Dviri Spams commited on
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,37 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- text-to-speech
|
| 7 |
+
- chatterbox
|
| 8 |
+
- lora
|
| 9 |
+
- onnx
|
| 10 |
+
base_model: ResembleAI/chatterbox
|
| 11 |
---
|
| 12 |
+
|
| 13 |
+
# Chatterbox-Turbo with LoRA input ports
|
| 14 |
+
|
| 15 |
+
Re-exported Chatterbox-Turbo ONNX with 192 optional LoRA adapter input
|
| 16 |
+
ports added to the language_model graph. Enables per-speaker fine-tuning
|
| 17 |
+
via ORT's `Ort::LoraAdapter` API.
|
| 18 |
+
|
| 19 |
+
## Files
|
| 20 |
+
|
| 21 |
+
- `language_model_q4f16.onnx` — graph file (~277 KB)
|
| 22 |
+
- `language_model_q4f16.onnx_data` — external weights file (~700 MB,
|
| 23 |
+
fp32 — quantization to q4f16 is a TODO)
|
| 24 |
+
|
| 25 |
+
## Modifications vs upstream
|
| 26 |
+
|
| 27 |
+
- 96 additive LoRA branches (`MatMul → MatMul → Add`) spliced into each
|
| 28 |
+
GPT2 block's c_attn / c_proj / c_fc / mlp.c_proj projections
|
| 29 |
+
- 192 named optional graph inputs: `lora.layers.{N}.{parent}_{child}.{A|B}`
|
| 30 |
+
- Zero-sized default initializers so the un-adapted graph runs identically
|
| 31 |
+
to the original
|
| 32 |
+
|
| 33 |
+
## Original work
|
| 34 |
+
|
| 35 |
+
Based on [ResembleAI/chatterbox](https://huggingface.co/ResembleAI/chatterbox)
|
| 36 |
+
by Resemble AI, licensed under MIT. This derivative is also MIT-licensed.
|
| 37 |
+
---
|