQuant 把「量化模型 = 重新校准」这个潜规则一脚踢开:训练侧量化不再绑死 checkpoint,运行时推理侧也摆脱了逐层调参的痛苦。DiT 想真正跑进端侧,这是少有的「一次量化,通用部署」方向。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.02461","7437aeb9-930c-4866-a2e9-48003c1a792b",[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},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"b49648f9-963e-4082-8684-3d085b7358fe","quantization",{"id":21,"name":22,"slug":22,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"f05ff212-ccac-46d9-a28a-7abd1e4b9cee","en","OrbitQuant: data-free DiT quantization via RPBH rotation","Diffusion Transformer (DiT) is the standard architecture for FLUX, Wan, CogVideoX and other image\u002Fvideo generation models of the moment, hitting SOTA for real, but inference is also expensive. Multi-step sampling stacked on top of parameter scale, DiT's deployment cost is always pressing at the door of \"real-time interaction\" and \"on-device\". Post-training quantization (PTQ) is generally seen as the optimal solution — but DiT's activation distribution drifts between timesteps, prompts, and guidance branches, so old methods need to re-run a round of calibration data for every checkpoint swap and every mode crossover, which is practically unusable. OrbitQuant (arXiv:2607.02461) sidesteps range estimation directly: it quantizes on a normalized, rotated basis. The core is a layer of randomized permuted block-Hadamard (RPBH) rotation, which can \"press\" any input's coordinates close to the same fixed, known marginal distribution — that is, a single Lloyd-Max codebook can cover all timesteps, prompts, and layers under the same input dimension; weight rows absorb the rotation offline the same way, with only one forward rotation on the activation side at runtime; the same set of recipes transfers from image to video without re-tuning, \"one quantization, cross-modality deployment\" thus has an engineering basis. Tested across FLUX.1, Z-Image-Turbo, Wan 2.1, CogVideoX, it refreshes PTQ SOTA across multiple low-bit tiers, with the most aggressive pushing image DiT's PTQ to W2A4 and preserving usable generation quality. For deployment, OrbitQuant kicks away the unwritten rule of \"quantized model = recalibration\": training-side quantization is no longer locked to the checkpoint, and runtime inference-side also escapes the pain of layer-by-layer tuning. For DiT to truly run on-device, this is one of the few \"one-time quantization, universal deployment\" directions.","orbitquant-rpbh-data-independent","2026-07-05T04:10:00Z","2026-07-05T04:11:17.291668Z","2026-08-19T02:08:40.142862Z",true,"agent",96,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"06110002-82fb-421e-9daf-5ab73ead7f27","FourTune：把扩散模型后训练压进 4-bit，W4A4G4 让 FLUX.1-dev 12B 内存砍半、吞吐翻倍","fourtune-4bit-flux-12b","2026-07-09T00:01:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"4b0d7ce2-ac3b-4d3c-a18b-83971308b1e2","从U-Net到DiT：扩散模型架构演进背后的「硬件彩票」","unet-dit-hardware-lottery-diffusion","2026-05-17T14:06:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"2bd17cfe-c17d-43d4-82ad-3688fa69fd3a","NVIDIA联手Black Forest Labs：FP8量化让FLUX.2进入RTX消费级显卡时代","nvidia-bfl-flux-2-fp8-rtx-40pct","2026-05-13T22:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"48e1c261-a40a-4c71-9cba-450a459e6ad3","4-bit 模型反超全精度:QAH 把量化从性能税变成第二次蒸馏","quantization-aware-healing-hypernova-60b","2026-08-25T17:20:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"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":64,"title":65,"news_slug":66,"published_at":67},"4147f71b-eaa7-4d39-91cf-c2c572105e7f","FlashPrefill V2:128K 长文本 prefill 提速 47 倍,块稀疏注意力走进生产框架","flashprefill-v2-block-sparse-prefill","2026-08-21T19:10:00+00:00"]