README.md

metadata

license: agpl-3.0

This repo catalogs my weights for use with my VALL-E implementation as I try and iron out the kinks.

The model currently is in a semi-usable state, and I'm releasing them now in hopes that it also helps jumpstart anyone else that wants to use them.

To reiterate, this is by no means complete. I am not passing this off as competitive.

Models

This repo contains the following configurations:

• config.retnet.yaml / ar+nar-retnet-8: The previously released weights.

  • This configuration utilizes a RetNet (retention based "transformer") as the underlying architecture due to a number of misleading interpretations with comparisons, for better or for worse.
    • Prompt and response embeddings are summed (further RVQ levels gets the previous RVQ levels' embeddings factored in).
    • Tokenizer is a homebrewed "naive" implementation.
  • This model received the most training time between my 4070Ti, 7900XTX, and a few rental rigs to training further progress, entirely at bfloat16 with prodigyopt (and a few optimizer restarts).
  • The later part of training aimed to shuffle between speakers rather than the global pool of utterances to better focus on zero-shot performance. Due to this, I feel it achieved decent zero-shot performance.
  • However, due to the dataset being aggressively trimmed under 12 seconds for memory savings during training, it suffers trying to inference non-short utterances. Additional training may fix this, the following models seemed to adapt well to longer utterances.
    • From the ar+nar-llama-8 experiment, I believe this can be "fixed" with additional training on the currently processed dataset.
  • Prior testing showed that longer prompt durations results in better utterances.
  • Can benefit from additional training, but I recall the average loss being around 1.9 to 2.1.
    • However, due to regressions (or bias from working under llama), I don't think I can optimially train with a RetNet again (both in terms of VRAM consumption and throughput).

• config.llama.yaml / ar+nar-llama-8: The most recent-ishly trained weights after learning from my mistakes.

  • This configuration utilizes Llama's attention-based transformer as the underlying architecture, making use of creature comforts like RoPE, GQA, and memory-efficient attention (trained under xformers, shouldn't really affect things).
    • Prompt and response embeddings are NOT summed (each RVQ level only attends to the current RVQ level).
    • Utilizes a HF tokenizer for "optimal" vocab.
    • The current RVQ level is included as a token as well to help guide NAR tasks better.
  • This model received a few days of training on my 4xV100s, stepping up the duration window to try and better make the model inference for longer utterances.
    • Some sessions end up training the current duration window for a few epochs, but I don't know how much it affected things.
  • However, it seems to only do well with long utterances. Short utterances fumble. I believe further training with a variety of durations should allow the AR to handle a variety of durations.
    • I believe the "slowly stepping up the context length" only works for text, and not audio.
    • Addendum: Additional brief training for a variety of duration lengths seemed to have mostly fixed this issue.
  • Zero-shot performance leaves a bit to be desired, as it did not receive the special training prioritizing shuffling between speakers rather than the global pool of utterances.
    • Addendum: Additional brief training for sampling based on speaker per "epoch" (per dataloader, not dataset) seemed to slightly improve it.
  • Testing showed that, despite also stepping up the prompt duration, it really likes three second prompts.
  • Definitely needs additional training, but the next way to go is unknown.
    • Naturally, training it on a "next RVQ level is half as likely" distribution introduces some crust as the later RVQ levels are less accurate, introducing noise and artifacts.
    • Naively training it on equally distributed RVQ levels does lobotomize the AR.
    • Additional training on the AR will see huge diminishing returns, so I don't know if it's worth doing so.
  • Seems to be a decent foundation for "distillation", at the very least for LoRA training.

• config.llama.split.yaml / ar-llama-1 + nar-llama-8: The above model, but split and trained a little bit more.

  • This experiment is to see whether the AR and NAR benefitted from being split up after enough pretraining, to un-"lobotomize" any penalties from attending to two different tasks (as the AR predicts the next token, and the NAR predicts the same token but a different level).
  • I believe I trained each separate model an additional extra day for another additional audio-duration window for similar training lengths.
  • I don't think audio quality differs a non-trivial amount to warrant splitting the model.
    • From recent experiments, it does seem a NAR-only model is beneficial.

Some additional configurations have been explored with, but experiments have not been fruitful:

• Exotic wrappers like BitNet seemed to yield little gains in inferencing, somehow.

• Mamba / Mamba2-based models have shown that it's really hard to have an AR+NAR model.

• A NAR only model has been experimented with, but seemed utterly useless in practice.

  • The underlying architecture will query the model for the duration, and then inference all RVQ levels in parallel (one level at a time).
  • Despite working in the overfitting test trainer and decent training metrics, inferencing will have the model fall completely flat.
  • I have zero ideas for which path to go with for further experimentation.

• A Descript-Audio-Codec based model has been experimented with, but has not seem fruitful.

  • This model would make use of 16 layers instead of the default 12 layers. I feel the performance hit is negligible, even with the additional tokens-per-frame increase with DAC.
  • This utilizes DAC's 44Khz model (erroniously at an actual 44KHz instead of 44.1KHz), as audio quantized through the 24KHz model will always diverge.
  • I imagine due to the nature of DAC leaving very little room for errors (a testament to how "optimized" the codes are), it's really hard to model an LM with it.
    • Output audio is rather crunchy and crusty from the later RVQ levels being inaccurate enough.
  • I'm not sure which path to go with it for further experimentation:
    • Utilizing the original model for embeddings or last hidden state as the input embeddings for the prompt/response.
      • I don't think this is the way to go. It seems negligible for additional complexity.
    • Training a dedicated NAR model in hopes to bolster the later RVQ levels' performance, as the issues come from the later RVQ levels.
    • Utilizing an interleaved pattern instead to make better use of attending to past tokens for all levels.