统一的 PyTorch 训练框架，让同一份代码在单卡到上万张 GPU 上都能跑，免去改代码就能做预训练和微调。
工业级零样本语音合成系统，只需几秒参考音频就能克隆音色并跨语种合成，兼顾可控性与推理效率。
让视频扩散模型在普通显卡上也能跑起来，通过帧打包技术大幅降低显存占用，实现实用的长视频生成。
面向大语言模型的高吞吐、低显存推理与服务引擎，通过 PagedAttention 等优化显著提升并发吞吐并降低显存占用，适合部署在线推理服务。
高性能大模型与多模态模型推理服务框架，通过优化注意力机制和调度策略提升吞吐与延迟表现，支持 Llama、DeepSeek、MoE 等主流模型部署。
基于 CTranslate2 重实现的 Whisper 语音转文字推理引擎，通过量化与高效推理大幅提速并降低显存占用，适合本地部署转录。
…tation (segment-everything) converted to Core ML, FP16. FastSAM-s (light, ~23 MB) and FastSAM-x (quality, ~138 MB). YOLOv8-seg architecture; outputs boxes + scores + mask
…etry estimation. Single CoreML mlpackage (~200 MB FP16, 504×504 fixed input). Predicts metric depth, surface normals, and confidence mask from a DINOv2 ViT-B/14 backbone.
…tps://github.com/huggingface/accelerate/pull/3688 FP16/BF16 Training for MPS devices BF16 and FP16 support for MPS devices is finally here. You can now pass mixed precisi
What's Changed fix sage fp16 triton repeat by @ITerydh in https://github.com/feifeibear/long-context-attention/pull/146 bump to 0.6.3.post1 by @feifeibear in https://gith
…0.onnx 326 MB English, full precision kokoro-v1.0.fp16.onnx 164 MB spectral correlation 0.999 against fp32 kokoro-v1.0.int8.onnx 114 MB spectral correlation 0.916 against
…rchu/kokoro-onnx/releases/tag/v0.2.0) kokoro-v1.0.fp16.onnx : (169MB): optimized f16 version from [taylorchu/kokoro-onnx](https://github.com/taylorchu/kokoro-onnx/release