CLIPA

petals 开源

在家用 BitTorrent 方式跑大模型，推理微调快 10 倍

Run LLMs at home, BitTorrent-style. Fine-tuning and inference up to 10x faster than ···

merge-models 开源

按比例融合两个扩散模型，快速生成新风格模型

Merges two latent diffusion models at a user-defined ratio

IPR 开源

逐步纠错，让LLM智能体每一步都走对

Watch Every Step! LLM Agent Learning via Iterative Step-level Process Refinement (E···

Soup 开源

一个YAML搞定大模型微调，4GB显卡也能训8B模型

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GP···

DPO-Shift 开源

修正 DPO 分布偏移，让模型对齐更稳更准

DPO-Shift: Shifting the Distribution of Direct Preference Optimization

model-optimization 开源

几行代码压缩模型，让大模型跑进手机和嵌入式设备

A toolkit to optimize ML models for deployment for Keras and TensorFlow, including q···

DDSM 开源

让扩散模型训练更快更省，按步骤智能分配学习资源

Denoising Diffusion Step-aware Models (ICLR2024)

EasyR1 开源

多模态强化学习训练，高效扩展不卡壳

EasyR1: An Efficient, Scalable, Multi-Modality RL Training Framework based on veRL

UniRL 开源

一套框架搞定多种数据训练，强化学习不再各搞各的

UniRL is a Framework for Unified Multimodal Model Reinforcement Learning