Skip to content

Releases: TeamAudiyo/Audiyo

0.3.4

Choose a tag to compare

@TheLimeDev TheLimeDev released this 18 Sep 18:27

Small fix release.

Fixed in 0.3.4: quantized music loads use a quantization config on current transformers instead of a bare flag that newer versions reject.

Run pip install -U audiyo to get it.

0.3.3

Choose a tag to compare

@TheLimeDev TheLimeDev released this 17 Sep 23:55

Small fix release.

Fixed in 0.3.3: the music loader no longer sends quantization flags to the pipeline itself, which rejected them. Quantization now applies to the language model on its own after loading. Also includes the GGUF test script under scripts/.

Run pip install -U audiyo to get it.

0.3.2

Choose a tag to compare

@TheLimeDev TheLimeDev released this 17 Sep 23:12

Small music fix release.

New in 0.3.2: GGUF quants now load from TeamAudiyo/Minimax-Music3-GGUF everywhere, the loader fills a missing tokenizer, language model, or depth decoder from the base repo and names what it filled, and each backend only gets the call arguments it understands so there are no more stray argument warnings. Songs run up to six minutes.

Run pip install -U audiyo to get it.

0.3.1

Choose a tag to compare

@TheLimeDev TheLimeDev released this 17 Sep 01:38

Easier to use day to day.

New in 0.3.1: generate variations and small batches in one go, tidy endings with fades plus normalize plus a soft limiter and silence trim, check levels with quality and compare, keep listening notes with sheets, reuse settings with config files, try the local web demo, and check size with estimate before loading.

PyPI is updated too. See README and docs/generation.md, docs/quality.md, docs/ui.md to get going.

0.2.1

Choose a tag to compare

@TheLimeDev TheLimeDev released this 11 Sep 22:09

MiniMax-Music3 out of the box.

New in 0.2.1: diffusers pinned to >=0.40 (ships MiniMaxMusic3ModularPipeline, verified); full Music3 checkpoint path loads against MiniMaxAI/MiniMax-Music3 modular components; GGUF DiT from TeamAudiyo/MM3-GGUF (default q4_k_m 1.49GB, peak under 4.5GB); README documents Music3 generation and dual-LoRA.

Full song generation needs a CUDA GPU. Model weights are not bundled.

0.2.0

Choose a tag to compare

@TheLimeDev TheLimeDev released this 11 Sep 21:41

Native GGUF DiT plus dual-component LoRA.

New in 0.2.0: GGUF DiT quants from TeamAudiyo/MM3-GGUF (default q4_k_m 1.49GB, peak under 4.5GB); quant/gguf_file/gguf_path in load(); lyrics in from_pretrained/generate(); runnable dual LoRA finetune for llm, transformer, both with frozen bases; gguf>=0.10.0 dep.

Model weights are not bundled.

0.1.1

Choose a tag to compare

@TheLimeDev TheLimeDev released this 11 Sep 19:49

Out-of-the-box MiniMax-Music3.

New in 0.1.1: real load() via MiniMaxMusic3ModularPipeline (needs diffusers newer than pinned 0.39.0); sequential stage offloading so the 8B LLM, 2.4B DiT, and vocoder swap in and out of GPU RAM; VRAM-aware auto-quant (12GB cards get int4 LLM, 16GB cards get int8, opt-in via bitsandbytes); scrubbed auth/OOM errors.

Model weights are not bundled.

0.1.0

Choose a tag to compare

@TheLimeDev TheLimeDev released this 11 Sep 18:40

Second release.

New in 0.1.0: backend dispatcher with stable-audio, testkit, and minimax-music backends; Minimax-Music3 evaluation plus stage-offloading plan; dual-component LoRA with llm, transformer, and both targets; multi-token decoding (speculative, Jacobi, chunk overlap); LoRA deps shipped by default.

Model weights are not bundled.

0.0.1

Choose a tag to compare

@TheLimeDev TheLimeDev released this 11 Sep 14:53

First release of Audiyo.

Small Python library for text-to-audio generation with Stable Audio Open, four memory presets, LoRA fine-tuning, and a command line with an interactive mode.

Model weights are not bundled. Accept the checkpoint gate on Hugging Face to generate audio.

See the README for install and quickstart.