Update official model card
Browse files
README.md
CHANGED
|
@@ -11,7 +11,8 @@ tags:
|
|
| 11 |
|
| 12 |
# FireRedTTS3-Base for SGLang Omni
|
| 13 |
|
| 14 |
-
|
|
|
|
| 15 |
[SGLang Omni](https://github.com/sgl-project/sglang-omni). It supports
|
| 16 |
zero-shot voice cloning from one reference recording and its transcript.
|
| 17 |
Audio output is 24 kHz mono, with WAV responses and streaming PCM available
|
|
@@ -39,20 +40,15 @@ source .venv/bin/activate
|
|
| 39 |
uv pip install --prerelease=allow -e '.[firered-tts3]'
|
| 40 |
```
|
| 41 |
|
| 42 |
-
|
| 43 |
-
is merged, check out the corresponding PR revision before installing; an older
|
| 44 |
-
PyPI release or checkout does not provide this model. Run the following commands
|
| 45 |
-
from the `sglang-omni` repository root.
|
| 46 |
|
| 47 |
## Download this checkpoint
|
| 48 |
|
| 49 |
-
|
| 50 |
-
[
|
| 51 |
-
After this PR, we plan to publish the official checkpoint at
|
| 52 |
-
[FireRedTeam/FireRedTTS3-Base](https://huggingface.co/FireRedTeam/FireRedTTS3-Base).
|
| 53 |
|
| 54 |
```bash
|
| 55 |
-
hf download
|
| 56 |
--local-dir ./models/FireRedTTS3-Base
|
| 57 |
```
|
| 58 |
|
|
@@ -182,7 +178,7 @@ if __name__ == "__main__":
|
|
| 182 |
## Serving notes
|
| 183 |
|
| 184 |
- The current deployment supports a single GPU (`tp_size=1`).
|
| 185 |
-
-
|
| 186 |
- Use an explicit language: `Chinese` (default), `English`, `Japanese`, `Korean`,
|
| 187 |
`German`, `French`, `Russian`, `Portuguese`, `Spanish`, or `Italian`.
|
| 188 |
Automatic language detection (`Auto`) is unsupported. These are the stock
|
|
@@ -193,6 +189,8 @@ if __name__ == "__main__":
|
|
| 193 |
- Put flow-solver settings under `stage_params.talker`, for example
|
| 194 |
`{"n_timesteps": 10, "inference_cfg": 2.0, "max_patches": 200}`. Unknown
|
| 195 |
top-level fields are not forwarded as model settings.
|
|
|
|
|
|
|
| 196 |
- `seed` controls the per-request flow sampler. Different precisions, hardware,
|
| 197 |
or batching schedules can produce different results with the same seed.
|
| 198 |
- For our published model-quality results, see the
|
|
|
|
| 11 |
|
| 12 |
# FireRedTTS3-Base for SGLang Omni
|
| 13 |
|
| 14 |
+
This repository contains the officially released HF-format FireRedTTS3-Base
|
| 15 |
+
checkpoint for inference with
|
| 16 |
[SGLang Omni](https://github.com/sgl-project/sglang-omni). It supports
|
| 17 |
zero-shot voice cloning from one reference recording and its transcript.
|
| 18 |
Audio output is 24 kHz mono, with WAV responses and streaming PCM available
|
|
|
|
| 40 |
uv pip install --prerelease=allow -e '.[firered-tts3]'
|
| 41 |
```
|
| 42 |
|
| 43 |
+
Run the following commands from the `sglang-omni` repository root.
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
## Download this checkpoint
|
| 46 |
|
| 47 |
+
Download the official HF-format checkpoint from
|
| 48 |
+
[FireRedTeam/FireRedTTS3-Base-HF](https://huggingface.co/FireRedTeam/FireRedTTS3-Base-HF).
|
|
|
|
|
|
|
| 49 |
|
| 50 |
```bash
|
| 51 |
+
hf download FireRedTeam/FireRedTTS3-Base-HF \
|
| 52 |
--local-dir ./models/FireRedTTS3-Base
|
| 53 |
```
|
| 54 |
|
|
|
|
| 178 |
## Serving notes
|
| 179 |
|
| 180 |
- The current deployment supports a single GPU (`tp_size=1`).
|
| 181 |
+
- The `input` text is limited to 400 characters.
|
| 182 |
- Use an explicit language: `Chinese` (default), `English`, `Japanese`, `Korean`,
|
| 183 |
`German`, `French`, `Russian`, `Portuguese`, `Spanish`, or `Italian`.
|
| 184 |
Automatic language detection (`Auto`) is unsupported. These are the stock
|
|
|
|
| 189 |
- Put flow-solver settings under `stage_params.talker`, for example
|
| 190 |
`{"n_timesteps": 10, "inference_cfg": 2.0, "max_patches": 200}`. Unknown
|
| 191 |
top-level fields are not forwarded as model settings.
|
| 192 |
+
- HTTP, batch, and WebSocket requests do not forward `extra_params`. The SDK
|
| 193 |
+
`GenerateRequest.extra_params` field remains available for SDK requests.
|
| 194 |
- `seed` controls the per-request flow sampler. Different precisions, hardware,
|
| 195 |
or batching schedules can produce different results with the same seed.
|
| 196 |
- For our published model-quality results, see the
|