cosmos3-quantization
Quantize Cosmos3 generator checkpoints to static-scale FP8 with the cosmos_framework.quantization recipe (NVIDIA TensorRT Model Optimizer): the Edge, Nano and Super video models and the Super Text-to-Image / Image-to-Video base and 4-step distilled students, exported as drop-in Diffusers-layout checkpoints with calibrated scales and an hf_quant_config.json that load on vLLM-Omni. Use when the user asks "quantize Cosmos3 to FP8", "make / calibrate / export my own FP8 Cosmos3 checkpoint", "run quantize_fp8_checkpoint", "DEMO=1 vs DEMO=0 calibration", or "what does the FP8 export contain". Do not use it to run inference with a published FP8 checkpoint (nvidia/Cosmos3-Edge-FP8, -Nano-FP8, -Super-FP8) — that is the FP8 route of cosmos3-inference-generate-media and of the Reasoner / Action / Transfer cookbook notebooks (their skills follow in later rows) — nor for the framework's load-time --quantization-method flags.
来源信息
- 仓库
- NVIDIA/cosmos-framework
- 最近来源活动
- 2026年11月11日 05:17
- 检测到的 SKILL.md 语言
- 英语
- 星标
- 648
- Forks
- 214
安装方式
默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。
检查来源文件
决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。
正在显示 SKILL.md
| name | cosmos3-quantization |
|---|---|
| description | Quantize Cosmos3 generator checkpoints to static-scale FP8 with the cosmos_framework.quantization recipe (NVIDIA TensorRT Model Optimizer): the Edge, Nano and Super video models and the Super Text-to-Image / Image-to-Video base and 4-step distilled students, exported as drop-in Diffusers-layout checkpoints with calibrated scales and an hf_quant_config.json that load on vLLM-Omni. Use when the user asks "quantize Cosmos3 to FP8", "make / calibrate / export my own FP8 Cosmos3 checkpoint", "run quantize_fp8_checkpoint", "DEMO=1 vs DEMO=0 calibration", or "what does the FP8 export contain". Do not use it to run inference with a published FP8 checkpoint (nvidia/Cosmos3-Edge-FP8, -Nano-FP8, -Super-FP8) — that is the FP8 route of cosmos3-inference-generate-media and of the Reasoner / Action / Transfer cookbook notebooks (their skills follow in later rows) — nor for the framework's load-time --quantization-method flags. |