Train llm (bloom, llama, baichuan2-7b, chatglm3-6b) with deepspeed pipeline mode. Faster than zero/zero++/fsdp.

Fine-tuning toolkit for Maya1, a 3B-parameter open-source TTS model by Maya Research. Supports full fine-tuning with SNAC neural codec, natural language voice design, and inline emotion tags. Features YAML-based config, offline preprocessing, and automatic audio sample generation at every checkpoint.