51 lines
2.2 KiB
Markdown
51 lines
2.2 KiB
Markdown
# vall_e.cpp
|
|
|
|
This is an implementation that makes use of [llama.cpp](https://github.com/ggerganov/llama.cpp/) and [encodec.cpp](https://github.com/PABannier/encodec.cpp).
|
|
|
|
At the moment it's ***very*** barebones as I try and wrestle with `llama.cpp`'s API without needing to modify its code.
|
|
|
|
## Build
|
|
|
|
Populate `./include/` with the `llama.cpp` and `encodec.cpp` headers.
|
|
|
|
Populate `./libs/` with the compiled libraries of `llama.cpp` and `encodec.cpp`.
|
|
|
|
Run `make`.
|
|
|
|
|
|
### Required Modifications
|
|
|
|
`encodec.cpp` requires updating its GGML copy to the latest version, which requires a few lines to get the CPU backend working.
|
|
`llama.cpp` *might* not require any modifications, but:
|
|
* `llm.build_vall_e` can mostly copy `llm.build_llama`, but with:
|
|
* `KQ_mask = build_inp_KQ_mask( lctx.cparams.causal_attn )`
|
|
* a unified output head (pain)
|
|
* OR adjusting the `model.output` to the correct classifier head
|
|
* OR slicing that tensor with the right range (`ggml_view_2d` confuses me)
|
|
* both require also require `*const_cast<uint32_t*>(&ctx->model.hparams.n_vocab) = output->ne[1];` because the logits are tied to `n_vocab`
|
|
* commenting out `GGML_ABORT("input/output layer tensor %s used with a layer number", tn.str().c_str());` because grabbing embeddings/classifiers require using `bid` to trick it thinking it's part of a layer
|
|
* some helper functions to retrieve the embeddings tensor from the model
|
|
* some helper functions to set the target classifier head
|
|
|
|
## To-Do
|
|
|
|
* [x] converted model to GGUF
|
|
* [ ] convert it without modifying any of the existing code, as the tokenizer requires some care
|
|
* [x] basic framework
|
|
* [x] load the quantized model
|
|
* [x] orchestrate the required embeddings
|
|
* [x] juggle the output head / classifier properly
|
|
* [ ] phonemize text
|
|
* with the help of espeak-ng
|
|
* [ ] tokenize phonemes
|
|
* the tokenizer is being a huge thorn on actual sequences
|
|
* [x] load audio from disk
|
|
* [x] encode audio
|
|
* [x] sum embeddings for the `prom` and prior `resp`s
|
|
* [x] `AR` sampling
|
|
* [ ] `NAR-len` demasking sampling
|
|
* [x] `NAR` sampling
|
|
* [x] decode audio to disk
|
|
* [ ] a functional CLI
|
|
* [ ] actually make it work
|
|
* it seems naively stitching the model together isn't good enough since the output is wrong, it most likely needs training with a glued together classifier |