只需几行代码就能把 PyTorch 模型跑在任意设备或分布式配置上，自动处理混合精度（含 fp8）和 FSDP、DeepSpeed 配置，省去手写分布式训练样板代码的麻烦。
从零实现的大模型蒸馏与推理服务引擎，包含自定义 Triton/CUDA 算子、FSDP 蒸馏、分页 KV 连续批处理、投机解码和 Rust 网关，适合研究推理加速与可解释性。
基于 PyTorch 的投机解码模型训练库，支持 EAGLE3、FSDP 等方案，帮你在 vLLM 等推理服务里加速大模型生成。
… API for sampling and training, sharded EMA under FSDP2, and a separate precision for method and sampling state. - Tooling: ComfyUI custom nodes, a VLM judge, FlashDreams
v1.15.0: FSDP2 activation memory, dtensor improvements FSDP2 A large batch of FSDP2 work this release: two fixes that cut activation memory at long sequence lengths, tied
…) New features - Extend data parallel support for FSDP/FSDP2 (PR 537) - Support trackio and wandb (PR 543) - Add accumulation mode with stack option (PR 536) Improvements
…ymFLUX.2 variant models. Changed - Change default FSDP buffer dtype to float32 . - Introduce various performance optimizations: - Use fused foreach implementation for EMA
…cture to lakonlab.models.architectures . - Update FSDP/DDP wrappers, checkpoint loading/saving, runner utilities, and S3 retry/refresh behavior. Fixed - Fix multiple bugs
FSDP2 Improvements This release brings a large batch of FSDP2 fixes and quality-of-life improvements: correct dtype handling on load, sharding of embeddings/norms, QLoRA