Repository: nari-labs/dia
Branch: main
Commit: 876125e461a0
Files: 24
Total size: 141.3 KB
Directory structure:
gitextract_f5lpo0wj/
├── .github/
│ └── workflows/
│ └── ci.yaml
├── .gitignore
├── .python-version
├── LICENSE
├── README.md
├── app.py
├── cli.py
├── dia/
│ ├── __init__.py
│ ├── audio.py
│ ├── config.py
│ ├── layers.py
│ ├── model.py
│ └── state.py
├── docker/
│ ├── Dockerfile.cpu
│ └── Dockerfile.gpu
├── example/
│ ├── benchmark.py
│ ├── simple-cpu.py
│ ├── simple-mac.py
│ ├── simple.py
│ ├── simple_batch.py
│ ├── voice_clone.py
│ └── voice_clone_batch.py
├── hf.py
└── pyproject.toml
================================================
FILE CONTENTS
================================================
================================================
FILE: .github/workflows/ci.yaml
================================================
name: Continuous Integration
on:
pull_request:
branches:
- main
jobs:
lint_and_format:
runs-on: ubuntu-latest
name: Lint and Format
steps:
- uses: actions/checkout@v4
- uses: astral-sh/ruff-action@v3
with:
version: latest
- name: Check Lint using Ruff
run: ruff check
- name: Check Format using Ruff
run: ruff format --check --diff
================================================
FILE: .gitignore
================================================
# Python-generated files
__pycache__/
*.py[oc]
build/
dist/
wheels/
*.egg-info
# Virtual environments
.venv
.idea/
.gradio
**/*.pth
**/*.safetensors
**/*.mp3
!example_prompt.mp3
**/*.txt
.ruff_cache
.ipynb_checkpoints
config.json
================================================
FILE: .python-version
================================================
3.10
================================================
FILE: LICENSE
================================================
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright 2025 Nari Labs
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
================================================
FILE: README.md
================================================
<p align="center">
<a href="https://github.com/nari-labs/dia">
<img src="./dia/static/images/banner.png">
</a>
</p>
<p align="center">
<a href="https://tally.so/r/meokbo" target="_blank"><img alt="Static Badge" src="https://img.shields.io/badge/Join-Waitlist-white?style=for-the-badge"></a>
<a href="https://discord.gg/bJq6vjRRKv" target="_blank"><img src="https://img.shields.io/badge/Discord-Join%20Chat-7289DA?logo=discord&style=for-the-badge"></a>
<a href="https://github.com/nari-labs/dia/blob/main/LICENSE" target="_blank"><img src="https://img.shields.io/badge/License-Apache_2.0-blue.svg?style=for-the-badge" alt="LICENSE"></a>
</p>
<p align="center">
<a href="https://huggingface.co/nari-labs/Dia-1.6B-0626"><img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/model-on-hf-lg-dark.svg" alt="Model on HuggingFace" height=42 ></a>
<a href="https://huggingface.co/spaces/nari-labs/Dia-1.6B"><img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/open-in-hf-spaces-lg-dark.svg" alt="Space on HuggingFace" height=38></a>
</p>
Dia is a 1.6B parameter text to speech model created by Nari Labs.
**UPDATE 🤗(06/27)**: Dia is now available through [Hugging Face Transformers](https://github.com/huggingface/transformers)!
**UPDATE 🚀(11/19)**: Dia2 is released on Github and HuggingFace [link](https://github.com/nari-labs/dia2)!
Dia **directly generates highly realistic dialogue from a transcript**. You can condition the output on audio, enabling emotion and tone control. The model can also produce nonverbal communications like laughter, coughing, clearing throat, etc.
To accelerate research, we are providing access to pretrained model checkpoints and inference code. The model weights are hosted on [Hugging Face](https://huggingface.co/nari-labs/Dia-1.6B-0626). The model only supports English generation at the moment.
We also provide a [demo page](https://yummy-fir-7a4.notion.site/dia) comparing our model to [ElevenLabs Studio](https://elevenlabs.io/studio) and [Sesame CSM-1B](https://github.com/SesameAILabs/csm).
- We have a ZeroGPU Space running! Try it now [here](https://huggingface.co/spaces/nari-labs/Dia-1.6B-0626). Thanks to the HF team for the support :)
- Join our [discord server](https://discord.gg/bJq6vjRRKv) for community support and access to new features.
- Play with a larger version of Dia: generate fun conversations, remix content, and share with friends. 🔮 Join the [waitlist](https://tally.so/r/meokbo) for early access.
## Generation Guidelines
- Keep input text length moderate
- Short input (corresponding to under 5s of audio) will sound unnatural
- Very long input (corresponding to over 20s of audio) will make the speech unnaturally fast.
- Use non-verbal tags sparingly, from the list in the README. Overusing or using unlisted non-verbals may cause weird artifacts.
- Always begin input text with `[S1]`, and always alternate between `[S1]` and `[S2]` (i.e. `[S1]`... `[S1]`... is not good)
- When using audio prompts (voice cloning), follow these instructions carefully:
- Provide the transcript of the to-be cloned audio before the generation text.
- Transcript must use `[S1]`, `[S2]` speaker tags correctly (i.e. single speaker: `[S1]`..., two speakers: `[S1]`... `[S2]`...)
- Duration of the to-be cloned audio should be 5~10 seconds for the best results.
(Keep in mind: 1 second ≈ 86 tokens)
- Put `[S1]` or `[S2]` (the second-to-last speaker's tag) at the end of the audio to improve audio quality at the end
## Quickstart
### Transformers Support
We now have a [Hugging Face Transformers](https://github.com/huggingface/transformers) implementation of Dia! You should install `main` branch of `transformers` to use it. See [hf.py](hf.py) for more information.
<details>
<summary>View more details</summary>
Install `main` branch of `transformers`
```bash
pip install git+https://github.com/huggingface/transformers.git
# or install with uv
uv pip install git+https://github.com/huggingface/transformers.git
```
Run `hf.py`. The file is as below.
```python
from transformers import AutoProcessor, DiaForConditionalGeneration
torch_device = "cuda"
model_checkpoint = "nari-labs/Dia-1.6B-0626"
text = [
"[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
]
processor = AutoProcessor.from_pretrained(model_checkpoint)
inputs = processor(text=text, padding=True, return_tensors="pt").to(torch_device)
model = DiaForConditionalGeneration.from_pretrained(model_checkpoint).to(torch_device)
outputs = model.generate(
**inputs, max_new_tokens=3072, guidance_scale=3.0, temperature=1.8, top_p=0.90, top_k=45
)
outputs = processor.batch_decode(outputs)
processor.save_audio(outputs, "example.mp3")
```
</details>
### Run with this repo
<details>
<summary> Install via pip </summary>
```bash
# Clone this repository
git clone https://github.com/nari-labs/dia.git
cd dia
# Optionally
python -m venv .venv && source .venv/bin/activate
# Install dia
pip install -e .
```
Or you can install without cloning.
```bash
# Install directly from GitHub
pip install git+https://github.com/nari-labs/dia.git
```
Now, run some examples.
```bash
python example/simple.py
```
</details>
<details>
<summary>Install via uv</summary>
You need [uv](https://docs.astral.sh/uv/) to be installed.
```bash
# Clone this repository
git clone https://github.com/nari-labs/dia.git
cd dia
```
Run some examples directly.
```bash
uv run example/simple.py
```
</details>
<details>
<summary>Run Gradio UI</summary>
```bash
python app.py
# Or if you have uv installed
uv run app.py
```
</details>
<details>
<summary>Run with CLI</summary>
```bash
python cli.py --help
# Or if you have uv installed
uv run cli.py --help
```
</details>
> [!NOTE]
> The model was not fine-tuned on a specific voice. Hence, you will get different voices every time you run the model.
> You can keep speaker consistency by either adding an audio prompt, or fixing the seed.
> [!IMPORTANT]
> If you are using 5000 series GPU, you should use torch 2.8 nightly. Look at the issue [#26](https://github.com/nari-labs/dia/issues/26) for more details.
## Features
- Generate dialogue via `[S1]` and `[S2]` tag
- Generate non-verbal like `(laughs)`, `(coughs)`, etc.
- Below verbal tags will be recognized, but might result in unexpected output.
- `(laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)`
- Voice cloning. See [`example/voice_clone.py`](example/voice_clone.py) for more information.
- In the Hugging Face space, you can upload the audio you want to clone and place its transcript before your script. Make sure the transcript follows the required format. The model will then output only the content of your script.
## 💻 Hardware and Inference Speed
Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon.
The initial run will take longer as the Descript Audio Codec also needs to be downloaded.
These are the speed we benchmarked in RTX 4090.
| precision | realtime factor w/ compile | realtime factor w/o compile | VRAM |
|:-:|:-:|:-:|:-:|
| `bfloat16` | x2.1 | x1.5 | ~4.4GB |
| `float16` | x2.2 | x1.3 | ~4.4GB |
| `float32` | x1 | x0.9 | ~7.9GB |
We will be adding a quantized version in the future.
If you don't have hardware available or if you want to play with bigger versions of our models, join the waitlist [here](https://tally.so/r/meokbo).
## 🪪 License
This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.
## ⚠️ Disclaimer
This project offers a high-fidelity speech generation model intended for research and educational use. The following uses are **strictly forbidden**:
- **Identity Misuse**: Do not produce audio resembling real individuals without permission.
- **Deceptive Content**: Do not use this model to generate misleading content (e.g. fake news)
- **Illegal or Malicious Use**: Do not use this model for activities that are illegal or intended to cause harm.
By using this model, you agree to uphold relevant legal standards and ethical responsibilities. We **are not responsible** for any misuse and firmly oppose any unethical usage of this technology.
## 🔭 TODO / Future Work
- Docker support for ARM architecture and MacOS.
- Optimize inference speed.
- Add quantization for memory efficiency.
## 🤝 Contributing
We are a tiny team of 1 full-time and 1 part-time research-engineers. We are extra-welcome to any contributions!
Join our [Discord Server](https://discord.gg/bJq6vjRRKv) for discussions.
## 🤗 Acknowledgements
- We thank the [Google TPU Research Cloud program](https://sites.research.google/trc/about/) for providing computation resources.
- Our work was heavily inspired by [SoundStorm](https://arxiv.org/abs/2305.09636), [Parakeet](https://jordandarefsky.com/blog/2024/parakeet/), and [Descript Audio Codec](https://github.com/descriptinc/descript-audio-codec).
- Hugging Face for providing the ZeroGPU Grant.
- "Nari" is a pure Korean word for lily.
- We thank Jason Y. for providing help with data filtering.
## ⭐ Star History
<a href="https://www.star-history.com/#nari-labs/dia&Date">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://api.star-history.com/svg?repos=nari-labs/dia&type=Date&theme=dark" />
<source media="(prefers-color-scheme: light)" srcset="https://api.star-history.com/svg?repos=nari-labs/dia&type=Date" />
<img alt="Star History Chart" src="https://api.star-history.com/svg?repos=nari-labs/dia&type=Date" />
</picture>
</a>
================================================
FILE: app.py
================================================
import argparse
import contextlib
import io
import random
import tempfile
import time
from pathlib import Path
from typing import Optional, Tuple
import gradio as gr
import numpy as np
import soundfile as sf
import torch
from dia.model import Dia
# --- Global Setup ---
parser = argparse.ArgumentParser(description="Gradio interface for Nari TTS")
parser.add_argument("--device", type=str, default=None, help="Force device (e.g., 'cuda', 'mps', 'cpu')")
parser.add_argument("--share", action="store_true", help="Enable Gradio sharing")
args = parser.parse_args()
# Determine device
if args.device:
device = torch.device(args.device)
elif torch.cuda.is_available():
device = torch.device("cuda")
# Simplified MPS check for broader compatibility
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
# Basic check is usually sufficient, detailed check can be problematic
device = torch.device("mps")
else:
device = torch.device("cpu")
print(f"Using device: {device}")
# Load Nari model and config
print("Loading Nari model...")
try:
dtype_map = {
"cpu": "float32",
"mps": "float32", # Apple M series – better with float32
"cuda": "float16", # NVIDIA – better with float16
}
dtype = dtype_map.get(device.type, "float16")
print(f"Using device: {device}, attempting to load model with {dtype}")
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype=dtype, device=device)
except Exception as e:
print(f"Error loading Nari model: {e}")
raise
def set_seed(seed: int):
"""Sets the random seed for reproducibility."""
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def run_inference(
text_input: str,
audio_prompt_text_input: str,
audio_prompt_input: Optional[Tuple[int, np.ndarray]],
max_new_tokens: int,
cfg_scale: float,
temperature: float,
top_p: float,
cfg_filter_top_k: int,
speed_factor: float,
seed: Optional[int] = None,
):
"""
Runs Nari inference using the globally loaded model and provided inputs.
Uses temporary files for text and audio prompt compatibility with inference.generate.
"""
global model, device # Access global model, config, device
console_output_buffer = io.StringIO()
with contextlib.redirect_stdout(console_output_buffer):
# Prepend transcript text if audio_prompt provided
if audio_prompt_input and audio_prompt_text_input and not audio_prompt_text_input.isspace():
text_input = audio_prompt_text_input + "\n" + text_input
text_input = text_input.strip()
if audio_prompt_input and (not audio_prompt_text_input or audio_prompt_text_input.isspace()):
raise gr.Error("Audio Prompt Text input cannot be empty.")
if not text_input or text_input.isspace():
raise gr.Error("Text input cannot be empty.")
# Preprocess Audio
temp_txt_file_path = None
temp_audio_prompt_path = None
output_audio = (44100, np.zeros(1, dtype=np.float32))
try:
prompt_path_for_generate = None
if audio_prompt_input is not None:
sr, audio_data = audio_prompt_input
# Check if audio_data is valid
if audio_data is None or audio_data.size == 0 or audio_data.max() == 0: # Check for silence/empty
gr.Warning("Audio prompt seems empty or silent, ignoring prompt.")
else:
# Save prompt audio to a temporary WAV file
with tempfile.NamedTemporaryFile(mode="wb", suffix=".wav", delete=False) as f_audio:
temp_audio_prompt_path = f_audio.name # Store path for cleanup
# Basic audio preprocessing for consistency
# Convert to float32 in [-1, 1] range if integer type
if np.issubdtype(audio_data.dtype, np.integer):
max_val = np.iinfo(audio_data.dtype).max
audio_data = audio_data.astype(np.float32) / max_val
elif not np.issubdtype(audio_data.dtype, np.floating):
gr.Warning(f"Unsupported audio prompt dtype {audio_data.dtype}, attempting conversion.")
# Attempt conversion, might fail for complex types
try:
audio_data = audio_data.astype(np.float32)
except Exception as conv_e:
raise gr.Error(f"Failed to convert audio prompt to float32: {conv_e}")
# Ensure mono (average channels if stereo)
if audio_data.ndim > 1:
if audio_data.shape[0] == 2: # Assume (2, N)
audio_data = np.mean(audio_data, axis=0)
elif audio_data.shape[1] == 2: # Assume (N, 2)
audio_data = np.mean(audio_data, axis=1)
else:
gr.Warning(
f"Audio prompt has unexpected shape {audio_data.shape}, taking first channel/axis."
)
audio_data = (
audio_data[0] if audio_data.shape[0] < audio_data.shape[1] else audio_data[:, 0]
)
audio_data = np.ascontiguousarray(audio_data) # Ensure contiguous after slicing/mean
# Write using soundfile
try:
sf.write(
temp_audio_prompt_path, audio_data, sr, subtype="FLOAT"
) # Explicitly use FLOAT subtype
prompt_path_for_generate = temp_audio_prompt_path
print(f"Created temporary audio prompt file: {temp_audio_prompt_path} (orig sr: {sr})")
except Exception as write_e:
print(f"Error writing temporary audio file: {write_e}")
raise gr.Error(f"Failed to save audio prompt: {write_e}")
# Set and Display Generation Seed
if seed is None or seed < 0:
seed = random.randint(0, 2**32 - 1)
print(f"\nNo seed provided, generated random seed: {seed}\n")
else:
print(f"\nUsing user-selected seed: {seed}\n")
set_seed(seed)
# Run Generation
print(f'Generating speech: \n"{text_input}"\n')
start_time = time.time()
# Use torch.inference_mode() context manager for the generation call
with torch.inference_mode():
output_audio_np = model.generate(
text_input,
max_tokens=max_new_tokens,
cfg_scale=cfg_scale,
temperature=temperature,
top_p=top_p,
cfg_filter_top_k=cfg_filter_top_k, # Pass the value here
use_torch_compile=False, # Keep False for Gradio stability
audio_prompt=prompt_path_for_generate,
verbose=True,
)
end_time = time.time()
print(f"Generation finished in {end_time - start_time:.2f} seconds.\n")
# 4. Convert Codes to Audio
if output_audio_np is not None:
# Get sample rate from the loaded DAC model
output_sr = 44100
# --- Slow down audio ---
original_len = len(output_audio_np)
# Ensure speed_factor is positive and not excessively small/large to avoid issues
speed_factor = max(0.1, min(speed_factor, 5.0))
target_len = int(original_len / speed_factor) # Target length based on speed_factor
if target_len != original_len and target_len > 0: # Only interpolate if length changes and is valid
x_original = np.arange(original_len)
x_resampled = np.linspace(0, original_len - 1, target_len)
resampled_audio_np = np.interp(x_resampled, x_original, output_audio_np)
output_audio = (
output_sr,
resampled_audio_np.astype(np.float32),
) # Use resampled audio
print(
f"Resampled audio from {original_len} to {target_len} samples for {speed_factor:.2f}x speed."
)
else:
output_audio = (
output_sr,
output_audio_np,
) # Keep original if calculation fails or no change
print(f"Skipping audio speed adjustment (factor: {speed_factor:.2f}).")
# --- End slowdown ---
print(f"Audio conversion successful. Final shape: {output_audio[1].shape}, Sample Rate: {output_sr}")
# Explicitly convert to int16 to prevent Gradio warning
if output_audio[1].dtype == np.float32 or output_audio[1].dtype == np.float64:
audio_for_gradio = np.clip(output_audio[1], -1.0, 1.0)
audio_for_gradio = (audio_for_gradio * 32767).astype(np.int16)
output_audio = (output_sr, audio_for_gradio)
print("Converted audio to int16 for Gradio output.")
else:
print("\nGeneration finished, but no valid tokens were produced.")
# Return default silence
gr.Warning("Generation produced no output.")
except Exception as e:
print(f"Error during inference: {e}")
import traceback
traceback.print_exc()
# Re-raise as Gradio error to display nicely in the UI
raise gr.Error(f"Inference failed: {e}")
finally:
# Cleanup Temporary Files defensively
if temp_txt_file_path and Path(temp_txt_file_path).exists():
try:
Path(temp_txt_file_path).unlink()
print(f"Deleted temporary text file: {temp_txt_file_path}")
except OSError as e:
print(f"Warning: Error deleting temporary text file {temp_txt_file_path}: {e}")
if temp_audio_prompt_path and Path(temp_audio_prompt_path).exists():
try:
Path(temp_audio_prompt_path).unlink()
print(f"Deleted temporary audio prompt file: {temp_audio_prompt_path}")
except OSError as e:
print(f"Warning: Error deleting temporary audio prompt file {temp_audio_prompt_path}: {e}")
# After generation, capture the printed output
console_output = console_output_buffer.getvalue()
return output_audio, seed, console_output
# --- Create Gradio Interface ---
css = """
#col-container {max-width: 90%; margin-left: auto; margin-right: auto;}
"""
# Attempt to load default text from example.txt
default_text = "[S1] Dia is an open weights text to dialogue model. \n[S2] You get full control over scripts and voices. \n[S1] Wow. Amazing. (laughs) \n[S2] Try it now on Git hub or Hugging Face."
example_txt_path = Path("./example.txt")
if example_txt_path.exists():
try:
default_text = example_txt_path.read_text(encoding="utf-8").strip()
if not default_text: # Handle empty example file
default_text = "Example text file was empty."
except Exception as e:
print(f"Warning: Could not read example.txt: {e}")
# Build Gradio UI
with gr.Blocks(css=css, theme="gradio/dark") as demo:
gr.Markdown("# Nari Text-to-Speech Synthesis")
with gr.Row(equal_height=False):
with gr.Column(scale=1):
with gr.Accordion("Audio Reference Prompt (Optional)", open=False):
audio_prompt_input = gr.Audio(
label="Audio Prompt (Optional)",
show_label=True,
sources=["upload", "microphone"],
type="numpy",
)
audio_prompt_text_input = gr.Textbox(
label="Transcript of Audio Prompt (Required if using Audio Prompt)",
placeholder="Enter text here...",
value="",
lines=5, # Increased lines
)
text_input = gr.Textbox(
label="Text To Generate",
placeholder="Enter text here...",
value=default_text,
lines=5, # Increased lines
)
with gr.Accordion("Generation Parameters", open=False):
max_new_tokens = gr.Slider(
label="Max New Tokens (Audio Length)",
minimum=860,
maximum=3072,
value=model.config.decoder_config.max_position_embeddings, # Use config default if available, else fallback
step=50,
info="Controls the maximum length of the generated audio (more tokens = longer audio).",
)
cfg_scale = gr.Slider(
label="CFG Scale (Guidance Strength)",
minimum=1.0,
maximum=5.0,
value=3.0, # Default from inference.py
step=0.1,
info="Higher values increase adherence to the text prompt.",
)
temperature = gr.Slider(
label="Temperature (Randomness)",
minimum=1.0,
maximum=2.5,
value=1.8, # Default from inference.py
step=0.05,
info="Lower values make the output more deterministic, higher values increase randomness.",
)
top_p = gr.Slider(
label="Top P (Nucleus Sampling)",
minimum=0.70,
maximum=1.0,
value=0.95, # Default from inference.py
step=0.01,
info="Filters vocabulary to the most likely tokens cumulatively reaching probability P.",
)
cfg_filter_top_k = gr.Slider(
label="CFG Filter Top K",
minimum=15,
maximum=100,
value=45,
step=1,
info="Top k filter for CFG guidance.",
)
speed_factor_slider = gr.Slider(
label="Speed Factor",
minimum=0.8,
maximum=1.0,
value=1.0,
step=0.02,
info="Adjusts the speed of the generated audio (1.0 = original speed).",
)
seed_input = gr.Number(
label="Generation Seed (Optional)",
value=-1,
precision=0, # No decimal points
step=1,
interactive=True,
info="Set a generation seed for reproducible outputs. Leave empty or -1 for random seed.",
)
run_button = gr.Button("Generate Audio", variant="primary")
with gr.Column(scale=1):
audio_output = gr.Audio(
label="Generated Audio",
type="numpy",
autoplay=False,
)
seed_output = gr.Textbox(label="Generation Seed", interactive=False)
console_output = gr.Textbox(label="Console Output Log", lines=10, interactive=False)
# Link button click to function
run_button.click(
fn=run_inference,
inputs=[
text_input,
audio_prompt_text_input,
audio_prompt_input,
max_new_tokens,
cfg_scale,
temperature,
top_p,
cfg_filter_top_k,
speed_factor_slider,
seed_input,
],
outputs=[
audio_output,
seed_output,
console_output,
], # Add status_output here if using it
api_name="generate_audio",
)
# Add examples (ensure the prompt path is correct or remove it if example file doesn't exist)
example_prompt_path = "./example_prompt.mp3" # Adjust if needed
examples_list = [
[
"[S1] Oh fire! Oh my goodness! What's the procedure? What to we do people? The smoke could be coming through an air duct! \n[S2] Oh my god! Okay.. it's happening. Everybody stay calm! \n[S1] What's the procedure... \n[S2] Everybody stay fucking calm!!!... Everybody fucking calm down!!!!! \n[S1] No! No! If you touch the handle, if its hot there might be a fire down the hallway! ",
None,
3072,
3.0,
1.8,
0.95,
45,
1.0,
],
[
"[S1] Open weights text to dialogue model. \n[S2] You get full control over scripts and voices. \n[S1] I'm biased, but I think we clearly won. \n[S2] Hard to disagree. (laughs) \n[S1] Thanks for listening to this demo. \n[S2] Try it now on Git hub and Hugging Face. \n[S1] If you liked our model, please give us a star and share to your friends. \n[S2] This was Nari Labs.",
example_prompt_path if Path(example_prompt_path).exists() else None,
3072,
3.0,
1.8,
0.95,
45,
1.0,
],
]
if examples_list:
gr.Examples(
examples=examples_list,
inputs=[
text_input,
audio_prompt_input,
max_new_tokens,
cfg_scale,
temperature,
top_p,
cfg_filter_top_k,
speed_factor_slider,
seed_input,
],
outputs=[audio_output],
fn=run_inference,
cache_examples=False,
label="Examples (Click to Run)",
)
else:
gr.Markdown("_(No examples configured or example prompt file missing)_")
# --- Launch the App ---
if __name__ == "__main__":
print("Launching Gradio interface...")
# set `GRADIO_SERVER_NAME`, `GRADIO_SERVER_PORT` env vars to override default values
# use `GRADIO_SERVER_NAME=0.0.0.0` for Docker
demo.launch(share=args.share)
================================================
FILE: cli.py
================================================
import argparse
import os
import random
import numpy as np
import soundfile as sf
import torch
from dia.model import Dia
def set_seed(seed: int):
"""Sets the random seed for reproducibility."""
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# Ensure deterministic behavior for cuDNN (if used)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def main():
parser = argparse.ArgumentParser(description="Generate audio using the Dia model.")
parser.add_argument("text", type=str, help="Input text for speech generation.")
parser.add_argument(
"--output", type=str, required=True, help="Path to save the generated audio file (e.g., output.wav)."
)
parser.add_argument(
"--repo-id",
type=str,
default="nari-labs/Dia-1.6B-0626",
help="Hugging Face repository ID (e.g., nari-labs/Dia-1.6B-0626).",
)
parser.add_argument(
"--local-paths", action="store_true", help="Load model from local config and checkpoint files."
)
parser.add_argument(
"--config", type=str, help="Path to local config.json file (required if --local-paths is set)."
)
parser.add_argument(
"--checkpoint", type=str, help="Path to local model checkpoint .pth file (required if --local-paths is set)."
)
parser.add_argument(
"--audio-prompt", type=str, default=None, help="Path to an optional audio prompt WAV file for voice cloning."
)
gen_group = parser.add_argument_group("Generation Parameters")
gen_group.add_argument(
"--max-tokens",
type=int,
default=None,
help="Maximum number of audio tokens to generate (defaults to config value).",
)
gen_group.add_argument(
"--cfg-scale", type=float, default=3.0, help="Classifier-Free Guidance scale (default: 3.0)."
)
gen_group.add_argument(
"--temperature", type=float, default=1.3, help="Sampling temperature (higher is more random, default: 0.7)."
)
gen_group.add_argument("--top-p", type=float, default=0.95, help="Nucleus sampling probability (default: 0.95).")
infra_group = parser.add_argument_group("Infrastructure")
infra_group.add_argument("--seed", type=int, default=None, help="Random seed for reproducibility.")
infra_group.add_argument(
"--device",
type=str,
default="cuda" if torch.cuda.is_available() else "cpu",
help="Device to run inference on (e.g., 'cuda', 'cpu', default: auto).",
)
args = parser.parse_args()
# Validation for local paths
if args.local_paths:
if not args.config:
parser.error("--config is required when --local-paths is set.")
if not args.checkpoint:
parser.error("--checkpoint is required when --local-paths is set.")
if not os.path.exists(args.config):
parser.error(f"Config file not found: {args.config}")
if not os.path.exists(args.checkpoint):
parser.error(f"Checkpoint file not found: {args.checkpoint}")
# Set seed if provided
if args.seed is not None:
set_seed(args.seed)
print(f"Using user-selected seed: {args.seed}")
# Determine device
device = torch.device(args.device)
print(f"Using device: {device}")
# Load model
print("Loading model...")
if args.local_paths:
print(f"Loading from local paths: config='{args.config}', checkpoint='{args.checkpoint}'")
try:
model = Dia.from_local(args.config, args.checkpoint, device=device)
except Exception as e:
print(f"Error loading local model: {e}")
exit(1)
else:
print(f"Loading from Hugging Face Hub: repo_id='{args.repo_id}'")
try:
model = Dia.from_pretrained(args.repo_id, device=device)
except Exception as e:
print(f"Error loading model from Hub: {e}")
exit(1)
print("Model loaded.")
# Generate audio
print("Generating audio...")
try:
sample_rate = 44100 # Default assumption
output_audio = model.generate(
text=args.text,
audio_prompt=args.audio_prompt,
max_tokens=args.max_tokens,
cfg_scale=args.cfg_scale,
temperature=args.temperature,
top_p=args.top_p,
)
print("Audio generation complete.")
print(f"Saving audio to {args.output}...")
os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
sf.write(args.output, output_audio, sample_rate)
print(f"Audio successfully saved to {args.output}")
except Exception as e:
print(f"Error during audio generation or saving: {e}")
exit(1)
if __name__ == "__main__":
main()
================================================
FILE: dia/__init__.py
================================================
from .model import Dia
__all__ = [
"Dia",
]
================================================
FILE: dia/audio.py
================================================
import typing as tp
import torch
def build_delay_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""
Precompute (t_idx_BxTxC, indices_BTCx3) so that out[t, c] = in[t - delay[c], c].
Negative t_idx => BOS; t_idx >= T => PAD.
"""
delay_arr = torch.tensor(delay_pattern, dtype=torch.int32)
t_idx_BxT = torch.broadcast_to(
torch.arange(T, dtype=torch.int32)[None, :],
[B, T],
)
t_idx_BxTx1 = t_idx_BxT[..., None]
t_idx_BxTxC = t_idx_BxTx1 - delay_arr.view(1, 1, C)
b_idx_BxTxC = torch.broadcast_to(
torch.arange(B, dtype=torch.int32).view(B, 1, 1),
[B, T, C],
)
c_idx_BxTxC = torch.broadcast_to(
torch.arange(C, dtype=torch.int32).view(1, 1, C),
[B, T, C],
)
# We must clamp time indices to [0..T-1] so gather_nd equivalent won't fail
t_clamped_BxTxC = torch.clamp(t_idx_BxTxC, 0, T - 1)
indices_BTCx3 = torch.stack(
[
b_idx_BxTxC.reshape(-1),
t_clamped_BxTxC.reshape(-1),
c_idx_BxTxC.reshape(-1),
],
dim=1,
).long() # Ensure indices are long type for indexing
return t_idx_BxTxC, indices_BTCx3
def apply_audio_delay(
audio_BxTxC: torch.Tensor,
pad_value: int,
bos_value: int,
precomp: tp.Tuple[torch.Tensor, torch.Tensor],
) -> torch.Tensor:
"""
Applies the delay pattern to batched audio tokens using precomputed indices,
inserting BOS where t_idx < 0 and PAD where t_idx >= T.
Args:
audio_BxTxC: [B, T, C] int16 audio tokens (or int32/float)
pad_value: the padding token
bos_value: the BOS token
precomp: (t_idx_BxTxC, indices_BTCx3) from build_delay_indices
Returns:
result_BxTxC: [B, T, C] delayed audio tokens
"""
device = audio_BxTxC.device # Get device from input tensor
t_idx_BxTxC, indices_BTCx3 = precomp
t_idx_BxTxC = t_idx_BxTxC.to(device) # Move precomputed indices to device
indices_BTCx3 = indices_BTCx3.to(device)
# Equivalent of tf.gather_nd using advanced indexing
# Ensure indices are long type if not already (build_delay_indices should handle this)
gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
gathered_BxTxC = gathered_flat.view(audio_BxTxC.shape)
# Create masks on the correct device
mask_bos = t_idx_BxTxC < 0 # => place bos_value
mask_pad = t_idx_BxTxC >= audio_BxTxC.shape[1] # => place pad_value
# Create scalar tensors on the correct device
bos_tensor = torch.tensor(bos_value, dtype=audio_BxTxC.dtype, device=device)
pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
# If mask_bos, BOS; else if mask_pad, PAD; else original gather
# All tensors should now be on the same device
result_BxTxC = torch.where(mask_bos, bos_tensor, torch.where(mask_pad, pad_tensor, gathered_BxTxC))
return result_BxTxC
def build_revert_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""
Precompute indices for the revert operation using PyTorch.
Returns:
A tuple (t_idx_BxTxC, indices_BTCx3) where:
- t_idx_BxTxC is a tensor of shape [B, T, C] computed as time indices plus the delay.
- indices_BTCx3 is a tensor of shape [B*T*C, 3] used for gathering, computed from:
batch indices, clamped time indices, and channel indices.
"""
# Use default device unless specified otherwise; assumes inputs might define device later
device = None # Or determine dynamically if needed, e.g., from a model parameter
delay_arr = torch.tensor(delay_pattern, dtype=torch.int32, device=device)
t_idx_BT1 = torch.broadcast_to(torch.arange(T, device=device).unsqueeze(0), [B, T])
t_idx_BT1 = t_idx_BT1.unsqueeze(-1)
t_idx_BxTxC = torch.minimum(
t_idx_BT1 + delay_arr.view(1, 1, C),
torch.tensor(T - 1, device=device),
)
b_idx_BxTxC = torch.broadcast_to(torch.arange(B, device=device).view(B, 1, 1), [B, T, C])
c_idx_BxTxC = torch.broadcast_to(torch.arange(C, device=device).view(1, 1, C), [B, T, C])
indices_BTCx3 = torch.stack(
[
b_idx_BxTxC.reshape(-1),
t_idx_BxTxC.reshape(-1),
c_idx_BxTxC.reshape(-1),
],
axis=1,
).long() # Ensure indices are long type
return t_idx_BxTxC, indices_BTCx3
def revert_audio_delay(
audio_BxTxC: torch.Tensor,
pad_value: int,
precomp: tp.Tuple[torch.Tensor, torch.Tensor],
T: int,
) -> torch.Tensor:
"""
Reverts a delay pattern from batched audio tokens using precomputed indices (PyTorch version).
Args:
audio_BxTxC: Input delayed audio tensor
pad_value: Padding value for out-of-bounds indices
precomp: Precomputed revert indices tuple containing:
- t_idx_BxTxC: Time offset indices tensor
- indices_BTCx3: Gather indices tensor for original audio
T: Original sequence length before padding
Returns:
Reverted audio tensor with same shape as input
"""
t_idx_BxTxC, indices_BTCx3 = precomp
device = audio_BxTxC.device # Get device from input tensor
# Move precomputed indices to the same device as audio_BxTxC if they aren't already
t_idx_BxTxC = t_idx_BxTxC.to(device)
indices_BTCx3 = indices_BTCx3.to(device)
# Using PyTorch advanced indexing (equivalent to tf.gather_nd or np equivalent)
gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
gathered_BxTxC = gathered_flat.view(audio_BxTxC.size()) # Use .size() for robust reshaping
# Create pad_tensor on the correct device
pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
# Create T tensor on the correct device for comparison
T_tensor = torch.tensor(T, device=device)
result_BxTxC = torch.where(t_idx_BxTxC >= T_tensor, pad_tensor, gathered_BxTxC) # Changed np.where to torch.where
return result_BxTxC
================================================
FILE: dia/config.py
================================================
"""Configuration management module for the Dia model.
This module provides comprehensive configuration management for the Dia model,
utilizing Pydantic for validation. It defines configurations for data processing,
model architecture (encoder and decoder), and training settings.
Key components:
- DataConfig: Parameters for data loading and preprocessing.
- EncoderConfig: Architecture details for the encoder module.
- DecoderConfig: Architecture details for the decoder module.
- ModelConfig: Combined model architecture settings.
- TrainingConfig: Training hyperparameters and settings.
- DiaConfig: Master configuration combining all components.
"""
import os
from pydantic import BaseModel, Field
class EncoderConfig(BaseModel, frozen=True):
"""Configuration for the encoder component of the Dia model.
Attributes:
model_type: Type of the model, defaults to "dia_encoder".
hidden_size: Size of the encoder layers, defaults to 1024.
intermediate_size: Size of the "intermediate" (i.e., feed-forward) layer in the encoder, defaults to 4096.
num_hidden_layers: Number of hidden layers in the encoder, defaults to 12.
num_attention_heads: Number of attention heads in the encoder, defaults to 16.
num_key_value_heads: Number of key-value heads in the encoder, defaults to 16.
head_dim: Dimension of each attention head, defaults to 128.
hidden_act: Activation function in the encoder, defaults to "silu".
max_position_embeddings: Maximum number of position embeddings, defaults to 1024.
initializer_range: Range for initializing weights, defaults to 0.02.
norm_eps: Epsilon value for normalization layers, defaults to 1e-5.
rope_theta: Theta value for RoPE, defaults to 10000.0.
rope_scaling: Optional scaling factor for RoPE.
vocab_size: Vocabulary size, defaults to 256.
"""
head_dim: int = Field(default=128, gt=0)
hidden_act: str = Field(default="silu")
hidden_size: int = Field(default=1024, gt=0)
initializer_range: float = Field(default=0.02)
intermediate_size: int = Field(default=4096, gt=0)
max_position_embeddings: int = Field(default=1024, gt=0)
model_type: str = Field(default="dia_encoder")
norm_eps: float = Field(default=1e-5)
num_attention_heads: int = Field(default=16, gt=0)
num_hidden_layers: int = Field(default=12, gt=0)
num_key_value_heads: int = Field(default=16, gt=0)
rope_scaling: float | None = Field(default=None)
rope_theta: float = Field(default=10000.0)
vocab_size: int = Field(default=256, gt=0)
class DecoderConfig(BaseModel, frozen=True):
"""Configuration for the decoder component of the Dia model.
Attributes:
model_type: Type of the model, defaults to "dia_decoder".
hidden_size: Size of the decoder layers, defaults to 2048.
intermediate_size: Size of the "intermediate" (i.e., feed-forward) layer in the decoder, defaults to 8192.
num_hidden_layers: Number of hidden layers in the decoder, defaults to 18.
num_attention_heads: Number of attention heads in the decoder, defaults to 16.
num_key_value_heads: Number of key-value heads in the decoder, defaults to 4.
head_dim: Dimension of each attention head, defaults to 128.
cross_hidden_size: Size of the cross-attention layers, defaults to 1024.
cross_num_attention_heads: Number of attention heads in the cross-attention mechanism, defaults to 16.
cross_num_key_value_heads: Number of key-value heads in the cross-attention mechanism, defaults to 16.
cross_head_dim: Dimension of each cross-attention head, defaults to 128.
hidden_act: Activation function in the decoder, defaults to "silu".
max_position_embeddings: Maximum number of position embeddings in the decoder, defaults to 3072.
initializer_range: Range for initializing weights in the decoder, defaults to 0.02.
norm_eps: Epsilon value for normalization layers in the decoder, defaults to 1e-5.
rope_theta: Theta value for RoPE in the decoder, defaults to 10000.0.
rope_scaling: Optional scaling factor for RoPE in the decoder.
vocab_size: Vocabulary size for the decoder, defaults to 1028.
num_channels: Number of channels in the decoder, defaults to 9.
"""
cross_head_dim: int = Field(default=128, gt=0)
cross_hidden_size: int = Field(default=1024, gt=0)
cross_num_attention_heads: int = Field(default=16, gt=0)
cross_num_key_value_heads: int = Field(default=16, gt=0)
head_dim: int = Field(default=128, gt=0)
hidden_act: str = Field(default="silu")
hidden_size: int = Field(default=2048, gt=0)
initializer_range: float = Field(default=0.02)
intermediate_size: int = Field(default=8192, gt=0)
max_position_embeddings: int = Field(default=3072, gt=0)
model_type: str = Field(default="dia_decoder")
norm_eps: float = Field(default=1e-5)
num_attention_heads: int = Field(default=16, gt=0)
num_channels: int = Field(default=9, gt=0)
num_hidden_layers: int = Field(default=18, gt=0)
num_key_value_heads: int = Field(default=4, gt=0)
rope_scaling: float | None = Field(default=None)
rope_theta: float = Field(default=10000.0)
vocab_size: int = Field(default=1028, gt=0)
class DiaConfig(BaseModel, frozen=True):
"""Main configuration container for the Dia model architecture.
Attributes:
model_type: Type of the model, defaults to "dia".
is_encoder_decoder: Flag indicating if the model is an encoder-decoder type, defaults to True.
encoder: Configuration for the encoder component.
decoder: Configuration for the decoder component.
src_vocab_size: Size of the source (text) vocabulary.
tgt_vocab_size: Size of the target (audio code) vocabulary.
initializer_range: Range for initializing weights, defaults to 0.02.
norm_eps: Epsilon value for normalization layers, defaults to 1e-5.
torch_dtype: Data type for model weights in PyTorch, defaults to "float32".
bos_token_id: Beginning-of-sequence token ID, defaults to 1026.
eos_token_id: End-of-sequence token ID, defaults to 1024.
pad_token_id: Padding token ID, defaults to 1025.
rope_theta: Theta value for RoPE, defaults to 10000.0.
rope_scaling: Optional scaling factor for RoPE.
transformers_version: Version of the transformers library, defaults to "4.53.0.dev0".
architectures: List of model architectures, defaults to ["DiaForConditionalGeneration"].
delay_pattern: List of delay values for each audio channel, defaults to [0,8,9,10,11,12,13,14,15].
"""
architectures: list[str] = Field(default_factory=lambda: ["DiaForConditionalGeneration"])
bos_token_id: int = Field(default=1026)
decoder_config: DecoderConfig
delay_pattern: list[int] = Field(default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15])
encoder_config: EncoderConfig
eos_token_id: int = Field(default=1024)
initializer_range: float = Field(default=0.02)
is_encoder_decoder: bool = Field(default=True)
model_type: str = Field(default="dia")
norm_eps: float = Field(default=1e-5)
pad_token_id: int = Field(default=1025)
torch_dtype: str = Field(default="float32")
transformers_version: str = Field(default="4.53.0.dev0")
def save(self, path: str) -> None:
"""Save the current configuration instance to a JSON file.
Ensures the parent directory exists and the file has a .json extension.
Args:
path: The target file path to save the configuration.
Raises:
ValueError: If the path is not a file with a .json extension.
"""
os.makedirs(os.path.dirname(path), exist_ok=True)
config_json = self.model_dump_json(indent=2)
with open(path, "w") as f:
f.write(config_json)
@classmethod
def load(cls, path: str) -> "DiaConfig | None":
"""Load and validate a Dia configuration from a JSON file.
Args:
path: The path to the configuration file.
Returns:
A validated DiaConfig instance if the file exists and is valid,
otherwise None if the file is not found.
Raises:
ValueError: If the path does not point to an existing .json file.
pydantic.ValidationError: If the JSON content fails validation against the DiaConfig schema.
"""
try:
with open(path, "r") as f:
content = f.read()
return cls.model_validate_json(content)
except FileNotFoundError:
return None
================================================
FILE: dia/layers.py
================================================
import torch
import torch.nn as nn
import torch.nn.functional as F
from huggingface_hub import PyTorchModelHubMixin
from torch import Tensor
from torch.nn import RMSNorm
from .config import DecoderConfig, DiaConfig, EncoderConfig
from .state import DecoderInferenceState, EncoderInferenceState, KVCache
def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
return tuple(ax if ax >= 0 else ndim + ax for ax in axes)
class DenseGeneral(nn.Module):
"""
PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
for the generalized matrix multiplication. Weight/bias shapes are calculated
and parameters created during initialization based on config.
`load_weights` validates shapes and copies data.
Attributes:
axis (Tuple[int, ...]): Input axis or axes to contract.
in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
out_features (Tuple[int, ...]): Shape of the output features (non-contracted dims).
use_bias (bool): Whether to add a bias term.
weight (nn.Parameter): The kernel parameter.
bias (Optional[nn.Parameter]): The bias parameter (if use_bias=True).
"""
def __init__(
self,
in_shapes: tuple[int, ...],
out_features: tuple[int, ...],
axis: tuple[int, ...] = (-1,),
weight_dtype: torch.dtype | None = None,
device: torch.device | None = None,
):
super().__init__()
self.in_shapes = in_shapes
self.out_features = out_features
self.axis = axis
self.kernel_shape = self.in_shapes + self.out_features
factory_kwargs = {"device": device, "dtype": weight_dtype}
self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
def forward(self, inputs: Tensor) -> Tensor:
norm_axis = _normalize_axes(self.axis, inputs.ndim)
kernel_contract_axes = tuple(range(len(norm_axis)))
output = torch.tensordot(
inputs.to(self.weight.dtype),
self.weight,
dims=(norm_axis, kernel_contract_axes),
).to(inputs.dtype)
return output
class MlpBlock(nn.Module):
"""MLP block using DenseGeneral."""
def __init__(self, embed_dim: int, intermediate_dim: int, compute_dtype: torch.dtype):
super().__init__()
self.dtype = compute_dtype
self.wi_fused = DenseGeneral(
in_shapes=(embed_dim,),
out_features=(2, intermediate_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.wo = DenseGeneral(
in_shapes=(intermediate_dim,),
out_features=(embed_dim,),
axis=(-1,),
weight_dtype=compute_dtype,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Forward pass."""
fused_x = self.wi_fused(x)
gate = fused_x[..., 0, :]
up = fused_x[..., 1, :]
hidden = torch.mul(F.silu(gate), up).to(self.dtype)
output = self.wo(hidden)
return output
class RotaryEmbedding(nn.Module):
"""Rotary Position Embedding (RoPE) implementation in PyTorch."""
def __init__(
self,
embedding_dims: int,
min_timescale: float = 1.0,
max_timescale: float = 10000.0,
dtype: torch.dtype = torch.float32,
):
super().__init__()
if embedding_dims % 2 != 0:
raise ValueError("Embedding dim must be even for RoPE.")
self.embedding_dims = embedding_dims
self.min_timescale = min_timescale
self.max_timescale = max_timescale
self.compute_dtype = dtype
half_embedding_dim = embedding_dims // 2
fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
timescale = (self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction).to(torch.float32)
self.register_buffer("timescale", timescale, persistent=False)
def forward(self, inputs: torch.Tensor, position: torch.Tensor):
"""Applies RoPE."""
position = position.unsqueeze(-1).unsqueeze(-1)
sinusoid_inp = position / self.timescale
sin = torch.sin(sinusoid_inp)
cos = torch.cos(sinusoid_inp)
first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
first_part = first_half * cos - second_half * sin
second_part = second_half * cos + first_half * sin
return torch.cat(
(first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)),
dim=-1,
)
def apply_rope(self, inputs: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor):
first_half, second_half = torch.chunk(inputs.to(torch.float32), 2, dim=-1)
first_part = first_half * cos - second_half * sin
second_part = second_half * cos + first_half * sin
return torch.cat((first_part.to(self.compute_dtype), second_part.to(self.compute_dtype)), dim=-1)
def custom_scaled_dot_product_attention(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
attn_mask: torch.Tensor | None = None,
scale: float = 1.0,
is_causal: bool = False,
num_gqa_groups: int = 1,
) -> torch.Tensor:
"""
Custom scaled dot-product attention with GQA support for MPS compatibility.
Args:
query: (B, N_q, T, H) - Query tensor, N_q = num_query_heads
key: (B, N_kv, S, H) - Key tensor, N_kv = num_kv_heads
value: (B, N_kv, S, H) - Value tensor
attn_mask: (B, 1, T, S) - Attention mask, optional
scale: Scaling factor for attention scores
is_causal: If True, apply causal masking
num_gqa_groups: Number of query groups per KV head (N_q / N_kv)
Returns:
output: (B, N_q, T, H) - Attention output
"""
B, N_q, T, H = query.shape
_, N_kv, S, _ = key.shape
# For GQA, repeat key and value tensors to match query heads
if num_gqa_groups > 1:
key = key.repeat_interleave(num_gqa_groups, dim=1) # (B, N_q, S, H)
value = value.repeat_interleave(num_gqa_groups, dim=1) # (B, N_q, S, H)
# Compute attention scores: (B, N_q, T, H) @ (B, N_q, H, S) -> (B, N_q, T, S)
scores = torch.matmul(query, key.transpose(-1, -2)) * scale
# Apply causal mask if needed
if is_causal:
causal_mask = torch.tril(torch.ones(T, S, dtype=torch.bool, device=query.device))
scores = scores.masked_fill(~causal_mask, float("-inf"))
# Apply attention mask if provided
if attn_mask is not None:
scores = scores.masked_fill(~attn_mask, float("-inf"))
# Softmax over the last dimension (S)
attn_weights = F.softmax(scores, dim=-1)
# Compute output: (B, N_q, T, S) @ (B, N_q, S, H) -> (B, N_q, T, H)
output = torch.matmul(attn_weights, value)
return output
class CrossAttention(nn.Module):
"""Cross-Attention using DenseGeneral."""
def __init__(
self,
config: EncoderConfig | DecoderConfig,
q_embed_dim: int,
kv_embed_dim: int,
num_query_heads: int,
num_kv_heads: int,
head_dim: int,
compute_dtype: torch.dtype,
out_embed_dim: int | None = None,
):
super().__init__()
self.num_query_heads = num_query_heads
self.num_kv_heads = num_kv_heads
self.head_dim = head_dim
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
self.projected_query_dim = num_query_heads * head_dim
if num_query_heads % num_kv_heads != 0:
raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
self.num_gqa_groups = num_query_heads // num_kv_heads
# --- Projection Layers using DenseGeneral ---
self.q_proj = DenseGeneral(
in_shapes=(q_embed_dim,),
out_features=(num_query_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.k_proj = DenseGeneral(
in_shapes=(kv_embed_dim,),
out_features=(num_kv_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.v_proj = DenseGeneral(
in_shapes=(kv_embed_dim,),
out_features=(num_kv_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.o_proj = DenseGeneral(
in_shapes=(num_query_heads, head_dim),
out_features=(self.output_dim,),
axis=(-2, -1),
weight_dtype=compute_dtype,
)
# --- Rotary Embedding ---
self.rotary_emb = RotaryEmbedding(
embedding_dims=self.head_dim,
max_timescale=config.rope_theta,
dtype=compute_dtype,
)
def forward(
self,
Xq: torch.Tensor, # (B, T, D) T = 1 in AR generation
q_positions: torch.Tensor, # (B, T)
kv_positions: torch.Tensor | None = None, # (B, S)
attn_mask: torch.Tensor | None = None, # None in Decoder Self Attention, Valid mask in Others
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
is_causal: bool = False,
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
"""
Performs attention calculation with optional KV caching.
Args:
Xq: Query tensor (B, T, D). T=1 during single-step decoding.
Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
q_positions: Positions for queries (B, T).
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
attn_mask: Attention mask.
cache: KVCache.
Returns:
A tuple containing:
- output: The attention output tensor (B, T, output_dim).
- present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
"""
if kv_positions is None:
kv_positions = q_positions
original_dtype = Xq.dtype
Xq_BxTxNxH = self.q_proj(Xq)
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
attn_k: torch.Tensor | None = cache.k if cache is not None else None
attn_v: torch.Tensor | None = cache.v if cache is not None else None
# Use custom attention for MPS backend, otherwise use optimized PyTorch function
is_mps = Xq.device.type == "mps" and torch.backends.mps.is_available()
if is_mps:
attn_output = custom_scaled_dot_product_attention(
query=Xq_BxNxTxH,
key=attn_k,
value=attn_v,
attn_mask=attn_mask if not is_causal else None,
scale=1.0,
is_causal=is_causal,
num_gqa_groups=self.num_gqa_groups,
)
else:
attn_output = F.scaled_dot_product_attention(
Xq_BxNxTxH,
attn_k,
attn_v,
attn_mask=attn_mask if not is_causal else None,
scale=1.0,
enable_gqa=self.num_gqa_groups > 1,
is_causal=is_causal,
)
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
output = self.o_proj(attn_output)
return output.to(original_dtype)
class FusedQKV(nn.Module):
def __init__(
self,
in_features: int,
out_features: int,
bias: bool = False,
num_q_heads: int = 1,
q_head_dim: int = 1,
num_kv_heads: int = 1,
kv_head_dim: int = 1,
):
super().__init__()
self.num_q_heads = num_q_heads
self.q_head_dim = q_head_dim
self.num_kv_heads = num_kv_heads
self.kv_head_dim = kv_head_dim
self.q_output_dim = num_q_heads * q_head_dim
self.kv_output_dim = num_kv_heads * kv_head_dim
self.linear = nn.Linear(in_features, out_features, bias=bias)
def forward(self, inputs: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
x = self.linear(inputs)
q, k, v = x.split([self.q_output_dim, self.kv_output_dim, self.kv_output_dim], dim=-1)
q = q.reshape(q.shape[:-1] + (self.num_q_heads, self.q_head_dim))
k = k.reshape(k.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))
v = v.reshape(v.shape[:-1] + (self.num_kv_heads, self.kv_head_dim))
return q, k, v
class SelfAttention(nn.Module):
"""Attention using DenseGeneral."""
def __init__(
self,
config: EncoderConfig | DecoderConfig,
q_embed_dim: int,
kv_embed_dim: int,
num_query_heads: int,
num_kv_heads: int,
head_dim: int,
compute_dtype: torch.dtype,
out_embed_dim: int | None = None,
):
super().__init__()
self.num_query_heads = num_query_heads
self.num_kv_heads = num_kv_heads
self.head_dim = head_dim
self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
self.projected_query_dim = num_query_heads * head_dim
if num_query_heads % num_kv_heads != 0:
raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
self.num_gqa_groups = num_query_heads // num_kv_heads
self.kv_embed_dim = kv_embed_dim
self.q_embed_dim = q_embed_dim
# --- Projection Layers using DenseGeneral ---
self.q_proj = DenseGeneral(
in_shapes=(q_embed_dim,),
out_features=(num_query_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.k_proj = DenseGeneral(
in_shapes=(kv_embed_dim,),
out_features=(num_kv_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.v_proj = DenseGeneral(
in_shapes=(kv_embed_dim,),
out_features=(num_kv_heads, head_dim),
axis=(-1,),
weight_dtype=compute_dtype,
)
self.o_proj = DenseGeneral(
in_shapes=(num_query_heads, head_dim),
out_features=(self.output_dim,),
axis=(-2, -1),
weight_dtype=compute_dtype,
)
# --- Rotary Embedding ---
self.rotary_emb = RotaryEmbedding(
embedding_dims=self.head_dim,
max_timescale=config.rope_theta,
dtype=compute_dtype,
)
self.is_fused_qkv = False
def get_linear_weight(self, dense: DenseGeneral):
W_dg = dense.weight.data
out_features = 1
input_features = 1
for dim in dense.out_features:
out_features *= dim
for dim in dense.in_shapes:
input_features *= dim
W_dg_reshaped_for_linear_T = W_dg.reshape(input_features, out_features)
linear_weight = W_dg_reshaped_for_linear_T.transpose(0, 1).contiguous()
return linear_weight
def patch_fused_qkv(self):
q_proj_weight = self.get_linear_weight(self.q_proj)
k_proj_weight = self.get_linear_weight(self.k_proj)
v_proj_weight = self.get_linear_weight(self.v_proj)
self.qkv = FusedQKV(
self.kv_embed_dim,
(self.num_query_heads * self.head_dim + 2 * (self.num_kv_heads * self.head_dim)),
bias=False,
num_q_heads=self.num_query_heads,
q_head_dim=self.head_dim,
num_kv_heads=self.num_kv_heads,
kv_head_dim=self.head_dim,
)
self.qkv.linear.weight.data = torch.cat([q_proj_weight, k_proj_weight, v_proj_weight], dim=0)
# print(f"qkv.weight.shape: {self.qkv.linear.weight.shape}")
self.is_fused_qkv = True
def forward(
self,
X: torch.Tensor, # (B, T, D) T = 1 in AR generation
q_positions: torch.Tensor, # (B, T)
kv_positions: torch.Tensor | None = None, # (B, S)
attn_mask: torch.Tensor | None = None, # None in Decoder Self Attention, Valid mask in Others
cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
prefill: bool = False,
is_causal: bool = False,
current_idx: torch.Tensor | None = None,
) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
"""
Performs attention calculation with optional KV caching.
Args:
Xq: Query tensor (B, T, D). T=1 during single-step decoding.
Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
q_positions: Positions for queries (B, T).
kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
attn_mask: Attention mask.
cache: KVCache.
prefill: If True, use prefill mode.
Returns:
A tuple containing:
- output: The attention output tensor (B, T, output_dim).
- present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
"""
if kv_positions is None:
kv_positions = q_positions
original_dtype = X.dtype
if self.is_fused_qkv:
Xq_BxTxNxH, Xk_BxSxKxH, Xv_BxSxKxH = self.qkv(X)
else:
Xq_BxTxNxH = self.q_proj(X)
Xk_BxSxKxH = self.k_proj(X)
Xv_BxSxKxH = self.v_proj(X)
position = q_positions.unsqueeze(-1).unsqueeze(-1)
sinusoid_inp = position / self.rotary_emb.timescale
sin = torch.sin(sinusoid_inp)
cos = torch.cos(sinusoid_inp)
Xq_BxTxNxH = self.rotary_emb.apply_rope(Xq_BxTxNxH, sin, cos)
Xk_BxSxKxH = self.rotary_emb.apply_rope(Xk_BxSxKxH, sin, cos)
Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
attn_k: torch.Tensor | None = cache.k if cache is not None else None
attn_v: torch.Tensor | None = cache.v if cache is not None else None
Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2) # (B, K, S, H)
Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2) # (B, K, S, H)
if cache is None:
attn_k = Xk_BxKxSxH
attn_v = Xv_BxKxSxH
elif prefill:
attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
cache.prefill(attn_k, attn_v)
else:
attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH, current_idx)
# Use custom attention for MPS backend, otherwise use optimized PyTorch function
is_mps = Xv_BxSxKxH.device.type == "mps" and torch.backends.mps.is_available()
if is_mps:
attn_output = custom_scaled_dot_product_attention(
query=Xq_BxNxTxH,
key=attn_k,
value=attn_v,
attn_mask=attn_mask if not is_causal else None,
scale=1.0,
is_causal=is_causal,
num_gqa_groups=self.num_gqa_groups,
)
else:
attn_output = F.scaled_dot_product_attention(
Xq_BxNxTxH,
attn_k,
attn_v,
attn_mask=attn_mask if not is_causal else None,
scale=1.0,
enable_gqa=self.num_gqa_groups > 1,
is_causal=is_causal,
)
attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
output = self.o_proj(attn_output)
return output.to(original_dtype)
class EncoderLayer(nn.Module):
"""Transformer Encoder Layer using DenseGeneral."""
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
super().__init__()
self.config = config
enc_config = config.encoder_config
embed_dim = enc_config.hidden_size
self.compute_dtype = compute_dtype
self.pre_sa_norm = RMSNorm(
embed_dim,
eps=enc_config.norm_eps,
dtype=torch.float32,
)
self.self_attention = SelfAttention(
enc_config,
q_embed_dim=embed_dim,
kv_embed_dim=embed_dim,
num_query_heads=enc_config.num_attention_heads,
num_kv_heads=enc_config.num_key_value_heads,
head_dim=enc_config.head_dim,
compute_dtype=compute_dtype,
out_embed_dim=embed_dim,
)
self.post_sa_norm = RMSNorm(
embed_dim,
eps=enc_config.norm_eps,
dtype=torch.float32,
)
self.mlp = MlpBlock(
embed_dim=embed_dim,
intermediate_dim=enc_config.intermediate_size,
compute_dtype=compute_dtype,
)
def forward(
self,
x: torch.Tensor,
state: EncoderInferenceState,
) -> torch.Tensor:
residual = x
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
sa_out = self.self_attention(
X=x_norm,
q_positions=state.positions,
kv_positions=state.positions,
attn_mask=state.attn_mask,
)
x = residual + sa_out
residual = x
x_norm = self.post_sa_norm(x).to(self.compute_dtype)
mlp_out = self.mlp(x_norm)
x = residual + mlp_out
return x
class Encoder(nn.Module):
"""Transformer Encoder Stack using DenseGeneral."""
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
super().__init__()
self.config = config
enc_config = config.encoder_config
self.compute_dtype = compute_dtype
self.embedding = nn.Embedding(
enc_config.vocab_size,
enc_config.hidden_size,
dtype=compute_dtype,
)
self.layers = nn.ModuleList([EncoderLayer(config, compute_dtype) for _ in range(enc_config.num_hidden_layers)])
self.norm = RMSNorm(
enc_config.hidden_size,
eps=enc_config.norm_eps,
dtype=torch.float32,
)
def forward(
self,
x_ids: torch.Tensor,
state: EncoderInferenceState,
) -> torch.Tensor:
x = self.embedding(x_ids)
for layer in self.layers:
x = layer(x, state)
x = self.norm(x).to(self.compute_dtype)
return x
class DecoderLayer(nn.Module):
"""Transformer Decoder Layer using DenseGeneral."""
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
super().__init__()
self.config = config
dec_config = config.decoder_config
enc_config = config.encoder_config
dec_embed_dim = dec_config.hidden_size
enc_embed_dim = enc_config.hidden_size
self.compute_dtype = compute_dtype
# Norms
self.pre_sa_norm = RMSNorm(
dec_embed_dim,
eps=dec_config.norm_eps,
dtype=torch.float32,
)
self.pre_ca_norm = RMSNorm(
dec_embed_dim,
eps=dec_config.norm_eps,
dtype=torch.float32,
)
self.pre_mlp_norm = RMSNorm(
dec_embed_dim,
eps=dec_config.norm_eps,
dtype=torch.float32,
)
# Self-Attention (GQA) with Causal Masking
self.self_attention = SelfAttention(
dec_config,
q_embed_dim=dec_embed_dim,
kv_embed_dim=dec_embed_dim,
num_query_heads=dec_config.num_attention_heads,
num_kv_heads=dec_config.num_key_value_heads,
head_dim=dec_config.head_dim,
compute_dtype=compute_dtype,
out_embed_dim=dec_embed_dim,
)
# Cross-Attention (MHA)
self.cross_attention = CrossAttention(
dec_config,
q_embed_dim=dec_embed_dim,
kv_embed_dim=enc_embed_dim, # Note kv_embed_dim
num_query_heads=dec_config.cross_num_attention_heads,
num_kv_heads=dec_config.cross_num_key_value_heads,
head_dim=dec_config.cross_head_dim,
compute_dtype=compute_dtype,
out_embed_dim=dec_embed_dim,
)
# MLP
self.mlp = MlpBlock(
embed_dim=dec_embed_dim,
intermediate_dim=dec_config.intermediate_size,
compute_dtype=compute_dtype,
)
def forward(
self,
x: torch.Tensor,
state: DecoderInferenceState,
self_attn_cache: KVCache | None = None,
cross_attn_cache: KVCache | None = None,
prefill: bool = False,
current_idx: int = 0,
) -> torch.Tensor:
residual = x
x_norm = self.pre_sa_norm(x).to(self.compute_dtype)
self_attn_mask = state.casual_attn_mask[None, None, current_idx]
sa_out = self.self_attention(
X=x_norm, # (2, 1, D)
q_positions=state.dec_positions, # (2, 1)
kv_positions=state.dec_positions, # (2, 1)
attn_mask=self_attn_mask,
cache=self_attn_cache,
prefill=prefill,
is_causal=prefill,
current_idx=current_idx,
)
x = residual + sa_out
residual = x
x_norm = self.pre_ca_norm(x).to(self.compute_dtype)
ca_out = self.cross_attention(
Xq=x_norm,
q_positions=state.dec_positions,
kv_positions=state.enc_positions,
attn_mask=state.cross_attn_mask,
cache=cross_attn_cache,
)
x = residual + ca_out
residual = x
x_norm = self.pre_mlp_norm(x).to(self.compute_dtype)
mlp_out = self.mlp(x_norm)
x = residual + mlp_out
return x
class Decoder(nn.Module):
"""Transformer Decoder Stack using DenseGeneral."""
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
super().__init__()
self.config = config
dec_config = config.decoder_config
self.num_channels = dec_config.num_channels
self.num_layers = dec_config.num_hidden_layers
self.embeddings = nn.ModuleList(
[
nn.Embedding(dec_config.vocab_size, dec_config.hidden_size, dtype=compute_dtype)
for _ in range(self.num_channels)
]
)
self.layers = nn.ModuleList(
[DecoderLayer(config=config, compute_dtype=compute_dtype) for _ in range(self.num_layers)]
)
self.norm = RMSNorm(
dec_config.hidden_size,
eps=dec_config.norm_eps,
dtype=torch.float32,
)
self.logits_dense = DenseGeneral(
in_shapes=(dec_config.hidden_size,),
out_features=(self.num_channels, dec_config.vocab_size),
axis=(-1,),
weight_dtype=compute_dtype,
)
def precompute_cross_attn_cache(
self,
enc_out: torch.Tensor, # (B, S, E)
) -> list[KVCache]:
"""
Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
"""
per_layer_kv_cache: list[KVCache] = []
for layer in self.layers:
cross_attn_module = layer.cross_attention
k_proj = cross_attn_module.k_proj(enc_out)
v_proj = cross_attn_module.v_proj(enc_out)
k = k_proj.transpose(1, 2)
v = v_proj.transpose(1, 2)
per_layer_kv_cache.append(KVCache.from_kv(k, v))
return per_layer_kv_cache
def decode_step(
self,
tgt_ids_Bx1xC: torch.Tensor, # [B, 1, C]
state: DecoderInferenceState,
current_idx: int,
) -> torch.Tensor:
"""
Performs a single decoding step, managing KV caches layer by layer.
Returns:
A tuple containing:
- logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
"""
x = None
for i in range(self.num_channels):
channel_tokens = tgt_ids_Bx1xC[..., i]
channel_embed = self.embeddings[i](channel_tokens)
x = channel_embed if x is None else x + channel_embed
for i, layer in enumerate(self.layers):
self_cache = state.self_attn_cache[i]
cross_cache = state.cross_attn_cache[i]
x = layer(
x, # (2, 1, D)
state,
self_attn_cache=self_cache,
cross_attn_cache=cross_cache,
current_idx=current_idx,
)
x = self.norm(x)
logits_Bx1xCxV = self.logits_dense(x)
return logits_Bx1xCxV.to(torch.float32)
def forward(self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInferenceState) -> torch.Tensor:
"""
Forward pass for the Decoder stack, managing KV caches.
Args:
tgt_ids_BxTxC: Target token IDs (B, T, C).
encoder_out: Output from the encoder (B, S, E).
tgt_positions: Positions for target sequence (B, T).
src_positions: Positions for source sequence (B, S).
self_attn_mask: Mask for self-attention.
cross_attn_mask: Mask for cross-attention.
past_key_values: List containing the self-attention KV cache for each layer
from the previous decoding step. `len(past_key_values)` should
equal `num_layers`.
precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
derived from `encoder_out`. This is passed identically
to all layers.
Returns:
A tuple containing:
- logits: The final output logits (B, T, C * V), cast to float32.
- present_key_values: A list containing the updated self-attention KV cache
for each layer for the *current* decoding step.
"""
_, _, num_channels_in = tgt_ids_BxTxC.shape
assert num_channels_in == self.num_channels, "Input channels mismatch"
# Embeddings
x = None
for i in range(self.num_channels):
channel_tokens = tgt_ids_BxTxC[..., i]
channel_embed = self.embeddings[i](channel_tokens)
x = channel_embed if x is None else x + channel_embed
for i, layer in enumerate(self.layers):
self_cache = state.self_attn_cache[i]
cross_cache = state.cross_attn_cache[i]
x = layer(
x,
state,
self_attn_cache=self_cache,
cross_attn_cache=cross_cache,
prefill=True,
)
# Final Norm
x = self.norm(x)
logits_BxTxCxV = self.logits_dense(x)
return logits_BxTxCxV.to(torch.float32)
class DiaModel(
nn.Module,
PyTorchModelHubMixin,
repo_url="https://github.com/nari-labs/dia",
pipeline_tag="text-to-speech",
license="apache-2.0",
coders={
DiaConfig: (
lambda x: x.model_dump(),
lambda data: DiaConfig.model_validate(data),
),
},
):
"""PyTorch Dia Model using DenseGeneral."""
def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
super().__init__()
self.config = config
self.encoder = Encoder(config, compute_dtype)
self.decoder = Decoder(config, compute_dtype)
================================================
FILE: dia/model.py
================================================
import time
from enum import Enum
from typing import Callable
import numpy as np
import torch
import torch.nn.functional as F
import torchaudio
from .audio import apply_audio_delay, build_delay_indices, build_revert_indices, revert_audio_delay
from .config import DiaConfig
from .layers import DiaModel
from .state import DecoderInferenceState, DecoderOutput, EncoderInferenceState
DEFAULT_SAMPLE_RATE = 44100
SAMPLE_RATE_RATIO = 512
def _get_default_device():
if torch.cuda.is_available():
return torch.device("cuda")
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
return torch.device("mps")
return torch.device("cpu")
def _sample_next_token(
logits_BCxV: torch.Tensor,
temperature: float,
top_p: float,
top_k: int | None,
audio_eos_value: int,
) -> torch.Tensor:
if temperature == 0.0:
return torch.argmax(logits_BCxV, dim=-1)
logits_BCxV = logits_BCxV / temperature
if audio_eos_value is not None and audio_eos_value >= 0:
top_logit_indices_BC = torch.argmax(logits_BCxV, dim=-1)
eos_not_highest_mask_BC = top_logit_indices_BC != audio_eos_value
mask_eos_unless_highest_BCxV = torch.zeros_like(logits_BCxV, dtype=torch.bool)
mask_eos_unless_highest_BCxV[eos_not_highest_mask_BC, audio_eos_value] = True
logits_BCxV = logits_BCxV.masked_fill(mask_eos_unless_highest_BCxV, -torch.inf)
eos_highest_mask_BC = top_logit_indices_BC == audio_eos_value
mask_eos_highest_BCxV = torch.zeros_like(logits_BCxV, dtype=torch.bool)
mask_eos_highest_BCxV[eos_highest_mask_BC, :audio_eos_value] = True
logits_BCxV = logits_BCxV.masked_fill(mask_eos_highest_BCxV, -torch.inf)
if top_k is not None:
_, top_k_indices_BCxV = torch.topk(logits_BCxV, k=top_k, dim=-1)
mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
mask = mask.scatter(dim=-1, index=top_k_indices_BCxV, value=False)
logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
if top_p < 1.0:
probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
sorted_probs_BCxV, sorted_indices_BCxV = torch.sort(probs_BCxV, dim=-1, descending=True)
cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
sorted_indices_to_remove_BCxV = torch.roll(sorted_indices_to_remove_BCxV, shifts=1, dims=-1)
sorted_indices_to_remove_BCxV[..., 0] = torch.zeros_like(sorted_indices_to_remove_BCxV[..., 0])
indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
indices_to_remove_BCxV = indices_to_remove_BCxV.scatter(
dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV
)
logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
final_probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
sampled_indices_BC = torch.multinomial(final_probs_BCxV, num_samples=1)
sampled_indices_C = sampled_indices_BC.squeeze(-1)
return sampled_indices_C
class ComputeDtype(str, Enum):
FLOAT32 = "float32"
FLOAT16 = "float16"
BFLOAT16 = "bfloat16"
def to_dtype(self) -> torch.dtype:
if self == ComputeDtype.FLOAT32:
return torch.float32
elif self == ComputeDtype.FLOAT16:
return torch.float16
elif self == ComputeDtype.BFLOAT16:
return torch.bfloat16
else:
raise ValueError(f"Unsupported compute dtype: {self}")
class Dia:
def __init__(
self,
config: DiaConfig,
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
device: torch.device | None = None,
load_dac: bool = True,
):
"""Initializes the Dia model.
Args:
config: The configuration object for the model.
compute_dtype: The computation dtype to use.
device: The device to load the model onto. If None, will automatically select the best available device.
load_dac: Whether to load the DAC model.
Raises:
RuntimeError: If there is an error loading the DAC model.
"""
super().__init__()
self.config = config
self.device = device if device is not None else _get_default_device()
if isinstance(compute_dtype, str):
compute_dtype = ComputeDtype(compute_dtype)
self.compute_dtype = compute_dtype.to_dtype()
self.model: DiaModel = DiaModel(config, self.compute_dtype)
self.dac_model = None
self._compiled_step = None
self.load_dac = load_dac
if not self.load_dac:
print("Warning: DAC model will not be loaded. This is not recommended.")
if torch.cuda.is_available():
torch.backends.cuda.matmul.allow_tf32 = True
@classmethod
def from_local(
cls,
config_path: str,
checkpoint_path: str,
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
device: torch.device | None = None,
load_dac: bool = True,
) -> "Dia":
"""Loads the Dia model from local configuration and checkpoint files.
Args:
config_path: Path to the configuration JSON file.
checkpoint_path: Path to the model checkpoint (.pth) file.
compute_dtype: The computation dtype to use.
device: The device to load the model onto. If None, will automatically select the best available device.
load_dac: Whether to load the DAC model.
Returns:
An instance of the Dia model loaded with weights and set to eval mode.
Raises:
FileNotFoundError: If the config or checkpoint file is not found.
RuntimeError: If there is an error loading the checkpoint.
"""
config = DiaConfig.load(config_path)
if config is None:
raise FileNotFoundError(f"Config file not found at {config_path}")
dia = cls(config, compute_dtype, device, load_dac)
try:
state_dict = torch.load(checkpoint_path, map_location=dia.device)
dia.model.load_state_dict(state_dict)
except FileNotFoundError:
raise FileNotFoundError(f"Checkpoint file not found at {checkpoint_path}")
except Exception as e:
raise RuntimeError(f"Error loading checkpoint from {checkpoint_path}") from e
dia.model.to(dia.device)
dia.model.eval()
if load_dac:
dia._load_dac_model()
return dia
@classmethod
def from_pretrained(
cls,
model_name: str = "nari-labs/Dia-1.6B-0626",
compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
device: torch.device | None = None,
load_dac: bool = True,
) -> "Dia":
"""Loads the Dia model from a Hugging Face Hub repository.
Downloads the configuration and checkpoint files from the specified
repository ID and then loads the model.
Args:
model_name: The Hugging Face Hub repository ID (e.g., "nari-labs/Dia-1.6B-0626").
compute_dtype: The computation dtype to use.
device: The device to load the model onto. If None, will automatically select the best available device.
load_dac: Whether to load the DAC model.
Returns:
An instance of the Dia model loaded with weights and set to eval mode.
Raises:
FileNotFoundError: If config or checkpoint download/loading fails.
RuntimeError: If there is an error loading the checkpoint.
"""
if isinstance(compute_dtype, str):
compute_dtype = ComputeDtype(compute_dtype)
# Load model directly using DiaModel's from_pretrained which handles HF download
try:
loaded_model = DiaModel.from_pretrained(model_name, compute_dtype=compute_dtype.to_dtype())
except Exception as e:
raise RuntimeError(f"Error loading model from Hugging Face Hub ({model_name})") from e
config = loaded_model.config # Get config from the loaded model
dia = cls(config, compute_dtype, device, load_dac)
dia.model = loaded_model # Assign the already loaded model
dia.model.to(dia.device)
dia.model.eval()
if load_dac:
dia._load_dac_model()
return dia
def _load_dac_model(self):
"""Loads the Descript Audio Codec (DAC) model.
Downloads the DAC model if necessary and loads it onto the specified device.
Sets the DAC model to evaluation mode.
Raises:
RuntimeError: If downloading or loading the DAC model fails.
"""
import dac
try:
dac_model_path = dac.utils.download()
dac_model = dac.DAC.load(dac_model_path).to(self.device)
dac_model.eval() # Ensure DAC is in eval mode
except Exception as e:
raise RuntimeError("Failed to load DAC model") from e
self.dac_model = dac_model
def _encode_text(self, text: str) -> torch.Tensor:
"""Encodes the input text string into a tensor of token IDs using byte-level encoding.
Special tokens [S1] and [S2] are replaced by their byte values. The resulting
sequence is truncated to the maximum configured text length.
Args:
text: The input text string.
Returns:
A tensor containing the encoded byte token IDs.
"""
max_len = self.config.encoder_config.max_position_embeddings
byte_text = text.encode("utf-8")
# Replace special tokens with their byte values if needed by the specific tokenizer/config
# Assuming byte values 1 and 2 are correct placeholders based on original code
replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
text_tokens = list(replaced_bytes)
return torch.tensor(
text_tokens[:max_len],
dtype=torch.long,
device=self.device,
)
def _pad_text_input(self, text_tokens: list[torch.Tensor]) -> torch.Tensor:
"""Pads the text input to the maximum length."""
text_pad_value = 0
max_len = self.config.encoder_config.max_position_embeddings
batch_size = len(text_tokens)
src_tokens = torch.full(
(batch_size, 1, max_len),
fill_value=text_pad_value,
dtype=torch.long,
device=self.device,
)
for i in range(batch_size):
current_len = len(text_tokens[i])
src_tokens[i, 0, :current_len] = text_tokens[i]
return src_tokens
def _prepare_audio_prompt(self, audio_prompts: list[torch.Tensor | None]) -> tuple[torch.Tensor, list[int]]:
"""Prepares the audio prompt tensor for the decoder.
Handles padding, adds the beginning-of-sequence (BOS) token, applies the
delay pattern, and determines the number of prefill steps for each item
in the batch.
Args:
audio_prompts: A list of audio prompt tensors (encoded DAC frames) or None.
Each tensor should have shape [T, C].
Returns:
A tuple containing:
- delayed_batch (torch.Tensor): The prepared audio prompt tensor with
delays applied, shape [B, T_max_padded, C].
- prefill_steps (list[int]): A list containing the number of valid
tokens (including BOS) for each prompt in the batch.
"""
num_channels = self.config.decoder_config.num_channels
audio_bos_value = self.config.bos_token_id
delay_pattern = self.config.delay_pattern
max_delay_pattern = max(delay_pattern)
batch_size = len(audio_prompts)
max_len = max(p.shape[0] if p is not None else 0 for p in audio_prompts) + max_delay_pattern
prefill_steps = []
prefill = torch.full(
(batch_size, max_len, num_channels),
fill_value=-1,
dtype=torch.int,
device=self.device,
)
prefill[:, 0, :] = audio_bos_value
for i in range(batch_size):
prompt = audio_prompts[i]
if prompt is not None:
prompt = prompt.to(device=self.device, dtype=torch.int)
prefill[i, 1 : prompt.shape[0] + 1, :] = prompt
prefill_steps.append(prompt.shape[0] + 1)
else:
prefill_steps.append(1)
delay_precomp = build_delay_indices(
B=batch_size,
T=max_len,
C=num_channels,
delay_pattern=delay_pattern,
)
delayed_batch = apply_audio_delay(
audio_BxTxC=prefill,
pad_value=-1,
bos_value=audio_bos_value,
precomp=delay_precomp,
)
return delayed_batch, prefill_steps
def _prepare_generation(
self,
text: torch.Tensor,
audio_prompts: list[torch.Tensor | None],
max_tokens: int | None = None,
attn_fn: Callable = F.scaled_dot_product_attention,
):
"""Initializes the model state for generation.
Encodes the text input (conditional and unconditional), prepares the
encoder and decoder states (including KV caches and cross-attention),
prepares the audio prompt, and performs the initial decoder prefill steps
based on the audio prompts.
Args:
text: The padded text input tensor, shape [B, 1, T_text].
audio_prompts: A list of prepared audio prompt tensors or None.
Returns:
A tuple containing:
- dec_state (DecoderInferenceState): The initialized decoder state.
- dec_output (DecoderOutput): The initialized decoder output manager,
containing the prefilled audio tokens.
"""
batch_size = text.shape[0]
enc_input_uncond = torch.zeros_like(text)
enc_input_cond = text
stacked_inputs = torch.stack([enc_input_uncond, enc_input_cond], dim=1)
enc_input = stacked_inputs.view(2 * batch_size, -1)
enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
encoder_out = self.model.encoder(enc_input, enc_state)
dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(encoder_out)
dec_state = DecoderInferenceState.new(
self.config,
enc_state,
encoder_out,
dec_cross_attn_cache,
self.compute_dtype,
max_generation_length=max_tokens,
)
prefill, prefill_steps = self._prepare_audio_prompt(audio_prompts)
dec_output = DecoderOutput.new(batch_size, self.config, self.device)
dec_output.prefill(prefill, prefill_steps)
dec_step = min(prefill_steps) - 1
if dec_step > 0:
dec_state.prepare_step(0, dec_step)
tokens_BxTxC = dec_output.get_tokens_at(0, dec_step).repeat_interleave(2, dim=0)
self.model.decoder.forward(tokens_BxTxC, dec_state)
return dec_state, dec_output
def _decoder_step(
self,
tokens_Bx1xC: torch.Tensor,
dec_state: DecoderInferenceState,
cfg_scale: float,
temperature: float,
top_p: float,
top_k: int,
current_idx: int,
) -> torch.Tensor:
"""Performs a single step of the decoder inference.
Takes the tokens from the previous step, runs them through the decoder
(for both conditional and unconditional paths), applies classifier-free
guidance (CFG), samples the next token using temperature, top-p, and top-k
sampling, and applies constraints (e.g., preventing EOS in certain channels).
Args:
tokens_Bx1xC: The input tokens for the current step, shape [2*B, 1, C].
Repeated for CFG (unconditional and conditional).
dec_state: The current state of the decoder (KV caches, etc.).
cfg_scale: The scale factor for classifier-free guidance.
temperature: The temperature for sampling.
top_p: The cumulative probability threshold for top-p sampling.
top_k: The number of top logits to consider for top-k sampling.
current_idx: The current generation step index.
Returns:
torch.Tensor: The sampled next tokens for each item in the batch,
shape [B, C].
"""
B = tokens_Bx1xC.shape[0] // 2
audio_eos_value = self.config.eos_token_id
logits_Bx1xCxV = self.model.decoder.decode_step(tokens_Bx1xC, dec_state, current_idx)
logits_last_2BxCxV = logits_Bx1xCxV[:, -1]
logits_last_Bx2xCxV = logits_last_2BxCxV.view(B, 2, *logits_last_2BxCxV.shape[1:])
uncond_logits_BxCxV = logits_last_Bx2xCxV[:, 0, :, :] # Shape [B, C, V]
cond_logits_BxCxV = logits_last_Bx2xCxV[:, 1, :, :] # Shape [B, C, V]
logits_BxCxV = cond_logits_BxCxV + cfg_scale * (cond_logits_BxCxV - uncond_logits_BxCxV)
_, top_k_indices_BxCxk = torch.topk(logits_BxCxV, k=top_k, dim=-1)
mask_BxCxV = torch.ones_like(logits_BxCxV, dtype=torch.bool)
mask_BxCxV = mask_BxCxV.scatter(dim=-1, index=top_k_indices_BxCxk, value=False)
logits_BxCxV = cond_logits_BxCxV.masked_fill(mask_BxCxV, -torch.inf)
logits_BxCxV[:, :, audio_eos_value + 1 :] = torch.full_like(
logits_BxCxV[:, :, audio_eos_value + 1 :],
fill_value=-torch.inf,
)
logits_BxCxV[:, 1:, audio_eos_value:] = torch.full_like(
logits_BxCxV[:, 1:, audio_eos_value:],
fill_value=-torch.inf,
)
flat_logits_BCxV = logits_BxCxV.view(B * self.config.decoder_config.num_channels, -1)
pred_BC = _sample_next_token(
flat_logits_BCxV.float(),
temperature=temperature,
top_p=top_p,
top_k=top_k,
audio_eos_value=audio_eos_value,
)
pred_BxC = pred_BC.view(B, self.config.decoder_config.num_channels)
return pred_BxC
def _generate_output(self, generated_codes: torch.Tensor, lengths_Bx: torch.Tensor) -> list[np.ndarray]:
"""Converts generated delayed codes into audio waveforms.
Reverts the delay pattern applied during generation, decodes the resulting
codebook using the DAC model (if loaded), and returns a list of audio
waveforms as NumPy arrays. If DAC is not loaded, returns the raw codebook indices.
Args:
generated_codes: The tensor of generated audio codes with delays,
shape [B, T_gen, C].
lengths_Bx: A tensor containing the valid length of generated codes
(excluding padding and BOS/EOS markers) for each item
in the batch, shape [B].
Returns:
A list of NumPy arrays, where each array represents the generated audio
waveform for one item in the batch. If DAC is not loaded, returns the
raw, reverted codebook indices as NumPy arrays.
"""
num_channels = self.config.decoder_config.num_channels
batch_size = generated_codes.shape[0]
seq_length = generated_codes.shape[1]
delay_pattern = self.config.delay_pattern
audio_pad_value = self.config.pad_token_id
max_delay_pattern = max(delay_pattern)
revert_precomp = build_revert_indices(
B=batch_size,
T=seq_length,
C=num_channels,
delay_pattern=delay_pattern,
)
codebook = revert_audio_delay(
audio_BxTxC=generated_codes,
pad_value=audio_pad_value,
precomp=revert_precomp,
T=seq_length,
)[:, :-max_delay_pattern, :]
min_valid_index = 0
max_valid_index = 1023
invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
codebook[invalid_mask] = 0
audios = []
if self.load_dac:
for i in range(batch_size):
audio = self._decode(codebook[i, : lengths_Bx[i], :])
audio_np = audio.cpu().numpy()
audios.append(audio_np)
else:
for i in range(batch_size):
audios.append(codebook[i, : lengths_Bx[i], :].cpu().numpy())
return audios
@torch.no_grad()
@torch.inference_mode()
def _encode(self, audio: torch.Tensor) -> torch.Tensor:
"""
Encodes the given audio waveform into a tensor of DAC codebook indices
"""
audio = audio.unsqueeze(0)
audio_data = self.dac_model.preprocess(audio, DEFAULT_SAMPLE_RATE)
_, encoded_frame, _, _, _ = self.dac_model.encode(audio_data)
encoded_frame: torch.Tensor
return encoded_frame.squeeze(0).transpose(0, 1)
@torch.no_grad()
@torch.inference_mode()
def _decode(self, audio_codes: torch.Tensor) -> torch.Tensor:
"""
Decodes the given frames into an output audio waveform
"""
audio_codes = audio_codes.unsqueeze(0).transpose(1, 2)
audio_values, _, _ = self.dac_model.quantizer.from_codes(audio_codes)
audio_values = self.dac_model.decode(audio_values)
audio_values: torch.Tensor
return audio_values.squeeze()
def load_audio(self, audio_path: str) -> torch.Tensor:
"""Loads and preprocesses an audio file for use as a prompt.
Loads the audio file, resamples it to the target sample rate if necessary,
preprocesses it using the DAC model's preprocessing, and encodes it into
DAC codebook indices.
Args:
audio_path: Path to the audio file.
Returns:
torch.Tensor: The encoded audio prompt as DAC codebook indices,
shape [T, C].
Raises:
RuntimeError: If the DAC model is not loaded (`load_dac=False` during init).
FileNotFoundError: If the audio file cannot be found.
Exception: If there's an error during loading or processing.
"""
if self.dac_model is None:
raise RuntimeError("DAC model is required for loading audio prompts but was not loaded.")
audio, sr = torchaudio.load(audio_path, channels_first=True) # C, T
if sr != DEFAULT_SAMPLE_RATE:
audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
# Convert to mono if stereo
if audio.shape[0] > 1:
audio = torch.mean(audio, dim=0, keepdim=True) # Average channels to get mono
return self._encode(audio.to(self.device))
def save_audio(self, path: str, audio: np.ndarray):
"""Saves the generated audio waveform to a file.
Uses the soundfile library to write the NumPy audio array to the specified
path with the default sample rate.
Args:
path: The path where the audio file will be saved.
audio: The audio waveform as a NumPy array.
"""
import soundfile as sf
sf.write(path, audio, DEFAULT_SAMPLE_RATE)
@torch.inference_mode()
def generate(
self,
text: str | list[str],
max_tokens: int = 3072,
cfg_scale: float = 3.0,
temperature: float = 1.2,
top_p: float = 0.95,
use_torch_compile: bool = False,
cfg_filter_top_k: int = 45,
audio_prompt: list[str | torch.Tensor | None] | str | torch.Tensor | None = None,
audio_prompt_path: list[str | torch.Tensor | None] | str | torch.Tensor | None = None,
use_cfg_filter: bool | None = None,
verbose: bool = False,
) -> np.ndarray | list[np.ndarray]:
"""Generates audio corresponding to the input text.
Args:
text: The input text prompt, or a list of text prompts for batch generation.
max_tokens: The maximum number of audio tokens to generate per prompt.
Defaults to the model's configured audio length if None.
cfg_scale: The scale factor for classifier-free guidance (CFG). Higher values
lead to stronger guidance towards the text prompt.
temperature: The temperature for sampling. Higher values increase randomness.
top_p: The cumulative probability threshold for nucleus (top-p) sampling.
use_torch_compile: Whether to compile the generation steps using torch.compile.
Can significantly speed up generation after the initial
compilation overhead. Defaults to False.
cfg_filter_top_k: The number of top logits to consider during CFG filtering.
(Note: This parameter name might be slightly misleading based
on the code; it's used in the `_sample_next_token` function.)
audio_prompt: An audio prompt or list of prompts to condition the generation.
Can be a file path (str), a pre-loaded tensor (DAC codes), or None.
If a list, its length must match the batch size of the text input.
audio_prompt_path: (Deprecated) Use `audio_prompt` instead.
use_cfg_filter: (Deprecated) This parameter is no longer used.
verbose: If True, prints progress information during generation, including
speed metrics.
Returns:
If a single text prompt was provided, returns a NumPy array containing the
generated audio waveform.
If a list of text prompts was provided, returns a list of NumPy arrays,
each corresponding to a prompt in the input list. Returns None for a
sequence if no audio was generated for it.
"""
batch_size = len(text) if isinstance(text, list) else 1
audio_eos_value = self.config.eos_token_id
audio_pad_value = self.config.pad_token_id
delay_pattern = self.config.delay_pattern
max_delay_pattern = max(delay_pattern)
delay_pattern_Cx = torch.tensor(delay_pattern, device=self.device, dtype=torch.long)
self.model.eval()
if audio_prompt_path:
print("Warning: audio_prompt_path is deprecated. Use audio_prompt instead.")
audio_prompt = audio_prompt_path
if use_cfg_filter is not None:
print("Warning: use_cfg_filter is deprecated.")
if verbose:
total_start_time = time.time()
if use_torch_compile and not hasattr(self, "_compiled"):
# Compilation can take about a minute.
self._prepare_generation = torch.compile(self._prepare_generation, dynamic=True, fullgraph=True)
self._decoder_step = torch.compile(self._decoder_step, fullgraph=True, mode="max-autotune")
self._compiled = True
if isinstance(audio_prompt, list):
audio_prompt = [self.load_audio(p) if isinstance(p, str) else p for p in audio_prompt]
elif isinstance(audio_prompt, str):
audio_prompt = [self.load_audio(audio_prompt)]
elif isinstance(audio_prompt, torch.Tensor):
audio_prompt = [audio_prompt]
elif audio_prompt is None:
audio_prompt = [None] * batch_size
assert len(audio_prompt) == batch_size, "Number of audio prompts must match batch size"
if isinstance(text, list):
text = [self._encode_text(t) for t in text]
else:
text = [self._encode_text(text)]
text = self._pad_text_input(text)
dec_state, dec_output = self._prepare_generation(text, audio_prompt, max_tokens=max_tokens)
dec_step = min(dec_output.prefill_steps) - 1
current_idx = torch.tensor([dec_step], device=self.device)
eos_detected_Bx = torch.zeros((batch_size,), dtype=torch.bool, device=self.device)
eos_countdown_Bx = torch.full((batch_size,), -1, dtype=torch.long, device=self.device)
finished_step_Bx = torch.full((batch_size,), -1, dtype=torch.long, device=self.device)
bos_over = False
if verbose:
print("generate: starting generation loop")
if use_torch_compile:
print("generate: using use_torch_compile=True, the first step may be slow")
start_time = time.time()
# --- Generation Loop ---
while dec_step < max_tokens:
if (eos_countdown_Bx == 0).all():
break
current_step_idx = dec_step + 1
torch.compiler.cudagraph_mark_step_begin()
dec_state.prepare_step(dec_step)
tokens_Bx1xC = dec_output.get_tokens_at(dec_step).repeat_interleave(2, dim=0) # Repeat for CFG
pred_BxC = self._decoder_step(
tokens_Bx1xC,
dec_state,
cfg_scale,
temperature,
top_p,
cfg_filter_top_k,
current_idx,
)
current_idx += 1
active_mask_Bx = eos_countdown_Bx != 0
eos_trigger_Bx = torch.zeros_like(active_mask_Bx)
if active_mask_Bx.any():
is_eos_token = (~eos_detected_Bx[active_mask_Bx]) & (pred_BxC[active_mask_Bx, 0] == audio_eos_value)
is_max_len = current_step_idx >= max_tokens - max_delay_pattern
eos_trigger_Bx[active_mask_Bx] = is_eos_token | is_max_len
eos_detected_Bx |= eos_trigger_Bx
start_countdown_mask_Bx = eos_trigger_Bx & (eos_countdown_Bx < 0)
if start_countdown_mask_Bx.any():
eos_countdown_Bx[start_countdown_mask_Bx] = max_delay_pattern
finished_step_Bx[start_countdown_mask_Bx] = current_step_idx
padding_mask_Bx = eos_countdown_Bx > 0
if padding_mask_Bx.any():
pred_active_BxC = pred_BxC[padding_mask_Bx].clone()
countdown_active_Bx = eos_countdown_Bx[padding_mask_Bx]
step_after_eos_Bx = max_delay_pattern - countdown_active_Bx
step_after_eos_Bx_ = step_after_eos_Bx.unsqueeze(1)
delay_pattern_Cx_ = delay_pattern_Cx.unsqueeze(0)
eos_mask_NxC = step_after_eos_Bx_ == delay_pattern_Cx_
pad_mask_NxC = step_after_eos_Bx_ > delay_pattern_Cx_
pred_active_BxC[eos_mask_NxC] = audio_eos_value
pred_active_BxC[pad_mask_NxC] = audio_pad_value
pred_BxC[padding_mask_Bx] = pred_active_BxC
eos_countdown_Bx[padding_mask_Bx] -= 1
# --- Update BOS flag (Original) ---
if not bos_over:
bos_over = all(
dec_step - prefill_step > max_delay_pattern for prefill_step in dec_output.prefill_steps
)
dec_output.update_one(pred_BxC, current_step_idx, not bos_over)
dec_step += 1
if verbose and dec_step % 86 == 0:
duration = time.time() - start_time
if duration > 0:
print(
f"generate step {dec_step}: speed={86 * batch_size / duration:.3f} tokens/s, realtime factor={batch_size / duration:.3f}x"
)
start_time = time.time()
# --- Finalize and Extract Output ---
final_step = dec_step + 1
finished_step_Bx[finished_step_Bx == -1] = final_step - max_delay_pattern
prefill_steps_tensor = torch.tensor(dec_output.prefill_steps, device=self.device)
lengths_Bx = finished_step_Bx - prefill_steps_tensor
lengths_Bx = torch.clamp(lengths_Bx, min=0)
max_len = lengths_Bx.max().item() + max_delay_pattern
outputs = []
if max_len > 0:
num_channels = self.config.decoder_config.num_channels
audio_pad_value = self.config.pad_token_id
generated_codes = torch.full(
(batch_size, max_len, num_channels),
fill_value=audio_pad_value,
dtype=torch.long,
device=self.device,
)
for i in range(batch_size):
start_step = dec_output.prefill_steps[i]
actual_len = lengths_Bx[i].item() + max_delay_pattern
if actual_len > 0:
tokens_to_copy = dec_output.generated_tokens[i, start_step : start_step + actual_len, :]
generated_codes[i, :actual_len, :] = tokens_to_copy
if verbose:
avg_steps = lengths_Bx.float().mean().item()
total_duration = time.time() - total_start_time
print(f"generate: avg steps={avg_steps:.1f}, total duration={total_duration:.3f}s")
del dec_state
outputs = self._generate_output(generated_codes, lengths_Bx)
else:
print("Warning: Nothing generated for any sequence in the batch.")
outputs = [None] * batch_size
return outputs if batch_size > 1 else outputs[0]
================================================
FILE: dia/state.py
================================================
from dataclasses import dataclass
from typing import Optional
import torch
from .config import DiaConfig
def create_attn_mask(
q_padding_mask_1d: torch.Tensor,
k_padding_mask_1d: torch.Tensor,
device: torch.device,
is_causal: bool = False,
) -> torch.Tensor:
"""
Creates the attention mask (self or cross) mimicking JAX segment ID logic.
"""
# B1, Tq = q_padding_mask_1d.shape
# B2, Tk = k_padding_mask_1d.shape
p_mask_q = q_padding_mask_1d.unsqueeze(2) # Shape [B, Tq, 1]
p_mask_k = k_padding_mask_1d.unsqueeze(1) # Shape [B, 1, Tk]
# Condition A: Non-padding query attends to non-padding key
non_pad_attends_non_pad = p_mask_q & p_mask_k # Shape [B, Tq, Tk]
# Condition B: Padding query attends to padding key
pad_attends_pad = (~p_mask_q) & (~p_mask_k) # Shape [B, Tq, Tk]
# Combine: True if padding status is compatible (both non-pad OR both pad)
mask = non_pad_attends_non_pad | pad_attends_pad # Shape [B, Tq, Tk]
if is_causal:
# assert Tq == Tk, "Causal mask requires query and key sequence lengths to be equal"
causal_mask_2d = torch.tril(torch.ones_like(mask[0], dtype=torch.bool, device=device)) # Shape [B, Tq, Tk]
causal_mask = mask & causal_mask_2d # Shape [B, Tq, Tk]
return causal_mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
else:
return mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
@dataclass
class EncoderInferenceState:
"""Parameters specifically for encoder inference."""
max_seq_len: int
device: torch.device
positions: torch.Tensor
padding_mask: torch.Tensor
attn_mask: torch.Tensor
@classmethod
def new(cls, config: DiaConfig, cond_src: torch.Tensor) -> "EncoderInferenceState":
"""Creates EtorchrInferenceParams from DiaConfig and a device."""
device = cond_src.device
positions = torch.arange(
config.encoder_config.max_position_embeddings, dtype=torch.float32, device=device
).unsqueeze(0)
padding_mask = (cond_src.squeeze(1) != 0).to(device).repeat_interleave(2, dim=0)
attn_mask = create_attn_mask(padding_mask, padding_mask, device, is_causal=False)
return cls(
max_seq_len=config.encoder_config.max_position_embeddings,
device=device,
positions=positions,
padding_mask=padding_mask,
attn_mask=attn_mask,
)
class KVCache(torch.nn.Module):
k: torch.Tensor
v: torch.Tensor
def __init__(
self,
batch_size: int,
num_heads: int,
max_len: int,
head_dim: int,
dtype: torch.dtype,
device: torch.device,
k: torch.Tensor | None = None,
v: torch.Tensor | None = None,
):
k = torch.zeros((2 * batch_size, num_heads, max_len, head_dim), dtype=dtype, device=device) if k is None else k
v = torch.zeros((2 * batch_size, num_heads, max_len, head_dim), dtype=dtype, device=device) if v is None else v
super().__init__()
self.register_buffer("k", k)
self.register_buffer("v", v)
@classmethod
def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
return cls(
batch_size=k.shape[0] // 2,
num_heads=k.shape[1],
max_len=k.shape[2],
head_dim=k.shape[3],
dtype=k.dtype,
device=k.device,
k=k,
v=v,
)
def update(self, k: torch.Tensor, v: torch.Tensor, current_idx: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
k_out, v_out = self.k, self.v
k_out[:, :, current_idx, :] = k
v_out[:, :, current_idx, :] = v
return self.k, self.v
def prefill(self, k: torch.Tensor, v: torch.Tensor):
prefill_len = k.shape[2]
self.k[:, :, :prefill_len, :] = k
self.v[:, :, :prefill_len, :] = v
@dataclass
class DecoderInferenceState:
"""Parameters specifically for decoder inference."""
device: torch.device
dtype: torch.dtype
enc_out: torch.Tensor
enc_positions: torch.Tensor
dec_positions: torch.Tensor
self_attn_cache: list[KVCache]
cross_attn_cache: list[KVCache]
casual_attn_mask: torch.Tensor
cross_attn_mask: torch.Tensor
@classmethod
def new(
cls,
config: DiaConfig,
enc_state: EncoderInferenceState,
enc_out: torch.Tensor,
dec_cross_attn_cache: list[KVCache],
compute_dtype: torch.dtype,
max_generation_length: Optional[int] = None,
) -> "DecoderInferenceState":
"""Creates DecoderInferenceParams from DiaConfig and a device."""
device = enc_out.device
max_audio_len = max_generation_length or config.decoder_config.max_position_embeddings
batch_size = enc_out.shape[0] // 2
dec_positions = torch.full((2 * batch_size, 1), fill_value=0, dtype=torch.int32, device=device)
causal_mask = torch.tril(torch.ones(max_audio_len, max_audio_len, dtype=torch.bool, device=device))
dec_mask = torch.ones((2 * batch_size, 1), dtype=torch.bool, device=device)
cross_attn_mask = create_attn_mask(dec_mask, enc_state.padding_mask, device, is_causal=False)
self_attn_cache = [
KVCache(
batch_size,
config.decoder_config.num_key_value_heads,
max_audio_len,
config.decoder_config.head_dim,
compute_dtype,
device,
)
for _ in range(config.decoder_config.num_hidden_layers)
]
return cls(
device=device,
dtype=compute_dtype,
enc_out=enc_out,
enc_positions=enc_state.positions,
dec_positions=dec_positions,
self_attn_cache=self_attn_cache,
cross_attn_cache=dec_cross_attn_cache,
casual_attn_mask=causal_mask,
cross_attn_mask=cross_attn_mask,
)
def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
if step_to is None:
step_to = step_from + 1
self.dec_positions = torch.arange(step_from, step_to, dtype=torch.int32, device=self.device).unsqueeze(0)
@dataclass
class DecoderOutput:
generated_tokens: torch.Tensor
prefill_steps: list[int]
@classmethod
def new(cls, batch_size: int, config: DiaConfig, device: torch.device) -> "DecoderOutput":
max_audio_len = config.decoder_config.max_position_embeddings
return cls(
generated_tokens=torch.full(
(batch_size, max_audio_len, config.decoder_config.num_channels),
fill_value=-1,
dtype=torch.int,
device=device,
),
prefill_steps=[],
)
def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
if step_to is None:
step_to = step_from + 1
return self.generated_tokens[:, step_from:step_to, :]
def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
dec_out = dec_out.to(self.generated_tokens.dtype)
if apply_mask:
mask = self.generated_tokens[:, step, :] == -1
self.generated_tokens[:, step, :] = torch.where(mask, dec_out, self.generated_tokens[:, step, :])
else:
self.generated_tokens[:, step, :] = dec_out
def prefill(self, dec_out: torch.Tensor, prefill_steps: list[int]):
length = dec_out.shape[1]
self.generated_tokens[:, :length, :] = dec_out
self.prefill_steps = prefill_steps
================================================
FILE: docker/Dockerfile.cpu
================================================
# Dockerfile.cpu - CPU-only deployment for DIA
# --------------------------------------------------
# Build: docker build . -f docker/Dockerfile.cpu -t dia-cpu
# Run: docker run --rm -p 7860:7860 dia-cpu
FROM python:3.10-slim
# Set non-interactive frontend
ENV DEBIAN_FRONTEND=noninteractive
# Install venv, and system dependencies
RUN apt-get update && apt-get install -y \
python3-venv \
libsndfile1 \
ffmpeg \
curl \
&& apt-get clean && rm -rf /var/lib/apt/lists/*
# Create non-root user and set up directories
RUN useradd -m -u 1001 appuser && \
mkdir -p /app/outputs /app && \
chown -R appuser:appuser /app
USER appuser
WORKDIR /app
# Copy all code (including pyproject.toml)
COPY --chown=appuser:appuser . .
# Create and activate virtual environment
RUN python3 -m venv /app/venv
ENV PATH="/app/venv/bin:$PATH"
# Install all project dependencies (CPU-only PyTorch)
RUN pip install --upgrade pip && \
pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu && \
pip install --no-cache-dir -e .[dev]
# Set environment variables
ENV PYTHONUNBUFFERED=1 \
PYTHONPATH=/app
# Expose Gradio default port
ENV GRADIO_SERVER_NAME="0.0.0.0"
EXPOSE 7860
# Entrypoint
CMD ["python3", "app.py"]
================================================
FILE: docker/Dockerfile.gpu
================================================
# Dockerfile.gpu - GPU deployment for DIA
# --------------------------------------------------
# Build: docker build . -f docker/Dockerfile.gpu -t dia-gpu
# Run: docker run --rm --gpus all -p 7860:7860 dia-gpu
# Requires NVIDIA Container Toolkit on host.
FROM pytorch/pytorch:2.1.2-cuda12.1-cudnn8-runtime
# Set non-interactive frontend
ENV DEBIAN_FRONTEND=noninteractive
# Install venv, and system dependencies
RUN apt-get update && apt-get install -y \
python3-venv \
libsndfile1 \
ffmpeg \
curl \
&& apt-get clean && rm -rf /var/lib/apt/lists/*
# Create non-root user and set up directories
RUN useradd -m -u 1001 appuser && \
mkdir -p /app/outputs /app && \
chown -R appuser:appuser /app
USER appuser
WORKDIR /app
# Copy all code (including pyproject.toml)
COPY --chown=appuser:appuser . .
# Create and activate virtual environment
RUN python3 -m venv /app/venv
ENV PATH="/app/venv/bin:$PATH"
# Install all project dependencies
RUN pip install --upgrade pip && pip install --no-cache-dir .
# Set environment variables
ENV PYTHONUNBUFFERED=1 \
PYTHONPATH=/app \
USE_GPU=true \
LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda-12.1/lib64:${LD_LIBRARY_PATH}
# Expose Gradio default port
ENV GRADIO_SERVER_NAME="0.0.0.0"
EXPOSE 7860
# Entrypoint
CMD ["python3", "app.py"]
================================================
FILE: example/benchmark.py
================================================
from random import choice
import torch
from dia.model import Dia
torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.triton.unique_kernel_names = True
torch._inductor.config.fx_graph_cache = True
# debugging
torch._logging.set_logs(graph_breaks=True, recompiles=True)
model_name = "nari-labs/Dia-1.6B-0626"
compute_dtype = "float16"
model = Dia.from_pretrained(model_name, compute_dtype=compute_dtype)
test_cases = [
"[S1] Dia is an open weights text to dialogue model.",
"[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face.",
"[S1] torch.compile is a new feature in PyTorch that allows you to compile your model with a single line of code.",
"[S1] torch.compile is a new feature in PyTorch that allows you to compile your model with a single line of code. [S2] It is a new feature in PyTorch that allows you to compile your model with a single line of code.",
]
# Wram up
for _ in range(2):
text = choice(test_cases)
output = model.generate(text, audio_prompt="./example_prompt.mp3", use_torch_compile=True, verbose=True)
output = model.generate(text, use_torch_compile=True, verbose=True)
# Benchmark
for _ in range(10):
text = choice(test_cases)
output = model.generate(text, use_torch_compile=True, verbose=True)
output = model.generate(text, audio_prompt="./example_prompt.mp3", use_torch_compile=True, verbose=True)
================================================
FILE: example/simple-cpu.py
================================================
import torch
from dia.model import Dia
# Select device: CPU
device = torch.device("cpu")
print(f"Using device: {device}")
# Load model
model = Dia.from_pretrained(
"nari-labs/Dia-1.6B-0626", compute_dtype="float32", device=device
) # Float32 works better than float16 on CPU - you can also test with float16
text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
output = model.generate(text, use_torch_compile=False, verbose=True)
model.save_audio("simple.mp3", output)
================================================
FILE: example/simple-mac.py
================================================
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
# It is important to set the `use_torch_compile` argument to `False` when using Dia on MacOS.
# This is because the `torch.compile` function is not supported on MacOS.
output = model.generate(text, use_torch_compile=False, verbose=True)
model.save_audio("simple.mp3", output)
================================================
FILE: example/simple.py
================================================
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
output = model.generate(
text,
use_torch_compile=False,
verbose=True,
cfg_scale=3.0,
temperature=1.8,
top_p=0.90,
cfg_filter_top_k=50,
)
model.save_audio("simple.mp3", output)
================================================
FILE: example/simple_batch.py
================================================
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
texts = [text for _ in range(10)]
output = model.generate(texts, use_torch_compile=True, verbose=True, max_tokens=1500)
for i, o in enumerate(output):
model.save_audio(f"simple_{i}.mp3", o)
================================================
FILE: example/voice_clone.py
================================================
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
# You should put the transcript of the voice you want to clone
# We will use the audio created by running simple.py as an example.
# Note that you will be REQUIRED TO RUN simple.py for the script to work as-is.
clone_from_text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
clone_from_audio = "simple.mp3"
# For your custom needs, replace above with below and add your audio file to this directory:
# clone_from_text = "[S1] ... [S2] ... [S1] ... corresponding to your_audio_name.mp3"
# clone_from_audio = "your_audio_name.mp3"
# Text to generate
text_to_generate = "[S1] Hello, how are you? [S2] I'm good, thank you. [S1] What's your name? [S2] My name is Dia. [S1] Nice to meet you. [S2] Nice to meet you too."
# It will only return the audio from the text_to_generate
output = model.generate(
clone_from_text + text_to_generate,
audio_prompt=clone_from_audio,
use_torch_compile=False,
verbose=True,
cfg_scale=4.0,
temperature=1.8,
top_p=0.90,
cfg_filter_top_k=50,
)
model.save_audio("voice_clone.mp3", output)
================================================
FILE: example/voice_clone_batch.py
================================================
from dia.model import Dia
model = Dia.from_pretrained("nari-labs/Dia-1.6B-0626", compute_dtype="float16")
# You should put the transcript of the voice you want to clone
# We will use the audio created by running simple.py as an example.
# Note that you will be REQUIRED TO RUN simple.py for the script to work as-is.
clone_from_text = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
# For your custom needs, replace above with below and add your audio file to this directory:
# clone_from_text = "[S1] ... [S2] ... [S1] ... corresponding to your_audio_name.mp3"
# clone_from_audio = "your_audio_name.mp3"
# Text to generate
text_to_generate = "[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
clone_from_audios = [f"simple_{i}.mp3" for i in range(10)]
texts = [clone_from_text + text_to_generate for _ in range(10)]
# It will only return the audio from the text_to_generate
output = model.generate(texts, audio_prompt=clone_from_audios, use_torch_compile=True, verbose=True, max_tokens=2000)
for i, o in enumerate(output):
model.save_audio(f"voice_clone_{i}.mp3", o)
================================================
FILE: hf.py
================================================
from transformers import AutoProcessor, DiaForConditionalGeneration
torch_device = "cuda"
model_checkpoint = "nari-labs/Dia-1.6B-0626"
text = [
"[S1] Dia is an open weights text to dialogue model. [S2] You get full control over scripts and voices. [S1] Wow. Amazing. (laughs) [S2] Try it now on Git hub or Hugging Face."
]
processor = AutoProcessor.from_pretrained(model_checkpoint)
inputs = processor(text=text, padding=True, return_tensors="pt").to(torch_device)
model = DiaForConditionalGeneration.from_pretrained(model_checkpoint).to(torch_device)
outputs = model.generate(**inputs, max_new_tokens=3072, guidance_scale=3.0, temperature=1.8, top_p=0.90, top_k=45)
outputs = processor.batch_decode(outputs)
processor.save_audio(outputs, "example.mp3")
================================================
FILE: pyproject.toml
================================================
[project]
name = "nari-tts"
version = "0.1.0"
description = "Dia - A text-to-speech model for dialogue generation"
readme = "README.md"
requires-python = ">=3.10"
license = {file = "LICENSE"}
authors = [
{name = "Nari Labs", email = "contact@narilabs.ai"}
]
dependencies = [
"descript-audio-codec>=1.0.0",
"gradio>=5.25.2",
"huggingface-hub>=0.30.2",
"numpy>=2.2.4",
"pydantic>=2.11.3",
"safetensors>=0.5.3",
"soundfile>=0.13.1",
"torch==2.6.0",
"torchaudio==2.6.0",
"triton==3.2.0 ; sys_platform == 'linux'",
"triton-windows==3.2.0.post18 ; sys_platform == 'win32'",
]
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[project.urls]
"Homepage" = "https://github.com/nari-labs/dia"
"Bug Tracker" = "https://github.com/nari-labs/dia/issues"
[tool.hatch.build.targets.wheel]
packages = ["dia"]
[tool.ruff]
# Never enforce `E501` (line length violations).
lint.ignore = ["C901", "E501", "E741", "W605"]
lint.select = ["C", "E", "F", "I", "W"]
line-length = 119
# Ignore import violations in all `__init__.py` files.
[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["E402", "F401", "F403", "F811"]
[tool.ruff.lint.isort]
lines-after-imports = 2
[tool.uv.sources]
torch = [
{ index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
]
torchaudio = [
{ index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
]
[[tool.uv.index]]
name = "pytorch-cu126"
url = "https://download.pytorch.org/whl/cu126"
explicit = true
[dependency-groups]
dev = [
"ninja>=1.11.1.4",
"packaging>=25.0",
]
gitextract_f5lpo0wj/ ├── .github/ │ └── workflows/ │ └── ci.yaml ├── .gitignore ├── .python-version ├── LICENSE ├── README.md ├── app.py ├── cli.py ├── dia/ │ ├── __init__.py │ ├── audio.py │ ├── config.py │ ├── layers.py │ ├── model.py │ └── state.py ├── docker/ │ ├── Dockerfile.cpu │ └── Dockerfile.gpu ├── example/ │ ├── benchmark.py │ ├── simple-cpu.py │ ├── simple-mac.py │ ├── simple.py │ ├── simple_batch.py │ ├── voice_clone.py │ └── voice_clone_batch.py ├── hf.py └── pyproject.toml
SYMBOL INDEX (88 symbols across 7 files)
FILE: app.py
function set_seed (line 57) | def set_seed(seed: int):
function run_inference (line 69) | def run_inference(
FILE: cli.py
function set_seed (line 12) | def set_seed(seed: int):
function main (line 25) | def main():
FILE: dia/audio.py
function build_delay_indices (line 6) | def build_delay_indices(B: int, T: int, C: int, delay_pattern: tp.List[i...
function apply_audio_delay (line 44) | def apply_audio_delay(
function build_revert_indices (line 88) | def build_revert_indices(B: int, T: int, C: int, delay_pattern: tp.List[...
function revert_audio_delay (line 125) | def revert_audio_delay(
FILE: dia/config.py
class EncoderConfig (line 21) | class EncoderConfig(BaseModel, frozen=True):
class DecoderConfig (line 57) | class DecoderConfig(BaseModel, frozen=True):
class DiaConfig (line 103) | class DiaConfig(BaseModel, frozen=True):
method save (line 140) | def save(self, path: str) -> None:
method load (line 157) | def load(cls, path: str) -> "DiaConfig | None":
FILE: dia/layers.py
function _normalize_axes (line 12) | def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
class DenseGeneral (line 16) | class DenseGeneral(nn.Module):
method __init__ (line 32) | def __init__(
method forward (line 49) | def forward(self, inputs: Tensor) -> Tensor:
class MlpBlock (line 61) | class MlpBlock(nn.Module):
method __init__ (line 64) | def __init__(self, embed_dim: int, intermediate_dim: int, compute_dtyp...
method forward (line 82) | def forward(self, x: torch.Tensor) -> torch.Tensor:
class RotaryEmbedding (line 95) | class RotaryEmbedding(nn.Module):
method __init__ (line 98) | def __init__(
method forward (line 118) | def forward(self, inputs: torch.Tensor, position: torch.Tensor):
method apply_rope (line 132) | def apply_rope(self, inputs: torch.Tensor, sin: torch.Tensor, cos: tor...
function custom_scaled_dot_product_attention (line 139) | def custom_scaled_dot_product_attention(
class CrossAttention (line 192) | class CrossAttention(nn.Module):
method __init__ (line 195) | def __init__(
method forward (line 249) | def forward(
class FusedQKV (line 313) | class FusedQKV(nn.Module):
method __init__ (line 314) | def __init__(
method forward (line 333) | def forward(self, inputs: torch.Tensor) -> tuple[torch.Tensor, torch.T...
class SelfAttention (line 345) | class SelfAttention(nn.Module):
method __init__ (line 348) | def __init__(
method get_linear_weight (line 406) | def get_linear_weight(self, dense: DenseGeneral):
method patch_fused_qkv (line 420) | def patch_fused_qkv(self):
method forward (line 439) | def forward(
class EncoderLayer (line 531) | class EncoderLayer(nn.Module):
method __init__ (line 534) | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
method forward (line 567) | def forward(
class Encoder (line 591) | class Encoder(nn.Module):
method __init__ (line 594) | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
method forward (line 612) | def forward(
class DecoderLayer (line 626) | class DecoderLayer(nn.Module):
method __init__ (line 629) | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
method forward (line 684) | def forward(
class Decoder (line 730) | class Decoder(nn.Module):
method __init__ (line 733) | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
method precompute_cross_attn_cache (line 763) | def precompute_cross_attn_cache(
method decode_step (line 784) | def decode_step(
method forward (line 819) | def forward(self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInference...
class DiaModel (line 869) | class DiaModel(
method __init__ (line 884) | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
FILE: dia/model.py
function _get_default_device (line 20) | def _get_default_device():
function _sample_next_token (line 28) | def _sample_next_token(
class ComputeDtype (line 79) | class ComputeDtype(str, Enum):
method to_dtype (line 84) | def to_dtype(self) -> torch.dtype:
class Dia (line 95) | class Dia:
method __init__ (line 96) | def __init__(
method from_local (line 132) | def from_local(
method from_pretrained (line 177) | def from_pretrained(
method _load_dac_model (line 221) | def _load_dac_model(self):
method _encode_text (line 240) | def _encode_text(self, text: str) -> torch.Tensor:
method _pad_text_input (line 265) | def _pad_text_input(self, text_tokens: list[torch.Tensor]) -> torch.Te...
method _prepare_audio_prompt (line 282) | def _prepare_audio_prompt(self, audio_prompts: list[torch.Tensor | Non...
method _prepare_generation (line 343) | def _prepare_generation(
method _decoder_step (line 399) | def _decoder_step(
method _generate_output (line 469) | def _generate_output(self, generated_codes: torch.Tensor, lengths_Bx: ...
method _encode (line 528) | def _encode(self, audio: torch.Tensor) -> torch.Tensor:
method _decode (line 540) | def _decode(self, audio_codes: torch.Tensor) -> torch.Tensor:
method load_audio (line 550) | def load_audio(self, audio_path: str) -> torch.Tensor:
method save_audio (line 579) | def save_audio(self, path: str, audio: np.ndarray):
method generate (line 594) | def generate(
FILE: dia/state.py
function create_attn_mask (line 9) | def create_attn_mask(
class EncoderInferenceState (line 43) | class EncoderInferenceState:
method new (line 53) | def new(cls, config: DiaConfig, cond_src: torch.Tensor) -> "EncoderInf...
class KVCache (line 72) | class KVCache(torch.nn.Module):
method __init__ (line 76) | def __init__(
method from_kv (line 95) | def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
method update (line 107) | def update(self, k: torch.Tensor, v: torch.Tensor, current_idx: torch....
method prefill (line 113) | def prefill(self, k: torch.Tensor, v: torch.Tensor):
class DecoderInferenceState (line 120) | class DecoderInferenceState:
method new (line 134) | def new(
method prepare_step (line 177) | def prepare_step(self, step_from: int, step_to: int | None = None) -> ...
class DecoderOutput (line 184) | class DecoderOutput:
method new (line 189) | def new(cls, batch_size: int, config: DiaConfig, device: torch.device)...
method get_tokens_at (line 201) | def get_tokens_at(self, step_from: int, step_to: int | None = None) ->...
method update_one (line 206) | def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: boo...
method prefill (line 214) | def prefill(self, dec_out: torch.Tensor, prefill_steps: list[int]):
Condensed preview — 24 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (151K chars).
[
{
"path": ".github/workflows/ci.yaml",
"chars": 420,
"preview": "name: Continuous Integration\n\non:\n pull_request:\n branches:\n - main\n\njobs:\n lint_and_format:\n runs-on: ubun"
},
{
"path": ".gitignore",
"chars": 232,
"preview": "# Python-generated files\n__pycache__/\n*.py[oc]\nbuild/\ndist/\nwheels/\n*.egg-info\n\n# Virtual environments\n.venv\n.idea/\n.gra"
},
{
"path": ".python-version",
"chars": 5,
"preview": "3.10\n"
},
{
"path": "LICENSE",
"chars": 11338,
"preview": " Apache License\n Version 2.0, January 2004\n "
},
{
"path": "README.md",
"chars": 9850,
"preview": "<p align=\"center\">\n<a href=\"https://github.com/nari-labs/dia\">\n<img src=\"./dia/static/images/banner.png\">\n</a>\n</p>\n<p a"
},
{
"path": "app.py",
"chars": 18992,
"preview": "import argparse\nimport contextlib\nimport io\nimport random\nimport tempfile\nimport time\nfrom pathlib import Path\nfrom typi"
},
{
"path": "cli.py",
"chars": 4934,
"preview": "import argparse\nimport os\nimport random\n\nimport numpy as np\nimport soundfile as sf\nimport torch\n\nfrom dia.model import D"
},
{
"path": "dia/__init__.py",
"chars": 50,
"preview": "from .model import Dia\n\n\n__all__ = [\n \"Dia\",\n]\n"
},
{
"path": "dia/audio.py",
"chars": 6142,
"preview": "import typing as tp\n\nimport torch\n\n\ndef build_delay_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.T"
},
{
"path": "dia/config.py",
"chars": 8746,
"preview": "\"\"\"Configuration management module for the Dia model.\n\nThis module provides comprehensive configuration management for t"
},
{
"path": "dia/layers.py",
"chars": 31662,
"preview": "import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom huggingface_hub import PyTorchModelHubMixin\nfrom"
},
{
"path": "dia/model.py",
"chars": 33337,
"preview": "import time\nfrom enum import Enum\nfrom typing import Callable\n\nimport numpy as np\nimport torch\nimport torch.nn.functiona"
},
{
"path": "dia/state.py",
"chars": 7659,
"preview": "from dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\n\nfrom .config import DiaConfig\n\n\ndef create_"
},
{
"path": "docker/Dockerfile.cpu",
"chars": 1261,
"preview": "# Dockerfile.cpu - CPU-only deployment for DIA\n# --------------------------------------------------\n# Build: docker buil"
},
{
"path": "docker/Dockerfile.gpu",
"chars": 1330,
"preview": "# Dockerfile.gpu - GPU deployment for DIA\n# --------------------------------------------------\n# Build: docker build . -"
},
{
"path": "example/benchmark.py",
"chars": 1524,
"preview": "from random import choice\n\nimport torch\n\nfrom dia.model import Dia\n\n\ntorch._inductor.config.coordinate_descent_tuning = "
},
{
"path": "example/simple-cpu.py",
"chars": 612,
"preview": "import torch\n\nfrom dia.model import Dia\n\n\n# Select device: CPU\ndevice = torch.device(\"cpu\")\nprint(f\"Using device: {devic"
},
{
"path": "example/simple-mac.py",
"chars": 571,
"preview": "from dia.model import Dia\n\n\nmodel = Dia.from_pretrained(\"nari-labs/Dia-1.6B-0626\", compute_dtype=\"float16\")\n\ntext = \"[S1"
},
{
"path": "example/simple.py",
"chars": 499,
"preview": "from dia.model import Dia\n\n\nmodel = Dia.from_pretrained(\"nari-labs/Dia-1.6B-0626\", compute_dtype=\"float16\")\n\ntext = \"[S1"
},
{
"path": "example/simple_batch.py",
"chars": 489,
"preview": "from dia.model import Dia\n\n\nmodel = Dia.from_pretrained(\"nari-labs/Dia-1.6B-0626\", compute_dtype=\"float16\")\n\ntext = \"[S1"
},
{
"path": "example/voice_clone.py",
"chars": 1290,
"preview": "from dia.model import Dia\n\n\nmodel = Dia.from_pretrained(\"nari-labs/Dia-1.6B-0626\", compute_dtype=\"float16\")\n\n# You shoul"
},
{
"path": "example/voice_clone_batch.py",
"chars": 1336,
"preview": "from dia.model import Dia\n\n\nmodel = Dia.from_pretrained(\"nari-labs/Dia-1.6B-0626\", compute_dtype=\"float16\")\n\n# You shoul"
},
{
"path": "hf.py",
"chars": 763,
"preview": "from transformers import AutoProcessor, DiaForConditionalGeneration\n\n\ntorch_device = \"cuda\"\nmodel_checkpoint = \"nari-lab"
},
{
"path": "pyproject.toml",
"chars": 1640,
"preview": "[project]\nname = \"nari-tts\"\nversion = \"0.1.0\"\ndescription = \"Dia - A text-to-speech model for dialogue generation\"\nreadm"
}
]
About this extraction
This page contains the full source code of the nari-labs/dia GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 24 files (141.3 KB), approximately 34.7k tokens, and a symbol index with 88 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.
Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.