Closed
Conversation
Owner
Owner
|
How about we close this and collaborate on #499? |
Contributor
Author
|
oops! closing in favor of #499 which already has this fully implemented with more robust token handling and audio features |
Owner
|
No worries! My bad, since sfx models are new and this model is quite complex(lots of moving pieces) I put it on hold for a couple weeks to push our Swift SDK and improve inference here. |
Owner
|
Will share updates in #499 later today or tomorrow. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Context
ACE-Step 1.5 (https://github.com/ace-step/ACE-Step-1.5) by StepFun. It is a state-of-the-art text-to-audio model capable of generating full songs with vocals and instrumentation directly from text prompts. It is highly efficient and designed for consumer hardware.
Description
adds ACE-Step to the tts pipeline. the model is a hybrid architecture utilizing an LLM (Qwen3) as a text-conditioner/planner, a Diffusion Transformer (DiT) to generate audio latents, and an Oobleck Autoencoder VAE to decode latents back to PCM audio.
the upstream repo provides a partial MLX backend for the DiT and VAE, but relies on PyTorch and
transformersfor theAceStepConditionEncoderand the text/lyric embedding phases. this PR fully ports the remaining components (AceStepConditionEncoderandAceStepLyricEncoder) to pure MLX and integrates the text prompting viamlx-lm, making the entire generation pipeline 100% native MLX with zero PyTorch runtime dependencies.also includes a conversion script since upstream only distributes pytorch weights.
Changes in the codebase
mlx_audio/tts/models/acestep/acestep.py- model implementation and pipeline logicmlx_audio/tts/models/acestep/conditioner.py- pure MLX port ofAceStepConditionEncoderandAceStepLyricEncodermlx_audio/tts/models/acestep/config.py- config dataclassesmlx_audio/tts/models/acestep/convert.py- pt -> safetensors conversion scriptmlx_audio/tts/models/acestep/dit.py- MLX DiT decodermlx_audio/tts/models/acestep/generate_utils.py- MLX diffusion loopsmlx_audio/tts/models/acestep/vae.py- MLX VAE decodermlx_audio/tts/models/acestep/README.md- setup + usagemlx_audio/tts/models/__init__.py,mlx_audio/tts/utils.py- registrationmlx_audio/tts/tests/test_acestep.py- unit tests for the pure MLX condition encoderChanges outside the codebase
none.
Additional information
transformersanddiffusersduring runtimeChecklist