├── .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
└── static
│ └── images
│ └── banner.png
├── docker
├── Dockerfile.cpu
└── Dockerfile.gpu
├── example
├── simple.py
└── voice_clone.py
├── example_prompt.mp3
├── pyproject.toml
└── uv.lock
/.github/workflows/ci.yaml:
--------------------------------------------------------------------------------
1 | name: Continuous Integration
2 |
3 | on:
4 | pull_request:
5 | branches:
6 | - main
7 |
8 | jobs:
9 | lint_and_format:
10 | runs-on: ubuntu-latest
11 | name: Lint and Format
12 | steps:
13 | - uses: actions/checkout@v4
14 | - uses: astral-sh/ruff-action@v3
15 | with:
16 | version: latest
17 |
18 | - name: Check Lint using Ruff
19 | run: ruff check
20 |
21 | - name: Check Format using Ruff
22 | run: ruff format --check --diff
23 |
--------------------------------------------------------------------------------
/.gitignore:
--------------------------------------------------------------------------------
1 | # Python-generated files
2 | __pycache__/
3 | *.py[oc]
4 | build/
5 | dist/
6 | wheels/
7 | *.egg-info
8 |
9 | # Virtual environments
10 | .venv
11 |
12 | .gradio
13 |
14 | **/*.pth
15 | **/*.mp3
16 | !example_prompt.mp3
17 | **/*.txt
18 |
19 | .ruff_cache
20 | .ipynb_checkpoints
21 | config.json
--------------------------------------------------------------------------------
/.python-version:
--------------------------------------------------------------------------------
1 | 3.10
2 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | Apache License
2 | Version 2.0, January 2004
3 | http://www.apache.org/licenses/
4 |
5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6 |
7 | 1. Definitions.
8 |
9 | "License" shall mean the terms and conditions for use, reproduction,
10 | and distribution as defined by Sections 1 through 9 of this document.
11 |
12 | "Licensor" shall mean the copyright owner or entity authorized by
13 | the copyright owner that is granting the License.
14 |
15 | "Legal Entity" shall mean the union of the acting entity and all
16 | other entities that control, are controlled by, or are under common
17 | control with that entity. For the purposes of this definition,
18 | "control" means (i) the power, direct or indirect, to cause the
19 | direction or management of such entity, whether by contract or
20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the
21 | outstanding shares, or (iii) beneficial ownership of such entity.
22 |
23 | "You" (or "Your") shall mean an individual or Legal Entity
24 | exercising permissions granted by this License.
25 |
26 | "Source" form shall mean the preferred form for making modifications,
27 | including but not limited to software source code, documentation
28 | source, and configuration files.
29 |
30 | "Object" form shall mean any form resulting from mechanical
31 | transformation or translation of a Source form, including but
32 | not limited to compiled object code, generated documentation,
33 | and conversions to other media types.
34 |
35 | "Work" shall mean the work of authorship, whether in Source or
36 | Object form, made available under the License, as indicated by a
37 | copyright notice that is included in or attached to the work
38 | (an example is provided in the Appendix below).
39 |
40 | "Derivative Works" shall mean any work, whether in Source or Object
41 | form, that is based on (or derived from) the Work and for which the
42 | editorial revisions, annotations, elaborations, or other modifications
43 | represent, as a whole, an original work of authorship. For the purposes
44 | of this License, Derivative Works shall not include works that remain
45 | separable from, or merely link (or bind by name) to the interfaces of,
46 | the Work and Derivative Works thereof.
47 |
48 | "Contribution" shall mean any work of authorship, including
49 | the original version of the Work and any modifications or additions
50 | to that Work or Derivative Works thereof, that is intentionally
51 | submitted to Licensor for inclusion in the Work by the copyright owner
52 | or by an individual or Legal Entity authorized to submit on behalf of
53 | the copyright owner. For the purposes of this definition, "submitted"
54 | means any form of electronic, verbal, or written communication sent
55 | to the Licensor or its representatives, including but not limited to
56 | communication on electronic mailing lists, source code control systems,
57 | and issue tracking systems that are managed by, or on behalf of, the
58 | Licensor for the purpose of discussing and improving the Work, but
59 | excluding communication that is conspicuously marked or otherwise
60 | designated in writing by the copyright owner as "Not a Contribution."
61 |
62 | "Contributor" shall mean Licensor and any individual or Legal Entity
63 | on behalf of whom a Contribution has been received by Licensor and
64 | subsequently incorporated within the Work.
65 |
66 | 2. Grant of Copyright License. Subject to the terms and conditions of
67 | this License, each Contributor hereby grants to You a perpetual,
68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69 | copyright license to reproduce, prepare Derivative Works of,
70 | publicly display, publicly perform, sublicense, and distribute the
71 | Work and such Derivative Works in Source or Object form.
72 |
73 | 3. Grant of Patent License. Subject to the terms and conditions of
74 | this License, each Contributor hereby grants to You a perpetual,
75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76 | (except as stated in this section) patent license to make, have made,
77 | use, offer to sell, sell, import, and otherwise transfer the Work,
78 | where such license applies only to those patent claims licensable
79 | by such Contributor that are necessarily infringed by their
80 | Contribution(s) alone or by combination of their Contribution(s)
81 | with the Work to which such Contribution(s) was submitted. If You
82 | institute patent litigation against any entity (including a
83 | cross-claim or counterclaim in a lawsuit) alleging that the Work
84 | or a Contribution incorporated within the Work constitutes direct
85 | or contributory patent infringement, then any patent licenses
86 | granted to You under this License for that Work shall terminate
87 | as of the date such litigation is filed.
88 |
89 | 4. Redistribution. You may reproduce and distribute copies of the
90 | Work or Derivative Works thereof in any medium, with or without
91 | modifications, and in Source or Object form, provided that You
92 | meet the following conditions:
93 |
94 | (a) You must give any other recipients of the Work or
95 | Derivative Works a copy of this License; and
96 |
97 | (b) You must cause any modified files to carry prominent notices
98 | stating that You changed the files; and
99 |
100 | (c) You must retain, in the Source form of any Derivative Works
101 | that You distribute, all copyright, patent, trademark, and
102 | attribution notices from the Source form of the Work,
103 | excluding those notices that do not pertain to any part of
104 | the Derivative Works; and
105 |
106 | (d) If the Work includes a "NOTICE" text file as part of its
107 | distribution, then any Derivative Works that You distribute must
108 | include a readable copy of the attribution notices contained
109 | within such NOTICE file, excluding those notices that do not
110 | pertain to any part of the Derivative Works, in at least one
111 | of the following places: within a NOTICE text file distributed
112 | as part of the Derivative Works; within the Source form or
113 | documentation, if provided along with the Derivative Works; or,
114 | within a display generated by the Derivative Works, if and
115 | wherever such third-party notices normally appear. The contents
116 | of the NOTICE file are for informational purposes only and
117 | do not modify the License. You may add Your own attribution
118 | notices within Derivative Works that You distribute, alongside
119 | or as an addendum to the NOTICE text from the Work, provided
120 | that such additional attribution notices cannot be construed
121 | as modifying the License.
122 |
123 | You may add Your own copyright statement to Your modifications and
124 | may provide additional or different license terms and conditions
125 | for use, reproduction, or distribution of Your modifications, or
126 | for any such Derivative Works as a whole, provided Your use,
127 | reproduction, and distribution of the Work otherwise complies with
128 | the conditions stated in this License.
129 |
130 | 5. Submission of Contributions. Unless You explicitly state otherwise,
131 | any Contribution intentionally submitted for inclusion in the Work
132 | by You to the Licensor shall be under the terms and conditions of
133 | this License, without any additional terms or conditions.
134 | Notwithstanding the above, nothing herein shall supersede or modify
135 | the terms of any separate license agreement you may have executed
136 | with Licensor regarding such Contributions.
137 |
138 | 6. Trademarks. This License does not grant permission to use the trade
139 | names, trademarks, service marks, or product names of the Licensor,
140 | except as required for reasonable and customary use in describing the
141 | origin of the Work and reproducing the content of the NOTICE file.
142 |
143 | 7. Disclaimer of Warranty. Unless required by applicable law or
144 | agreed to in writing, Licensor provides the Work (and each
145 | Contributor provides its Contributions) on an "AS IS" BASIS,
146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147 | implied, including, without limitation, any warranties or conditions
148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149 | PARTICULAR PURPOSE. You are solely responsible for determining the
150 | appropriateness of using or redistributing the Work and assume any
151 | risks associated with Your exercise of permissions under this License.
152 |
153 | 8. Limitation of Liability. In no event and under no legal theory,
154 | whether in tort (including negligence), contract, or otherwise,
155 | unless required by applicable law (such as deliberate and grossly
156 | negligent acts) or agreed to in writing, shall any Contributor be
157 | liable to You for damages, including any direct, indirect, special,
158 | incidental, or consequential damages of any character arising as a
159 | result of this License or out of the use or inability to use the
160 | Work (including but not limited to damages for loss of goodwill,
161 | work stoppage, computer failure or malfunction, or any and all
162 | other commercial damages or losses), even if such Contributor
163 | has been advised of the possibility of such damages.
164 |
165 | 9. Accepting Warranty or Additional Liability. While redistributing
166 | the Work or Derivative Works thereof, You may choose to offer,
167 | and charge a fee for, acceptance of support, warranty, indemnity,
168 | or other liability obligations and/or rights consistent with this
169 | License. However, in accepting such obligations, You may act only
170 | on Your own behalf and on Your sole responsibility, not on behalf
171 | of any other Contributor, and only if You agree to indemnify,
172 | defend, and hold each Contributor harmless for any liability
173 | incurred by, or claims asserted against, such Contributor by reason
174 | of your accepting any such warranty or additional liability.
175 |
176 | END OF TERMS AND CONDITIONS
177 |
178 | APPENDIX: How to apply the Apache License to your work.
179 |
180 | To apply the Apache License to your work, attach the following
181 | boilerplate notice, with the fields enclosed by brackets "[]"
182 | replaced with your own identifying information. (Don't include
183 | the brackets!) The text should be enclosed in the appropriate
184 | comment syntax for the file format. We also recommend that a
185 | file or class name and description of purpose be included on the
186 | same "printed page" as the copyright notice for easier
187 | identification within third-party archives.
188 |
189 | Copyright 2025 Nari Labs
190 |
191 | Licensed under the Apache License, Version 2.0 (the "License");
192 | you may not use this file except in compliance with the License.
193 | You may obtain a copy of the License at
194 |
195 | http://www.apache.org/licenses/LICENSE-2.0
196 |
197 | Unless required by applicable law or agreed to in writing, software
198 | distributed under the License is distributed on an "AS IS" BASIS,
199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200 | See the License for the specific language governing permissions and
201 | limitations under the License.
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 |
2 |
3 |
5 |
6 |
7 |
11 |
12 |
15 |
16 | Dia is a 1.6B parameter text to speech model created by Nari Labs.
17 |
18 | 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.
19 |
20 | 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). The model only supports English generation at the moment.
21 |
22 | 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).
23 |
24 | - (Update) We have a ZeroGPU Space running! Try it now [here](https://huggingface.co/spaces/nari-labs/Dia-1.6B). Thanks to the HF team for the support :)
25 | - Join our [discord server](https://discord.gg/pgdB5YRe) for community support and access to new features.
26 | - 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.
27 |
28 | ## ⚡️ Quickstart
29 |
30 | ### Install via pip
31 |
32 | ```bash
33 | # Install directly from GitHub
34 | pip install git+https://github.com/nari-labs/dia.git
35 | ```
36 |
37 | ### Run the Gradio UI
38 |
39 | This will open a Gradio UI that you can work on.
40 |
41 | ```bash
42 | git clone https://github.com/nari-labs/dia.git
43 | cd dia && uv run app.py
44 | ```
45 |
46 | or if you do not have `uv` pre-installed:
47 |
48 | ```bash
49 | git clone https://github.com/nari-labs/dia.git
50 | cd dia
51 | python -m venv .venv
52 | source .venv/bin/activate
53 | pip install -e .
54 | python app.py
55 | ```
56 |
57 | Note that the model was not fine-tuned on a specific voice. Hence, you will get different voices every time you run the model.
58 | You can keep speaker consistency by either adding an audio prompt (a guide coming VERY soon - try it with the second example on Gradio for now), or fixing the seed.
59 |
60 | ## Features
61 |
62 | - Generate dialogue via `[S1]` and `[S2]` tag
63 | - Generate non-verbal like `(laughs)`, `(coughs)`, etc.
64 | - Below verbal tags will be recognized, but might result in unexpected output.
65 | - `(laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)`
66 | - Voice cloning. See [`example/voice_clone.py`](example/voice_clone.py) for more information.
67 | - 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.
68 |
69 | ## ⚙️ Usage
70 |
71 | ### As a Python Library
72 |
73 | ```python
74 | import soundfile as sf
75 |
76 | from dia.model import Dia
77 |
78 |
79 | model = Dia.from_pretrained("nari-labs/Dia-1.6B")
80 |
81 | 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."
82 |
83 | output = model.generate(text)
84 |
85 | sf.write("simple.mp3", output, 44100)
86 | ```
87 |
88 | A pypi package and a working CLI tool will be available soon.
89 |
90 | ## 💻 Hardware and Inference Speed
91 |
92 | Dia has been tested on only GPUs (pytorch 2.0+, CUDA 12.6). CPU support is to be added soon.
93 | The initial run will take longer as the Descript Audio Codec also needs to be downloaded.
94 |
95 | On enterprise GPUs, Dia can generate audio in real-time. On older GPUs, inference time will be slower.
96 | For reference, on a A4000 GPU, Dia roughly generates 40 tokens/s (86 tokens equals 1 second of audio).
97 | `torch.compile` will increase speeds for supported GPUs.
98 |
99 | The full version of Dia requires around 12-13GB of VRAM to run. We will be adding a quantized version in the future.
100 |
101 | 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).
102 |
103 | ## 🪪 License
104 |
105 | This project is licensed under the Apache License 2.0 - see the [LICENSE](LICENSE) file for details.
106 |
107 | ## ⚠️ Disclaimer
108 |
109 | This project offers a high-fidelity speech generation model intended for research and educational use. The following uses are **strictly forbidden**:
110 |
111 | - **Identity Misuse**: Do not produce audio resembling real individuals without permission.
112 | - **Deceptive Content**: Do not use this model to generate misleading content (e.g. fake news)
113 | - **Illegal or Malicious Use**: Do not use this model for activities that are illegal or intended to cause harm.
114 |
115 | 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.
116 |
117 | ## 🔭 TODO / Future Work
118 |
119 | - Docker support for ARM architecture and MacOS.
120 | - Optimize inference speed.
121 | - Add quantization for memory efficiency.
122 |
123 | ## 🤝 Contributing
124 |
125 | We are a tiny team of 1 full-time and 1 part-time research-engineers. We are extra-welcome to any contributions!
126 | Join our [Discord Server](https://discord.gg/pgdB5YRe) for discussions.
127 |
128 | ## 🤗 Acknowledgements
129 |
130 | - We thank the [Google TPU Research Cloud program](https://sites.research.google/trc/about/) for providing computation resources.
131 | - 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).
132 | - Hugging Face for providing the ZeroGPU Grant.
133 | - "Nari" is a pure Korean word for lily.
134 | - We thank Jason Y. for providing help with data filtering.
135 |
136 |
137 | ## ⭐ Star History
138 |
139 |
140 |
141 |
145 |
146 |
--------------------------------------------------------------------------------
/app.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import tempfile
3 | import time
4 | from pathlib import Path
5 | from typing import Optional, Tuple
6 |
7 | import gradio as gr
8 | import numpy as np
9 | import soundfile as sf
10 | import torch
11 |
12 | from dia.model import Dia
13 |
14 |
15 | # --- Global Setup ---
16 | parser = argparse.ArgumentParser(description="Gradio interface for Nari TTS")
17 | parser.add_argument("--device", type=str, default=None, help="Force device (e.g., 'cuda', 'mps', 'cpu')")
18 | parser.add_argument("--share", action="store_true", help="Enable Gradio sharing")
19 |
20 | args = parser.parse_args()
21 |
22 |
23 | # Determine device
24 | if args.device:
25 | device = torch.device(args.device)
26 | elif torch.cuda.is_available():
27 | device = torch.device("cuda")
28 | # Simplified MPS check for broader compatibility
29 | elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
30 | # Basic check is usually sufficient, detailed check can be problematic
31 | device = torch.device("mps")
32 | else:
33 | device = torch.device("cpu")
34 |
35 | print(f"Using device: {device}")
36 |
37 | # Load Nari model and config
38 | print("Loading Nari model...")
39 | try:
40 | # Use the function from inference.py
41 | model = Dia.from_pretrained("nari-labs/Dia-1.6B", device=device)
42 | except Exception as e:
43 | print(f"Error loading Nari model: {e}")
44 | raise
45 |
46 |
47 | def run_inference(
48 | text_input: str,
49 | audio_prompt_input: Optional[Tuple[int, np.ndarray]],
50 | max_new_tokens: int,
51 | cfg_scale: float,
52 | temperature: float,
53 | top_p: float,
54 | cfg_filter_top_k: int,
55 | speed_factor: float,
56 | ):
57 | """
58 | Runs Nari inference using the globally loaded model and provided inputs.
59 | Uses temporary files for text and audio prompt compatibility with inference.generate.
60 | """
61 | global model, device # Access global model, config, device
62 |
63 | if not text_input or text_input.isspace():
64 | raise gr.Error("Text input cannot be empty.")
65 |
66 | temp_txt_file_path = None
67 | temp_audio_prompt_path = None
68 | output_audio = (44100, np.zeros(1, dtype=np.float32))
69 |
70 | try:
71 | prompt_path_for_generate = None
72 | if audio_prompt_input is not None:
73 | sr, audio_data = audio_prompt_input
74 | # Check if audio_data is valid
75 | if audio_data is None or audio_data.size == 0 or audio_data.max() == 0: # Check for silence/empty
76 | gr.Warning("Audio prompt seems empty or silent, ignoring prompt.")
77 | else:
78 | # Save prompt audio to a temporary WAV file
79 | with tempfile.NamedTemporaryFile(mode="wb", suffix=".wav", delete=False) as f_audio:
80 | temp_audio_prompt_path = f_audio.name # Store path for cleanup
81 |
82 | # Basic audio preprocessing for consistency
83 | # Convert to float32 in [-1, 1] range if integer type
84 | if np.issubdtype(audio_data.dtype, np.integer):
85 | max_val = np.iinfo(audio_data.dtype).max
86 | audio_data = audio_data.astype(np.float32) / max_val
87 | elif not np.issubdtype(audio_data.dtype, np.floating):
88 | gr.Warning(f"Unsupported audio prompt dtype {audio_data.dtype}, attempting conversion.")
89 | # Attempt conversion, might fail for complex types
90 | try:
91 | audio_data = audio_data.astype(np.float32)
92 | except Exception as conv_e:
93 | raise gr.Error(f"Failed to convert audio prompt to float32: {conv_e}")
94 |
95 | # Ensure mono (average channels if stereo)
96 | if audio_data.ndim > 1:
97 | if audio_data.shape[0] == 2: # Assume (2, N)
98 | audio_data = np.mean(audio_data, axis=0)
99 | elif audio_data.shape[1] == 2: # Assume (N, 2)
100 | audio_data = np.mean(audio_data, axis=1)
101 | else:
102 | gr.Warning(
103 | f"Audio prompt has unexpected shape {audio_data.shape}, taking first channel/axis."
104 | )
105 | audio_data = (
106 | audio_data[0] if audio_data.shape[0] < audio_data.shape[1] else audio_data[:, 0]
107 | )
108 | audio_data = np.ascontiguousarray(audio_data) # Ensure contiguous after slicing/mean
109 |
110 | # Write using soundfile
111 | try:
112 | sf.write(
113 | temp_audio_prompt_path, audio_data, sr, subtype="FLOAT"
114 | ) # Explicitly use FLOAT subtype
115 | prompt_path_for_generate = temp_audio_prompt_path
116 | print(f"Created temporary audio prompt file: {temp_audio_prompt_path} (orig sr: {sr})")
117 | except Exception as write_e:
118 | print(f"Error writing temporary audio file: {write_e}")
119 | raise gr.Error(f"Failed to save audio prompt: {write_e}")
120 |
121 | # 3. Run Generation
122 |
123 | start_time = time.time()
124 |
125 | # Use torch.inference_mode() context manager for the generation call
126 | with torch.inference_mode():
127 | output_audio_np = model.generate(
128 | text_input,
129 | max_tokens=max_new_tokens,
130 | cfg_scale=cfg_scale,
131 | temperature=temperature,
132 | top_p=top_p,
133 | cfg_filter_top_k=cfg_filter_top_k, # Pass the value here
134 | use_torch_compile=False, # Keep False for Gradio stability
135 | audio_prompt=prompt_path_for_generate,
136 | )
137 |
138 | end_time = time.time()
139 | print(f"Generation finished in {end_time - start_time:.2f} seconds.")
140 |
141 | # 4. Convert Codes to Audio
142 | if output_audio_np is not None:
143 | # Get sample rate from the loaded DAC model
144 | output_sr = 44100
145 |
146 | # --- Slow down audio ---
147 | original_len = len(output_audio_np)
148 | # Ensure speed_factor is positive and not excessively small/large to avoid issues
149 | speed_factor = max(0.1, min(speed_factor, 5.0))
150 | target_len = int(original_len / speed_factor) # Target length based on speed_factor
151 | if target_len != original_len and target_len > 0: # Only interpolate if length changes and is valid
152 | x_original = np.arange(original_len)
153 | x_resampled = np.linspace(0, original_len - 1, target_len)
154 | resampled_audio_np = np.interp(x_resampled, x_original, output_audio_np)
155 | output_audio = (
156 | output_sr,
157 | resampled_audio_np.astype(np.float32),
158 | ) # Use resampled audio
159 | print(f"Resampled audio from {original_len} to {target_len} samples for {speed_factor:.2f}x speed.")
160 | else:
161 | output_audio = (
162 | output_sr,
163 | output_audio_np,
164 | ) # Keep original if calculation fails or no change
165 | print(f"Skipping audio speed adjustment (factor: {speed_factor:.2f}).")
166 | # --- End slowdown ---
167 |
168 | print(f"Audio conversion successful. Final shape: {output_audio[1].shape}, Sample Rate: {output_sr}")
169 |
170 | # Explicitly convert to int16 to prevent Gradio warning
171 | if output_audio[1].dtype == np.float32 or output_audio[1].dtype == np.float64:
172 | audio_for_gradio = np.clip(output_audio[1], -1.0, 1.0)
173 | audio_for_gradio = (audio_for_gradio * 32767).astype(np.int16)
174 | output_audio = (output_sr, audio_for_gradio)
175 | print("Converted audio to int16 for Gradio output.")
176 |
177 | else:
178 | print("\nGeneration finished, but no valid tokens were produced.")
179 | # Return default silence
180 | gr.Warning("Generation produced no output.")
181 |
182 | except Exception as e:
183 | print(f"Error during inference: {e}")
184 | import traceback
185 |
186 | traceback.print_exc()
187 | # Re-raise as Gradio error to display nicely in the UI
188 | raise gr.Error(f"Inference failed: {e}")
189 |
190 | finally:
191 | # 5. Cleanup Temporary Files defensively
192 | if temp_txt_file_path and Path(temp_txt_file_path).exists():
193 | try:
194 | Path(temp_txt_file_path).unlink()
195 | print(f"Deleted temporary text file: {temp_txt_file_path}")
196 | except OSError as e:
197 | print(f"Warning: Error deleting temporary text file {temp_txt_file_path}: {e}")
198 | if temp_audio_prompt_path and Path(temp_audio_prompt_path).exists():
199 | try:
200 | Path(temp_audio_prompt_path).unlink()
201 | print(f"Deleted temporary audio prompt file: {temp_audio_prompt_path}")
202 | except OSError as e:
203 | print(f"Warning: Error deleting temporary audio prompt file {temp_audio_prompt_path}: {e}")
204 |
205 | return output_audio
206 |
207 |
208 | # --- Create Gradio Interface ---
209 | css = """
210 | #col-container {max-width: 90%; margin-left: auto; margin-right: auto;}
211 | """
212 | # Attempt to load default text from example.txt
213 | 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."
214 | example_txt_path = Path("./example.txt")
215 | if example_txt_path.exists():
216 | try:
217 | default_text = example_txt_path.read_text(encoding="utf-8").strip()
218 | if not default_text: # Handle empty example file
219 | default_text = "Example text file was empty."
220 | except Exception as e:
221 | print(f"Warning: Could not read example.txt: {e}")
222 |
223 |
224 | # Build Gradio UI
225 | with gr.Blocks(css=css) as demo:
226 | gr.Markdown("# Nari Text-to-Speech Synthesis")
227 |
228 | with gr.Row(equal_height=False):
229 | with gr.Column(scale=1):
230 | text_input = gr.Textbox(
231 | label="Input Text",
232 | placeholder="Enter text here...",
233 | value=default_text,
234 | lines=5, # Increased lines
235 | )
236 | audio_prompt_input = gr.Audio(
237 | label="Audio Prompt (Optional)",
238 | show_label=True,
239 | sources=["upload", "microphone"],
240 | type="numpy",
241 | )
242 | with gr.Accordion("Generation Parameters", open=False):
243 | max_new_tokens = gr.Slider(
244 | label="Max New Tokens (Audio Length)",
245 | minimum=860,
246 | maximum=3072,
247 | value=model.config.data.audio_length, # Use config default if available, else fallback
248 | step=50,
249 | info="Controls the maximum length of the generated audio (more tokens = longer audio).",
250 | )
251 | cfg_scale = gr.Slider(
252 | label="CFG Scale (Guidance Strength)",
253 | minimum=1.0,
254 | maximum=5.0,
255 | value=3.0, # Default from inference.py
256 | step=0.1,
257 | info="Higher values increase adherence to the text prompt.",
258 | )
259 | temperature = gr.Slider(
260 | label="Temperature (Randomness)",
261 | minimum=1.0,
262 | maximum=1.5,
263 | value=1.3, # Default from inference.py
264 | step=0.05,
265 | info="Lower values make the output more deterministic, higher values increase randomness.",
266 | )
267 | top_p = gr.Slider(
268 | label="Top P (Nucleus Sampling)",
269 | minimum=0.80,
270 | maximum=1.0,
271 | value=0.95, # Default from inference.py
272 | step=0.01,
273 | info="Filters vocabulary to the most likely tokens cumulatively reaching probability P.",
274 | )
275 | cfg_filter_top_k = gr.Slider(
276 | label="CFG Filter Top K",
277 | minimum=15,
278 | maximum=50,
279 | value=30,
280 | step=1,
281 | info="Top k filter for CFG guidance.",
282 | )
283 | speed_factor_slider = gr.Slider(
284 | label="Speed Factor",
285 | minimum=0.8,
286 | maximum=1.0,
287 | value=0.94,
288 | step=0.02,
289 | info="Adjusts the speed of the generated audio (1.0 = original speed).",
290 | )
291 |
292 | run_button = gr.Button("Generate Audio", variant="primary")
293 |
294 | with gr.Column(scale=1):
295 | audio_output = gr.Audio(
296 | label="Generated Audio",
297 | type="numpy",
298 | autoplay=False,
299 | )
300 |
301 | # Link button click to function
302 | run_button.click(
303 | fn=run_inference,
304 | inputs=[
305 | text_input,
306 | audio_prompt_input,
307 | max_new_tokens,
308 | cfg_scale,
309 | temperature,
310 | top_p,
311 | cfg_filter_top_k,
312 | speed_factor_slider,
313 | ],
314 | outputs=[audio_output], # Add status_output here if using it
315 | api_name="generate_audio",
316 | )
317 |
318 | # Add examples (ensure the prompt path is correct or remove it if example file doesn't exist)
319 | example_prompt_path = "./example_prompt.mp3" # Adjust if needed
320 | examples_list = [
321 | [
322 | "[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! ",
323 | None,
324 | 3072,
325 | 3.0,
326 | 1.3,
327 | 0.95,
328 | 35,
329 | 0.94,
330 | ],
331 | [
332 | "[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.",
333 | example_prompt_path if Path(example_prompt_path).exists() else None,
334 | 3072,
335 | 3.0,
336 | 1.3,
337 | 0.95,
338 | 35,
339 | 0.94,
340 | ],
341 | ]
342 |
343 | if examples_list:
344 | gr.Examples(
345 | examples=examples_list,
346 | inputs=[
347 | text_input,
348 | audio_prompt_input,
349 | max_new_tokens,
350 | cfg_scale,
351 | temperature,
352 | top_p,
353 | cfg_filter_top_k,
354 | speed_factor_slider,
355 | ],
356 | outputs=[audio_output],
357 | fn=run_inference,
358 | cache_examples=False,
359 | label="Examples (Click to Run)",
360 | )
361 | else:
362 | gr.Markdown("_(No examples configured or example prompt file missing)_")
363 |
364 | # --- Launch the App ---
365 | if __name__ == "__main__":
366 | print("Launching Gradio interface...")
367 |
368 | # set `GRADIO_SERVER_NAME`, `GRADIO_SERVER_PORT` env vars to override default values
369 | # use `GRADIO_SERVER_NAME=0.0.0.0` for Docker
370 | demo.launch(share=args.share)
371 |
--------------------------------------------------------------------------------
/cli.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import random
4 |
5 | import numpy as np
6 | import soundfile as sf
7 | import torch
8 |
9 | from dia.model import Dia
10 |
11 |
12 | def set_seed(seed: int):
13 | """Sets the random seed for reproducibility."""
14 | random.seed(seed)
15 | np.random.seed(seed)
16 | torch.manual_seed(seed)
17 | if torch.cuda.is_available():
18 | torch.cuda.manual_seed(seed)
19 | torch.cuda.manual_seed_all(seed)
20 | # Ensure deterministic behavior for cuDNN (if used)
21 | torch.backends.cudnn.deterministic = True
22 | torch.backends.cudnn.benchmark = False
23 |
24 |
25 | def main():
26 | parser = argparse.ArgumentParser(description="Generate audio using the Dia model.")
27 |
28 | parser.add_argument("text", type=str, help="Input text for speech generation.")
29 | parser.add_argument(
30 | "--output", type=str, required=True, help="Path to save the generated audio file (e.g., output.wav)."
31 | )
32 |
33 | parser.add_argument(
34 | "--repo-id",
35 | type=str,
36 | default="nari-labs/Dia-1.6B",
37 | help="Hugging Face repository ID (e.g., nari-labs/Dia-1.6B).",
38 | )
39 | parser.add_argument(
40 | "--local-paths", action="store_true", help="Load model from local config and checkpoint files."
41 | )
42 |
43 | parser.add_argument(
44 | "--config", type=str, help="Path to local config.json file (required if --local-paths is set)."
45 | )
46 | parser.add_argument(
47 | "--checkpoint", type=str, help="Path to local model checkpoint .pth file (required if --local-paths is set)."
48 | )
49 | parser.add_argument(
50 | "--audio-prompt", type=str, default=None, help="Path to an optional audio prompt WAV file for voice cloning."
51 | )
52 |
53 | gen_group = parser.add_argument_group("Generation Parameters")
54 | gen_group.add_argument(
55 | "--max-tokens",
56 | type=int,
57 | default=None,
58 | help="Maximum number of audio tokens to generate (defaults to config value).",
59 | )
60 | gen_group.add_argument(
61 | "--cfg-scale", type=float, default=3.0, help="Classifier-Free Guidance scale (default: 3.0)."
62 | )
63 | gen_group.add_argument(
64 | "--temperature", type=float, default=1.3, help="Sampling temperature (higher is more random, default: 0.7)."
65 | )
66 | gen_group.add_argument("--top-p", type=float, default=0.95, help="Nucleus sampling probability (default: 0.95).")
67 |
68 | infra_group = parser.add_argument_group("Infrastructure")
69 | infra_group.add_argument("--seed", type=int, default=None, help="Random seed for reproducibility.")
70 | infra_group.add_argument(
71 | "--device",
72 | type=str,
73 | default="cuda" if torch.cuda.is_available() else "cpu",
74 | help="Device to run inference on (e.g., 'cuda', 'cpu', default: auto).",
75 | )
76 |
77 | args = parser.parse_args()
78 |
79 | # Validation for local paths
80 | if args.local_paths:
81 | if not args.config:
82 | parser.error("--config is required when --local-paths is set.")
83 | if not args.checkpoint:
84 | parser.error("--checkpoint is required when --local-paths is set.")
85 | if not os.path.exists(args.config):
86 | parser.error(f"Config file not found: {args.config}")
87 | if not os.path.exists(args.checkpoint):
88 | parser.error(f"Checkpoint file not found: {args.checkpoint}")
89 |
90 | # Set seed if provided
91 | if args.seed is not None:
92 | set_seed(args.seed)
93 | print(f"Using random seed: {args.seed}")
94 |
95 | # Determine device
96 | device = torch.device(args.device)
97 | print(f"Using device: {device}")
98 |
99 | # Load model
100 | print("Loading model...")
101 | if args.local_paths:
102 | print(f"Loading from local paths: config='{args.config}', checkpoint='{args.checkpoint}'")
103 | try:
104 | model = Dia.from_local(args.config, args.checkpoint, device=device)
105 | except Exception as e:
106 | print(f"Error loading local model: {e}")
107 | exit(1)
108 | else:
109 | print(f"Loading from Hugging Face Hub: repo_id='{args.repo_id}'")
110 | try:
111 | model = Dia.from_pretrained(args.repo_id, device=device)
112 | except Exception as e:
113 | print(f"Error loading model from Hub: {e}")
114 | exit(1)
115 | print("Model loaded.")
116 |
117 | # Generate audio
118 | print("Generating audio...")
119 | try:
120 | sample_rate = 44100 # Default assumption
121 |
122 | output_audio = model.generate(
123 | text=args.text,
124 | audio_prompt=args.audio_prompt,
125 | max_tokens=args.max_tokens,
126 | cfg_scale=args.cfg_scale,
127 | temperature=args.temperature,
128 | top_p=args.top_p,
129 | )
130 | print("Audio generation complete.")
131 |
132 | print(f"Saving audio to {args.output}...")
133 | os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
134 |
135 | sf.write(args.output, output_audio, sample_rate)
136 | print(f"Audio successfully saved to {args.output}")
137 |
138 | except Exception as e:
139 | print(f"Error during audio generation or saving: {e}")
140 | exit(1)
141 |
142 |
143 | if __name__ == "__main__":
144 | main()
145 |
--------------------------------------------------------------------------------
/dia/__init__.py:
--------------------------------------------------------------------------------
1 | from .model import Dia
2 |
3 |
4 | __all__ = [
5 | "Dia",
6 | ]
7 |
--------------------------------------------------------------------------------
/dia/audio.py:
--------------------------------------------------------------------------------
1 | import typing as tp
2 |
3 | import torch
4 |
5 |
6 | def build_delay_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
7 | """
8 | Precompute (t_idx_BxTxC, indices_BTCx3) so that out[t, c] = in[t - delay[c], c].
9 | Negative t_idx => BOS; t_idx >= T => PAD.
10 | """
11 | delay_arr = torch.tensor(delay_pattern, dtype=torch.int32)
12 |
13 | t_idx_BxT = torch.broadcast_to(
14 | torch.arange(T, dtype=torch.int32)[None, :],
15 | [B, T],
16 | )
17 | t_idx_BxTx1 = t_idx_BxT[..., None]
18 | t_idx_BxTxC = t_idx_BxTx1 - delay_arr.view(1, 1, C)
19 |
20 | b_idx_BxTxC = torch.broadcast_to(
21 | torch.arange(B, dtype=torch.int32).view(B, 1, 1),
22 | [B, T, C],
23 | )
24 | c_idx_BxTxC = torch.broadcast_to(
25 | torch.arange(C, dtype=torch.int32).view(1, 1, C),
26 | [B, T, C],
27 | )
28 |
29 | # We must clamp time indices to [0..T-1] so gather_nd equivalent won't fail
30 | t_clamped_BxTxC = torch.clamp(t_idx_BxTxC, 0, T - 1)
31 |
32 | indices_BTCx3 = torch.stack(
33 | [
34 | b_idx_BxTxC.reshape(-1),
35 | t_clamped_BxTxC.reshape(-1),
36 | c_idx_BxTxC.reshape(-1),
37 | ],
38 | dim=1,
39 | ).long() # Ensure indices are long type for indexing
40 |
41 | return t_idx_BxTxC, indices_BTCx3
42 |
43 |
44 | def apply_audio_delay(
45 | audio_BxTxC: torch.Tensor,
46 | pad_value: int,
47 | bos_value: int,
48 | precomp: tp.Tuple[torch.Tensor, torch.Tensor],
49 | ) -> torch.Tensor:
50 | """
51 | Applies the delay pattern to batched audio tokens using precomputed indices,
52 | inserting BOS where t_idx < 0 and PAD where t_idx >= T.
53 |
54 | Args:
55 | audio_BxTxC: [B, T, C] int16 audio tokens (or int32/float)
56 | pad_value: the padding token
57 | bos_value: the BOS token
58 | precomp: (t_idx_BxTxC, indices_BTCx3) from build_delay_indices
59 |
60 | Returns:
61 | result_BxTxC: [B, T, C] delayed audio tokens
62 | """
63 | device = audio_BxTxC.device # Get device from input tensor
64 | t_idx_BxTxC, indices_BTCx3 = precomp
65 | t_idx_BxTxC = t_idx_BxTxC.to(device) # Move precomputed indices to device
66 | indices_BTCx3 = indices_BTCx3.to(device)
67 |
68 | # Equivalent of tf.gather_nd using advanced indexing
69 | # Ensure indices are long type if not already (build_delay_indices should handle this)
70 | gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
71 | gathered_BxTxC = gathered_flat.view(audio_BxTxC.shape)
72 |
73 | # Create masks on the correct device
74 | mask_bos = t_idx_BxTxC < 0 # => place bos_value
75 | mask_pad = t_idx_BxTxC >= audio_BxTxC.shape[1] # => place pad_value
76 |
77 | # Create scalar tensors on the correct device
78 | bos_tensor = torch.tensor(bos_value, dtype=audio_BxTxC.dtype, device=device)
79 | pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
80 |
81 | # If mask_bos, BOS; else if mask_pad, PAD; else original gather
82 | # All tensors should now be on the same device
83 | result_BxTxC = torch.where(mask_bos, bos_tensor, torch.where(mask_pad, pad_tensor, gathered_BxTxC))
84 |
85 | return result_BxTxC
86 |
87 |
88 | def build_revert_indices(B: int, T: int, C: int, delay_pattern: tp.List[int]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
89 | """
90 | Precompute indices for the revert operation using PyTorch.
91 |
92 | Returns:
93 | A tuple (t_idx_BxTxC, indices_BTCx3) where:
94 | - t_idx_BxTxC is a tensor of shape [B, T, C] computed as time indices plus the delay.
95 | - indices_BTCx3 is a tensor of shape [B*T*C, 3] used for gathering, computed from:
96 | batch indices, clamped time indices, and channel indices.
97 | """
98 | # Use default device unless specified otherwise; assumes inputs might define device later
99 | device = None # Or determine dynamically if needed, e.g., from a model parameter
100 |
101 | delay_arr = torch.tensor(delay_pattern, dtype=torch.int32, device=device)
102 |
103 | t_idx_BT1 = torch.broadcast_to(torch.arange(T, device=device).unsqueeze(0), [B, T])
104 | t_idx_BT1 = t_idx_BT1.unsqueeze(-1)
105 |
106 | t_idx_BxTxC = torch.minimum(
107 | t_idx_BT1 + delay_arr.view(1, 1, C),
108 | torch.tensor(T - 1, device=device),
109 | )
110 | b_idx_BxTxC = torch.broadcast_to(torch.arange(B, device=device).view(B, 1, 1), [B, T, C])
111 | c_idx_BxTxC = torch.broadcast_to(torch.arange(C, device=device).view(1, 1, C), [B, T, C])
112 |
113 | indices_BTCx3 = torch.stack(
114 | [
115 | b_idx_BxTxC.reshape(-1),
116 | t_idx_BxTxC.reshape(-1),
117 | c_idx_BxTxC.reshape(-1),
118 | ],
119 | axis=1,
120 | ).long() # Ensure indices are long type
121 |
122 | return t_idx_BxTxC, indices_BTCx3
123 |
124 |
125 | def revert_audio_delay(
126 | audio_BxTxC: torch.Tensor,
127 | pad_value: int,
128 | precomp: tp.Tuple[torch.Tensor, torch.Tensor],
129 | T: int,
130 | ) -> torch.Tensor:
131 | """
132 | Reverts a delay pattern from batched audio tokens using precomputed indices (PyTorch version).
133 |
134 | Args:
135 | audio_BxTxC: Input delayed audio tensor
136 | pad_value: Padding value for out-of-bounds indices
137 | precomp: Precomputed revert indices tuple containing:
138 | - t_idx_BxTxC: Time offset indices tensor
139 | - indices_BTCx3: Gather indices tensor for original audio
140 | T: Original sequence length before padding
141 |
142 | Returns:
143 | Reverted audio tensor with same shape as input
144 | """
145 | t_idx_BxTxC, indices_BTCx3 = precomp
146 | device = audio_BxTxC.device # Get device from input tensor
147 |
148 | # Move precomputed indices to the same device as audio_BxTxC if they aren't already
149 | t_idx_BxTxC = t_idx_BxTxC.to(device)
150 | indices_BTCx3 = indices_BTCx3.to(device)
151 |
152 | # Using PyTorch advanced indexing (equivalent to tf.gather_nd or np equivalent)
153 | gathered_flat = audio_BxTxC[indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]]
154 | gathered_BxTxC = gathered_flat.view(audio_BxTxC.size()) # Use .size() for robust reshaping
155 |
156 | # Create pad_tensor on the correct device
157 | pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
158 | # Create T tensor on the correct device for comparison
159 | T_tensor = torch.tensor(T, device=device)
160 |
161 | result_BxTxC = torch.where(t_idx_BxTxC >= T_tensor, pad_tensor, gathered_BxTxC) # Changed np.where to torch.where
162 |
163 | return result_BxTxC
164 |
165 |
166 | @torch.no_grad()
167 | @torch.inference_mode()
168 | def decode(
169 | model,
170 | audio_codes,
171 | ):
172 | """
173 | Decodes the given frames into an output audio waveform
174 | """
175 | if len(audio_codes) != 1:
176 | raise ValueError(f"Expected one frame, got {len(audio_codes)}")
177 |
178 | try:
179 | audio_values = model.quantizer.from_codes(audio_codes)
180 | audio_values = model.decode(audio_values[0])
181 |
182 | return audio_values
183 | except Exception as e:
184 | print(f"Error in decode method: {str(e)}")
185 | raise
186 |
--------------------------------------------------------------------------------
/dia/config.py:
--------------------------------------------------------------------------------
1 | """Configuration management module for the Dia model.
2 |
3 | This module provides comprehensive configuration management for the Dia model,
4 | utilizing Pydantic for validation. It defines configurations for data processing,
5 | model architecture (encoder and decoder), and training settings.
6 |
7 | Key components:
8 | - DataConfig: Parameters for data loading and preprocessing.
9 | - EncoderConfig: Architecture details for the encoder module.
10 | - DecoderConfig: Architecture details for the decoder module.
11 | - ModelConfig: Combined model architecture settings.
12 | - TrainingConfig: Training hyperparameters and settings.
13 | - DiaConfig: Master configuration combining all components.
14 | """
15 |
16 | import os
17 | from typing import Annotated
18 |
19 | from pydantic import BaseModel, BeforeValidator, Field
20 |
21 |
22 | class DataConfig(BaseModel, frozen=True):
23 | """Configuration for data loading and preprocessing.
24 |
25 | Attributes:
26 | text_length: Maximum length of text sequences (must be multiple of 128).
27 | audio_length: Maximum length of audio sequences (must be multiple of 128).
28 | channels: Number of audio channels.
29 | text_pad_value: Value used for padding text sequences.
30 | audio_eos_value: Value representing the end of audio sequences.
31 | audio_bos_value: Value representing the beginning of audio sequences.
32 | audio_pad_value: Value used for padding audio sequences.
33 | delay_pattern: List of delay values for each audio channel.
34 | """
35 |
36 | text_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = Field(gt=0, multiple_of=128)
37 | audio_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = Field(gt=0, multiple_of=128)
38 | channels: int = Field(default=9, gt=0, multiple_of=1)
39 | text_pad_value: int = Field(default=0)
40 | audio_eos_value: int = Field(default=1024)
41 | audio_pad_value: int = Field(default=1025)
42 | audio_bos_value: int = Field(default=1026)
43 | delay_pattern: list[Annotated[int, Field(ge=0)]] = Field(default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15])
44 |
45 | def __hash__(self) -> int:
46 | """Generate a hash based on all fields of the config."""
47 | return hash(
48 | (
49 | self.text_length,
50 | self.audio_length,
51 | self.channels,
52 | self.text_pad_value,
53 | self.audio_pad_value,
54 | self.audio_bos_value,
55 | self.audio_eos_value,
56 | tuple(self.delay_pattern),
57 | )
58 | )
59 |
60 |
61 | class EncoderConfig(BaseModel, frozen=True):
62 | """Configuration for the encoder component of the Dia model.
63 |
64 | Attributes:
65 | n_layer: Number of transformer layers.
66 | n_embd: Embedding dimension.
67 | n_hidden: Hidden dimension size in the MLP layers.
68 | n_head: Number of attention heads.
69 | head_dim: Dimension per attention head.
70 | """
71 |
72 | n_layer: int = Field(gt=0)
73 | n_embd: int = Field(gt=0)
74 | n_hidden: int = Field(gt=0)
75 | n_head: int = Field(gt=0)
76 | head_dim: int = Field(gt=0)
77 |
78 |
79 | class DecoderConfig(BaseModel, frozen=True):
80 | """Configuration for the decoder component of the Dia model.
81 |
82 | Attributes:
83 | n_layer: Number of transformer layers.
84 | n_embd: Embedding dimension.
85 | n_hidden: Hidden dimension size in the MLP layers.
86 | gqa_query_heads: Number of query heads for grouped-query self-attention.
87 | kv_heads: Number of key/value heads for grouped-query self-attention.
88 | gqa_head_dim: Dimension per query head for grouped-query self-attention.
89 | cross_query_heads: Number of query heads for cross-attention.
90 | cross_head_dim: Dimension per cross-attention head.
91 | """
92 |
93 | n_layer: int = Field(gt=0)
94 | n_embd: int = Field(gt=0)
95 | n_hidden: int = Field(gt=0)
96 | gqa_query_heads: int = Field(gt=0)
97 | kv_heads: int = Field(gt=0)
98 | gqa_head_dim: int = Field(gt=0)
99 | cross_query_heads: int = Field(gt=0)
100 | cross_head_dim: int = Field(gt=0)
101 |
102 |
103 | class ModelConfig(BaseModel, frozen=True):
104 | """Main configuration container for the Dia model architecture.
105 |
106 | Attributes:
107 | encoder: Configuration for the encoder component.
108 | decoder: Configuration for the decoder component.
109 | src_vocab_size: Size of the source (text) vocabulary.
110 | tgt_vocab_size: Size of the target (audio code) vocabulary.
111 | dropout: Dropout probability applied within the model.
112 | normalization_layer_epsilon: Epsilon value for normalization layers (e.g., LayerNorm).
113 | weight_dtype: Data type for model weights (e.g., "float32", "bfloat16").
114 | rope_min_timescale: Minimum timescale for Rotary Positional Embeddings (RoPE).
115 | rope_max_timescale: Maximum timescale for Rotary Positional Embeddings (RoPE).
116 | """
117 |
118 | encoder: EncoderConfig
119 | decoder: DecoderConfig
120 | src_vocab_size: int = Field(default=128, gt=0)
121 | tgt_vocab_size: int = Field(default=1028, gt=0)
122 | dropout: float = Field(default=0.0, ge=0.0, lt=1.0)
123 | normalization_layer_epsilon: float = Field(default=1.0e-5, ge=0.0)
124 | weight_dtype: str = Field(default="float32", description="Weight precision")
125 | rope_min_timescale: int = Field(default=1, description="Timescale For global Attention")
126 | rope_max_timescale: int = Field(default=10_000, description="Timescale For global Attention")
127 |
128 |
129 | class TrainingConfig(BaseModel, frozen=True):
130 | pass
131 |
132 |
133 | class DiaConfig(BaseModel, frozen=True):
134 | """Master configuration for the Dia model.
135 |
136 | Combines all sub-configurations into a single validated object.
137 |
138 | Attributes:
139 | version: Configuration version string.
140 | model: Model architecture configuration.
141 | training: Training process configuration (precision settings).
142 | data: Data loading and processing configuration.
143 | """
144 |
145 | version: str = Field(default="1.0")
146 | model: ModelConfig
147 | # TODO: remove training. this is just for backwards-compatability
148 | training: TrainingConfig
149 | data: DataConfig
150 |
151 | def save(self, path: str) -> None:
152 | """Save the current configuration instance to a JSON file.
153 |
154 | Ensures the parent directory exists and the file has a .json extension.
155 |
156 | Args:
157 | path: The target file path to save the configuration.
158 |
159 | Raises:
160 | ValueError: If the path is not a file with a .json extension.
161 | """
162 | os.makedirs(os.path.dirname(path), exist_ok=True)
163 | config_json = self.model_dump_json(indent=2)
164 | with open(path, "w") as f:
165 | f.write(config_json)
166 |
167 | @classmethod
168 | def load(cls, path: str) -> "DiaConfig | None":
169 | """Load and validate a Dia configuration from a JSON file.
170 |
171 | Args:
172 | path: The path to the configuration file.
173 |
174 | Returns:
175 | A validated DiaConfig instance if the file exists and is valid,
176 | otherwise None if the file is not found.
177 |
178 | Raises:
179 | ValueError: If the path does not point to an existing .json file.
180 | pydantic.ValidationError: If the JSON content fails validation against the DiaConfig schema.
181 | """
182 | try:
183 | with open(path, "r") as f:
184 | content = f.read()
185 | return cls.model_validate_json(content)
186 | except FileNotFoundError:
187 | return None
188 |
--------------------------------------------------------------------------------
/dia/layers.py:
--------------------------------------------------------------------------------
1 | import torch
2 | import torch.nn as nn
3 | import torch.nn.functional as F
4 | from torch import Tensor
5 | from torch.nn import RMSNorm
6 |
7 | from .config import DiaConfig
8 | from .state import DecoderInferenceState, EncoderInferenceState, KVCache
9 |
10 |
11 | def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
12 | return tuple(ax if ax >= 0 else ndim + ax for ax in axes)
13 |
14 |
15 | class DenseGeneral(nn.Module):
16 | """
17 | PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
18 |
19 | Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
20 | for the generalized matrix multiplication. Weight/bias shapes are calculated
21 | and parameters created during initialization based on config.
22 | `load_weights` validates shapes and copies data.
23 |
24 | Attributes:
25 | axis (Tuple[int, ...]): Input axis or axes to contract.
26 | in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
27 | out_features (Tuple[int, ...]): Shape of the output features (non-contracted dims).
28 | use_bias (bool): Whether to add a bias term.
29 | weight (nn.Parameter): The kernel parameter.
30 | bias (Optional[nn.Parameter]): The bias parameter (if use_bias=True).
31 | """
32 |
33 | def __init__(
34 | self,
35 | in_shapes: tuple[int, ...],
36 | out_features: tuple[int, ...],
37 | axis: tuple[int, ...] = (-1,),
38 | weight_dtype: torch.dtype | None = None,
39 | device: torch.device | None = None,
40 | ):
41 | super().__init__()
42 | self.in_shapes = in_shapes
43 | self.out_features = out_features
44 | self.axis = axis
45 | self.kernel_shape = self.in_shapes + self.out_features
46 |
47 | factory_kwargs = {"device": device, "dtype": weight_dtype}
48 | self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
49 | self.register_parameter("bias", None)
50 |
51 | def forward(self, inputs: Tensor) -> Tensor:
52 | norm_axis = _normalize_axes(self.axis, inputs.ndim)
53 | kernel_contract_axes = tuple(range(len(norm_axis)))
54 |
55 | output = torch.tensordot(
56 | inputs.to(self.weight.dtype),
57 | self.weight,
58 | dims=(norm_axis, kernel_contract_axes),
59 | ).to(inputs.dtype)
60 | return output
61 |
62 |
63 | class MlpBlock(nn.Module):
64 | """MLP block using DenseGeneral."""
65 |
66 | def __init__(self, embed_dim: int, intermediate_dim: int, compute_dtype: torch.dtype):
67 | super().__init__()
68 | self.dtype = compute_dtype
69 |
70 | self.wi_fused = DenseGeneral(
71 | in_shapes=(embed_dim,),
72 | out_features=(2, intermediate_dim),
73 | axis=(-1,),
74 | weight_dtype=compute_dtype,
75 | )
76 |
77 | self.wo = DenseGeneral(
78 | in_shapes=(intermediate_dim,),
79 | out_features=(embed_dim,),
80 | axis=(-1,),
81 | weight_dtype=compute_dtype,
82 | )
83 |
84 | def forward(self, x: torch.Tensor) -> torch.Tensor:
85 | """Forward pass."""
86 | fused_x = self.wi_fused(x)
87 |
88 | gate = fused_x[..., 0, :]
89 | up = fused_x[..., 1, :]
90 |
91 | hidden = torch.mul(F.silu(gate), up).to(self.dtype)
92 |
93 | output = self.wo(hidden)
94 | return output
95 |
96 |
97 | class RotaryEmbedding(nn.Module):
98 | """Rotary Position Embedding (RoPE) implementation in PyTorch."""
99 |
100 | def __init__(
101 | self,
102 | embedding_dims: int,
103 | min_timescale: int = 1,
104 | max_timescale: int = 10000,
105 | dtype: torch.dtype = torch.float32,
106 | ):
107 | super().__init__()
108 | if embedding_dims % 2 != 0:
109 | raise ValueError("Embedding dim must be even for RoPE.")
110 | self.embedding_dims = embedding_dims
111 | self.min_timescale = min_timescale
112 | self.max_timescale = max_timescale
113 | self.dtype = dtype
114 |
115 | half_embedding_dim = embedding_dims // 2
116 | fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
117 | self.register_buffer(
118 | "timescale",
119 | self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction,
120 | persistent=False,
121 | )
122 |
123 | def extra_repr(self) -> str:
124 | s = f"{self.timescale.shape}"
125 | return s
126 |
127 | def forward(self, inputs: torch.Tensor, position: torch.Tensor):
128 | """Applies RoPE."""
129 | position = position.unsqueeze(-1).unsqueeze(-1)
130 | timescale = self.timescale.to(inputs.device)
131 | sinusoid_inp = position / timescale
132 | sin = torch.sin(sinusoid_inp).to(inputs.dtype)
133 | cos = torch.cos(sinusoid_inp).to(inputs.dtype)
134 | first_half, second_half = torch.chunk(inputs, 2, dim=-1)
135 | first_part = first_half * cos - second_half * sin
136 | second_part = second_half * cos + first_half * sin
137 | return torch.cat((first_part, second_part), dim=-1)
138 |
139 |
140 | class Attention(nn.Module):
141 | """Attention using DenseGeneral."""
142 |
143 | def __init__(
144 | self,
145 | config: DiaConfig,
146 | q_embed_dim: int,
147 | kv_embed_dim: int,
148 | num_query_heads: int,
149 | num_kv_heads: int,
150 | head_dim: int,
151 | compute_dtype: torch.dtype,
152 | is_cross_attn: bool = False,
153 | out_embed_dim: int | None = None,
154 | ):
155 | super().__init__()
156 | self.num_query_heads = num_query_heads
157 | self.num_kv_heads = num_kv_heads
158 | self.head_dim = head_dim
159 | self.is_cross_attn = is_cross_attn
160 | self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
161 | self.projected_query_dim = num_query_heads * head_dim
162 | if num_query_heads % num_kv_heads != 0:
163 | raise ValueError(f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})")
164 | self.num_gqa_groups = num_query_heads // num_kv_heads
165 |
166 | # --- Projection Layers using DenseGeneral ---
167 | self.q_proj = DenseGeneral(
168 | in_shapes=(q_embed_dim,),
169 | out_features=(num_query_heads, head_dim),
170 | axis=(-1,),
171 | weight_dtype=compute_dtype,
172 | )
173 | self.k_proj = DenseGeneral(
174 | in_shapes=(kv_embed_dim,),
175 | out_features=(num_kv_heads, head_dim),
176 | axis=(-1,),
177 | weight_dtype=compute_dtype,
178 | )
179 | self.v_proj = DenseGeneral(
180 | in_shapes=(kv_embed_dim,),
181 | out_features=(num_kv_heads, head_dim),
182 | axis=(-1,),
183 | weight_dtype=compute_dtype,
184 | )
185 | self.o_proj = DenseGeneral(
186 | in_shapes=(num_query_heads, head_dim),
187 | out_features=(self.output_dim,),
188 | axis=(-2, -1),
189 | weight_dtype=compute_dtype,
190 | )
191 |
192 | # --- Rotary Embedding ---
193 | self.rotary_emb = RotaryEmbedding(
194 | embedding_dims=self.head_dim,
195 | min_timescale=config.model.rope_min_timescale,
196 | max_timescale=config.model.rope_max_timescale,
197 | dtype=compute_dtype,
198 | )
199 |
200 | def forward(
201 | self,
202 | Xq: torch.Tensor, # (B, T, D) T = 1 in AR generation
203 | Xkv: torch.Tensor, # (B, S, E) S = 1 in AR generation
204 | q_positions: torch.Tensor, # (B, T)
205 | kv_positions: torch.Tensor | None = None, # (B, S)
206 | attn_mask: torch.Tensor | None = None, # None in Decoder Self Attention, Valid mask in Others
207 | cache: KVCache | None = None, # None in Encoder, KVCache in Decoder
208 | prefill: bool = False,
209 | is_causal: bool = False,
210 | ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
211 | """
212 | Performs attention calculation with optional KV caching.
213 |
214 | Args:
215 | Xq: Query tensor (B, T, D). T=1 during single-step decoding.
216 | Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
217 | q_positions: Positions for queries (B, T).
218 | kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
219 | attn_mask: Attention mask.
220 | cache: KVCache.
221 | prefill: If True, use prefill mode.
222 |
223 | Returns:
224 | A tuple containing:
225 | - output: The attention output tensor (B, T, output_dim).
226 | - 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.
227 | """
228 | if kv_positions is None:
229 | kv_positions = q_positions
230 | original_dtype = Xq.dtype
231 |
232 | Xq_BxTxNxH = self.q_proj(Xq)
233 | Xq_BxTxNxH = self.rotary_emb(Xq_BxTxNxH, position=q_positions)
234 | Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
235 |
236 | attn_k: torch.Tensor | None = None
237 | attn_v: torch.Tensor | None = None
238 |
239 | if self.is_cross_attn:
240 | attn_k, attn_v = cache.k, cache.v
241 | else:
242 | Xk_BxSxKxH = self.k_proj(Xkv) # (B, S, K, H)
243 | Xv_BxSxKxH = self.v_proj(Xkv) # (B, S, K, H)
244 | Xk_BxSxKxH = self.rotary_emb(Xk_BxSxKxH, position=kv_positions) # (B, S, K, H)
245 |
246 | Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2) # (B, K, S, H)
247 | Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2) # (B, K, S, H)
248 |
249 | if cache is None:
250 | attn_k = Xk_BxKxSxH
251 | attn_v = Xv_BxKxSxH
252 | else:
253 | if prefill:
254 | attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
255 | cache.prefill(attn_k, attn_v)
256 | else:
257 | attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH)
258 |
259 | attn_output = F.scaled_dot_product_attention(
260 | Xq_BxNxTxH,
261 | attn_k,
262 | attn_v,
263 | attn_mask=attn_mask,
264 | scale=1.0,
265 | enable_gqa=self.num_gqa_groups > 1,
266 | is_causal=is_causal,
267 | )
268 |
269 | attn_output = attn_output.transpose(1, 2).contiguous() # (B, T, N, H)
270 | output = self.o_proj(attn_output)
271 |
272 | return output.to(original_dtype)
273 |
274 |
275 | class EncoderLayer(nn.Module):
276 | """Transformer Encoder Layer using DenseGeneral."""
277 |
278 | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
279 | super().__init__()
280 | self.config = config
281 | model_config = config.model
282 | enc_config = config.model.encoder
283 | embed_dim = enc_config.n_embd
284 |
285 | self.pre_sa_norm = RMSNorm(
286 | embed_dim,
287 | eps=model_config.normalization_layer_epsilon,
288 | dtype=torch.float32,
289 | )
290 | self.self_attention = Attention(
291 | config,
292 | q_embed_dim=embed_dim,
293 | kv_embed_dim=embed_dim,
294 | num_query_heads=enc_config.n_head,
295 | num_kv_heads=enc_config.n_head,
296 | head_dim=enc_config.head_dim,
297 | compute_dtype=compute_dtype,
298 | is_cross_attn=False,
299 | out_embed_dim=embed_dim,
300 | )
301 | self.post_sa_norm = RMSNorm(
302 | embed_dim,
303 | eps=model_config.normalization_layer_epsilon,
304 | dtype=torch.float32,
305 | )
306 | self.mlp = MlpBlock(embed_dim=embed_dim, intermediate_dim=enc_config.n_hidden, compute_dtype=compute_dtype)
307 |
308 | def forward(
309 | self,
310 | x: torch.Tensor,
311 | state: EncoderInferenceState,
312 | ) -> torch.Tensor:
313 | residual = x
314 | x_norm = self.pre_sa_norm(x)
315 | sa_out = self.self_attention(
316 | Xq=x_norm,
317 | Xkv=x_norm,
318 | q_positions=state.positions,
319 | kv_positions=state.positions,
320 | attn_mask=state.attn_mask,
321 | )
322 | x = residual + sa_out
323 |
324 | residual = x
325 | x_norm = self.post_sa_norm(x)
326 | mlp_out = self.mlp(x_norm)
327 | x = residual + mlp_out
328 |
329 | return x
330 |
331 |
332 | class Encoder(nn.Module):
333 | """Transformer Encoder Stack using DenseGeneral."""
334 |
335 | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
336 | super().__init__()
337 | self.config = config
338 | model_config = config.model
339 | enc_config = config.model.encoder
340 |
341 | self.embedding = nn.Embedding(
342 | model_config.src_vocab_size,
343 | enc_config.n_embd,
344 | dtype=compute_dtype,
345 | )
346 | self.layers = nn.ModuleList([EncoderLayer(config, compute_dtype) for _ in range(enc_config.n_layer)])
347 | self.norm = RMSNorm(
348 | enc_config.n_embd,
349 | eps=model_config.normalization_layer_epsilon,
350 | dtype=torch.float32,
351 | )
352 |
353 | def forward(
354 | self,
355 | x_ids: torch.Tensor,
356 | state: EncoderInferenceState,
357 | ) -> torch.Tensor:
358 | x = self.embedding(x_ids)
359 |
360 | for layer in self.layers:
361 | x = layer(x, state)
362 |
363 | x = self.norm(x)
364 | return x
365 |
366 |
367 | class DecoderLayer(nn.Module):
368 | """Transformer Decoder Layer using DenseGeneral."""
369 |
370 | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
371 | super().__init__()
372 | self.config = config
373 | model_config = config.model
374 | dec_config = config.model.decoder
375 | enc_config = config.model.encoder
376 | dec_embed_dim = dec_config.n_embd
377 | enc_embed_dim = enc_config.n_embd
378 |
379 | # Norms
380 | self.pre_sa_norm = RMSNorm(
381 | dec_embed_dim,
382 | eps=model_config.normalization_layer_epsilon,
383 | dtype=torch.float32,
384 | )
385 | self.pre_ca_norm = RMSNorm(
386 | dec_embed_dim,
387 | eps=model_config.normalization_layer_epsilon,
388 | dtype=torch.float32,
389 | )
390 | self.pre_mlp_norm = RMSNorm(
391 | dec_embed_dim,
392 | eps=model_config.normalization_layer_epsilon,
393 | dtype=torch.float32,
394 | )
395 |
396 | # Self-Attention (GQA) with Causal Masking
397 | self.self_attention = Attention(
398 | config,
399 | q_embed_dim=dec_embed_dim,
400 | kv_embed_dim=dec_embed_dim,
401 | num_query_heads=dec_config.gqa_query_heads,
402 | num_kv_heads=dec_config.kv_heads,
403 | head_dim=dec_config.gqa_head_dim,
404 | compute_dtype=compute_dtype,
405 | is_cross_attn=False,
406 | out_embed_dim=dec_embed_dim,
407 | )
408 | # Cross-Attention (MHA)
409 | self.cross_attention = Attention(
410 | config=config,
411 | q_embed_dim=dec_embed_dim,
412 | kv_embed_dim=enc_embed_dim, # Note kv_embed_dim
413 | num_query_heads=dec_config.cross_query_heads,
414 | num_kv_heads=dec_config.cross_query_heads,
415 | head_dim=dec_config.cross_head_dim,
416 | compute_dtype=compute_dtype,
417 | is_cross_attn=True,
418 | out_embed_dim=dec_embed_dim,
419 | )
420 | # MLP
421 | self.mlp = MlpBlock(
422 | embed_dim=dec_embed_dim,
423 | intermediate_dim=dec_config.n_hidden,
424 | compute_dtype=compute_dtype,
425 | )
426 |
427 | def forward(
428 | self,
429 | x: torch.Tensor,
430 | state: DecoderInferenceState,
431 | self_attn_cache: KVCache | None = None,
432 | cross_attn_cache: KVCache | None = None,
433 | prefill: bool = False,
434 | ) -> torch.Tensor:
435 | residual = x
436 | x_norm = self.pre_sa_norm(x)
437 |
438 | sa_out = self.self_attention(
439 | Xq=x_norm, # (2, 1, D)
440 | Xkv=x_norm, # (2, 1, D)
441 | q_positions=state.dec_positions, # (2, 1)
442 | kv_positions=state.dec_positions, # (2, 1)
443 | attn_mask=None,
444 | cache=self_attn_cache,
445 | prefill=prefill,
446 | is_causal=prefill,
447 | )
448 |
449 | x = residual + sa_out
450 |
451 | residual = x
452 | x_norm = self.pre_ca_norm(x)
453 | ca_out = self.cross_attention(
454 | Xq=x_norm,
455 | Xkv=state.enc_out,
456 | q_positions=state.dec_positions,
457 | kv_positions=state.enc_positions,
458 | attn_mask=state.dec_cross_attn_mask,
459 | cache=cross_attn_cache,
460 | )
461 | x = residual + ca_out
462 |
463 | residual = x
464 | x_norm = self.pre_mlp_norm(x)
465 | mlp_out = self.mlp(x_norm)
466 | x = residual + mlp_out
467 |
468 | return x
469 |
470 |
471 | class Decoder(nn.Module):
472 | """Transformer Decoder Stack using DenseGeneral."""
473 |
474 | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
475 | super().__init__()
476 | self.config = config
477 | model_config = config.model
478 | dec_config = config.model.decoder
479 | data_config = config.data
480 | self.num_channels = data_config.channels
481 | self.num_layers = dec_config.n_layer
482 |
483 | self.embeddings = nn.ModuleList(
484 | [
485 | nn.Embedding(model_config.tgt_vocab_size, dec_config.n_embd, dtype=compute_dtype)
486 | for _ in range(self.num_channels)
487 | ]
488 | )
489 | self.layers = nn.ModuleList(
490 | [DecoderLayer(config=config, compute_dtype=compute_dtype) for _ in range(self.num_layers)]
491 | )
492 |
493 | self.norm = RMSNorm(
494 | dec_config.n_embd,
495 | eps=model_config.normalization_layer_epsilon,
496 | dtype=torch.float32,
497 | )
498 |
499 | self.logits_dense = DenseGeneral(
500 | in_shapes=(dec_config.n_embd,),
501 | out_features=(self.num_channels, model_config.tgt_vocab_size),
502 | axis=(-1,),
503 | weight_dtype=compute_dtype,
504 | )
505 |
506 | def precompute_cross_attn_cache(
507 | self,
508 | enc_out: torch.Tensor, # (B, S, E)
509 | enc_positions: torch.Tensor, # (B, S)
510 | ) -> list[KVCache]:
511 | """
512 | Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
513 | """
514 | per_layer_kv_cache: list[KVCache] = []
515 |
516 | for layer in self.layers:
517 | cross_attn_module = layer.cross_attention
518 | k_proj = cross_attn_module.k_proj(enc_out)
519 | v_proj = cross_attn_module.v_proj(enc_out)
520 |
521 | k_proj = cross_attn_module.rotary_emb(k_proj, position=enc_positions)
522 | k = k_proj.transpose(1, 2)
523 | v = v_proj.transpose(1, 2)
524 |
525 | per_layer_kv_cache.append(KVCache.from_kv(k, v))
526 |
527 | return per_layer_kv_cache
528 |
529 | def decode_step(
530 | self,
531 | tgt_ids_Bx1xC: torch.Tensor, # [B, 1, C]
532 | state: DecoderInferenceState,
533 | ) -> torch.Tensor:
534 | """
535 | Performs a single decoding step, managing KV caches layer by layer.
536 |
537 | Returns:
538 | A tuple containing:
539 | - logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
540 | """
541 |
542 | x = None
543 | for i in range(self.num_channels):
544 | channel_tokens = tgt_ids_Bx1xC[..., i]
545 | channel_embed = self.embeddings[i](channel_tokens)
546 | x = channel_embed if x is None else x + channel_embed
547 |
548 | for i, layer in enumerate(self.layers):
549 | self_cache = state.self_attn_cache[i]
550 | cross_cache = state.cross_attn_cache[i]
551 | x = layer(
552 | x, # (2, 1, D)
553 | state,
554 | self_attn_cache=self_cache,
555 | cross_attn_cache=cross_cache,
556 | )
557 |
558 | x = self.norm(x)
559 | logits_Bx1xCxV = self.logits_dense(x)
560 |
561 | return logits_Bx1xCxV.to(torch.float32)
562 |
563 | def forward(self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInferenceState) -> torch.Tensor:
564 | """
565 | Forward pass for the Decoder stack, managing KV caches.
566 |
567 | Args:
568 | tgt_ids_BxTxC: Target token IDs (B, T, C).
569 | encoder_out: Output from the encoder (B, S, E).
570 | tgt_positions: Positions for target sequence (B, T).
571 | src_positions: Positions for source sequence (B, S).
572 | self_attn_mask: Mask for self-attention.
573 | cross_attn_mask: Mask for cross-attention.
574 | past_key_values: List containing the self-attention KV cache for each layer
575 | from the previous decoding step. `len(past_key_values)` should
576 | equal `num_layers`.
577 | precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
578 | derived from `encoder_out`. This is passed identically
579 | to all layers.
580 |
581 | Returns:
582 | A tuple containing:
583 | - logits: The final output logits (B, T, C * V), cast to float32.
584 | - present_key_values: A list containing the updated self-attention KV cache
585 | for each layer for the *current* decoding step.
586 | """
587 | _, _, num_channels_in = tgt_ids_BxTxC.shape
588 | assert num_channels_in == self.num_channels, "Input channels mismatch"
589 |
590 | # Embeddings
591 | x = None
592 | for i in range(self.num_channels):
593 | channel_tokens = tgt_ids_BxTxC[..., i]
594 | channel_embed = self.embeddings[i](channel_tokens)
595 | x = channel_embed if x is None else x + channel_embed
596 |
597 | for i, layer in enumerate(self.layers):
598 | self_cache = state.self_attn_cache[i]
599 | cross_cache = state.cross_attn_cache[i]
600 | x = layer(x, state, self_attn_cache=self_cache, cross_attn_cache=cross_cache, prefill=True)
601 |
602 | # Final Norm
603 | x = self.norm(x)
604 | logits_BxTxCxV = self.logits_dense(x)
605 |
606 | return logits_BxTxCxV.to(torch.float32)
607 |
608 |
609 | class DiaModel(nn.Module):
610 | """PyTorch Dia Model using DenseGeneral."""
611 |
612 | def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
613 | super().__init__()
614 | self.config = config
615 | self.encoder = Encoder(config, compute_dtype)
616 | self.decoder = Decoder(config, compute_dtype)
617 |
--------------------------------------------------------------------------------
/dia/model.py:
--------------------------------------------------------------------------------
1 | import time
2 | from enum import Enum
3 |
4 | import dac
5 | import numpy as np
6 | import torch
7 | import torchaudio
8 | from huggingface_hub import hf_hub_download
9 |
10 | from .audio import apply_audio_delay, build_delay_indices, build_revert_indices, decode, revert_audio_delay
11 | from .config import DiaConfig
12 | from .layers import DiaModel
13 | from .state import DecoderInferenceState, DecoderOutput, EncoderInferenceState
14 |
15 |
16 | DEFAULT_SAMPLE_RATE = 44100
17 |
18 |
19 | def _get_default_device():
20 | if torch.cuda.is_available():
21 | return torch.device("cuda")
22 | elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
23 | return torch.device("mps")
24 | return torch.device("cpu")
25 |
26 |
27 | def _sample_next_token(
28 | logits_BCxV: torch.Tensor,
29 | temperature: float,
30 | top_p: float,
31 | cfg_filter_top_k: int | None = None,
32 | ) -> torch.Tensor:
33 | if temperature == 0.0:
34 | return torch.argmax(logits_BCxV, dim=-1)
35 |
36 | logits_BCxV = logits_BCxV / temperature
37 | if cfg_filter_top_k is not None:
38 | _, top_k_indices_BCxV = torch.topk(logits_BCxV, k=cfg_filter_top_k, dim=-1)
39 | mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
40 | mask.scatter_(dim=-1, index=top_k_indices_BCxV, value=False)
41 | logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
42 |
43 | if top_p < 1.0:
44 | probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
45 | sorted_probs_BCxV, sorted_indices_BCxV = torch.sort(probs_BCxV, dim=-1, descending=True)
46 | cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
47 |
48 | sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
49 | sorted_indices_to_remove_BCxV[..., 1:] = sorted_indices_to_remove_BCxV[..., :-1].clone()
50 | sorted_indices_to_remove_BCxV[..., 0] = 0
51 |
52 | indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
53 | indices_to_remove_BCxV.scatter_(dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV)
54 | logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
55 |
56 | final_probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
57 |
58 | sampled_indices_BC = torch.multinomial(final_probs_BCxV, num_samples=1)
59 | sampled_indices_C = sampled_indices_BC.squeeze(-1)
60 | return sampled_indices_C
61 |
62 |
63 | class ComputeDtype(str, Enum):
64 | FLOAT32 = "float32"
65 | FLOAT16 = "float16"
66 | BFLOAT16 = "bfloat16"
67 |
68 | def to_dtype(self) -> torch.dtype:
69 | if self == ComputeDtype.FLOAT32:
70 | return torch.float32
71 | elif self == ComputeDtype.FLOAT16:
72 | return torch.float16
73 | elif self == ComputeDtype.BFLOAT16:
74 | return torch.bfloat16
75 | else:
76 | raise ValueError(f"Unsupported compute dtype: {self}")
77 |
78 |
79 | class Dia:
80 | def __init__(
81 | self,
82 | config: DiaConfig,
83 | compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
84 | device: torch.device | None = None,
85 | ):
86 | """Initializes the Dia model.
87 |
88 | Args:
89 | config: The configuration object for the model.
90 | device: The device to load the model onto. If None, will automatically select the best available device.
91 |
92 | Raises:
93 | RuntimeError: If there is an error loading the DAC model.
94 | """
95 | super().__init__()
96 | self.config = config
97 | self.device = device if device is not None else _get_default_device()
98 | if isinstance(compute_dtype, str):
99 | compute_dtype = ComputeDtype(compute_dtype)
100 | self.compute_dtype = compute_dtype.to_dtype()
101 | self.model = DiaModel(config, self.compute_dtype)
102 | self.dac_model = None
103 |
104 | @classmethod
105 | def from_local(
106 | cls,
107 | config_path: str,
108 | checkpoint_path: str,
109 | compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
110 | device: torch.device | None = None,
111 | ) -> "Dia":
112 | """Loads the Dia model from local configuration and checkpoint files.
113 |
114 | Args:
115 | config_path: Path to the configuration JSON file.
116 | checkpoint_path: Path to the model checkpoint (.pth) file.
117 | device: The device to load the model onto. If None, will automatically select the best available device.
118 |
119 | Returns:
120 | An instance of the Dia model loaded with weights and set to eval mode.
121 |
122 | Raises:
123 | FileNotFoundError: If the config or checkpoint file is not found.
124 | RuntimeError: If there is an error loading the checkpoint.
125 | """
126 | config = DiaConfig.load(config_path)
127 | if config is None:
128 | raise FileNotFoundError(f"Config file not found at {config_path}")
129 |
130 | dia = cls(config, compute_dtype, device)
131 |
132 | try:
133 | state_dict = torch.load(checkpoint_path, map_location=dia.device)
134 | dia.model.load_state_dict(state_dict)
135 | except FileNotFoundError:
136 | raise FileNotFoundError(f"Checkpoint file not found at {checkpoint_path}")
137 | except Exception as e:
138 | raise RuntimeError(f"Error loading checkpoint from {checkpoint_path}") from e
139 |
140 | dia.model.to(dia.device)
141 | dia.model.eval()
142 | dia._load_dac_model()
143 | return dia
144 |
145 | @classmethod
146 | def from_pretrained(
147 | cls,
148 | model_name: str = "nari-labs/Dia-1.6B",
149 | compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
150 | device: torch.device | None = None,
151 | ) -> "Dia":
152 | """Loads the Dia model from a Hugging Face Hub repository.
153 |
154 | Downloads the configuration and checkpoint files from the specified
155 | repository ID and then loads the model.
156 |
157 | Args:
158 | model_name: The Hugging Face Hub repository ID (e.g., "NariLabs/Dia-1.6B").
159 | device: The device to load the model onto. If None, will automatically select the best available device.
160 |
161 | Returns:
162 | An instance of the Dia model loaded with weights and set to eval mode.
163 |
164 | Raises:
165 | FileNotFoundError: If config or checkpoint download/loading fails.
166 | RuntimeError: If there is an error loading the checkpoint.
167 | """
168 | config_path = hf_hub_download(repo_id=model_name, filename="config.json")
169 | checkpoint_path = hf_hub_download(repo_id=model_name, filename="dia-v0_1.pth")
170 | return cls.from_local(config_path, checkpoint_path, compute_dtype, device)
171 |
172 | def _load_dac_model(self):
173 | try:
174 | dac_model_path = dac.utils.download()
175 | dac_model = dac.DAC.load(dac_model_path).to(self.device)
176 | except Exception as e:
177 | raise RuntimeError("Failed to load DAC model") from e
178 | self.dac_model = dac_model
179 |
180 | def _prepare_text_input(self, text: str) -> torch.Tensor:
181 | """Encodes text prompt, pads, and creates attention mask and positions."""
182 | text_pad_value = self.config.data.text_pad_value
183 | max_len = self.config.data.text_length
184 |
185 | byte_text = text.encode("utf-8")
186 | replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
187 | text_tokens = list(replaced_bytes)
188 |
189 | current_len = len(text_tokens)
190 | padding_needed = max_len - current_len
191 | if padding_needed <= 0:
192 | text_tokens = text_tokens[:max_len]
193 | padded_text_np = np.array(text_tokens, dtype=np.uint8)
194 | else:
195 | padded_text_np = np.pad(
196 | text_tokens,
197 | (0, padding_needed),
198 | mode="constant",
199 | constant_values=text_pad_value,
200 | ).astype(np.uint8)
201 |
202 | src_tokens = torch.from_numpy(padded_text_np).to(torch.long).to(self.device).unsqueeze(0) # [1, S]
203 | return src_tokens
204 |
205 | def _prepare_audio_prompt(self, audio_prompt: torch.Tensor | None) -> tuple[torch.Tensor, int]:
206 | num_channels = self.config.data.channels
207 | audio_bos_value = self.config.data.audio_bos_value
208 | audio_pad_value = self.config.data.audio_pad_value
209 | delay_pattern = self.config.data.delay_pattern
210 | max_delay_pattern = max(delay_pattern)
211 |
212 | prefill = torch.full(
213 | (1, num_channels),
214 | fill_value=audio_bos_value,
215 | dtype=torch.int,
216 | device=self.device,
217 | )
218 |
219 | prefill_step = 1
220 |
221 | if audio_prompt is not None:
222 | prefill_step += audio_prompt.shape[0]
223 | prefill = torch.cat([prefill, audio_prompt], dim=0)
224 |
225 | delay_pad_tensor = torch.full(
226 | (max_delay_pattern, num_channels), fill_value=-1, dtype=torch.int, device=self.device
227 | )
228 | prefill = torch.cat([prefill, delay_pad_tensor], dim=0)
229 |
230 | delay_precomp = build_delay_indices(
231 | B=1,
232 | T=prefill.shape[0],
233 | C=num_channels,
234 | delay_pattern=delay_pattern,
235 | )
236 |
237 | prefill = apply_audio_delay(
238 | audio_BxTxC=prefill.unsqueeze(0),
239 | pad_value=audio_pad_value,
240 | bos_value=audio_bos_value,
241 | precomp=delay_precomp,
242 | ).squeeze(0)
243 |
244 | return prefill, prefill_step
245 |
246 | def _prepare_generation(self, text: str, audio_prompt: str | torch.Tensor | None, verbose: bool):
247 | enc_input_cond = self._prepare_text_input(text)
248 | enc_input_uncond = torch.zeros_like(enc_input_cond)
249 | enc_input = torch.cat([enc_input_uncond, enc_input_cond], dim=0)
250 |
251 | if isinstance(audio_prompt, str):
252 | audio_prompt = self.load_audio(audio_prompt)
253 | prefill, prefill_step = self._prepare_audio_prompt(audio_prompt)
254 |
255 | if verbose:
256 | print("generate: data loaded")
257 |
258 | enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
259 | encoder_out = self.model.encoder(enc_input, enc_state)
260 |
261 | dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(encoder_out, enc_state.positions)
262 | dec_state = DecoderInferenceState.new(
263 | self.config, enc_state, encoder_out, dec_cross_attn_cache, self.compute_dtype
264 | )
265 | dec_output = DecoderOutput.new(self.config, self.device)
266 | dec_output.prefill(prefill, prefill_step)
267 |
268 | dec_step = prefill_step - 1
269 | if dec_step > 0:
270 | dec_state.prepare_step(0, dec_step)
271 | tokens_BxTxC = dec_output.get_tokens_at(0, dec_step).unsqueeze(0).expand(2, -1, -1)
272 | self.model.decoder.forward(tokens_BxTxC, dec_state)
273 |
274 | return dec_state, dec_output
275 |
276 | def _decoder_step(
277 | self,
278 | tokens_Bx1xC: torch.Tensor,
279 | dec_state: DecoderInferenceState,
280 | cfg_scale: float,
281 | temperature: float,
282 | top_p: float,
283 | cfg_filter_top_k: int,
284 | ) -> torch.Tensor:
285 | audio_eos_value = self.config.data.audio_eos_value
286 | logits_Bx1xCxV = self.model.decoder.decode_step(tokens_Bx1xC, dec_state)
287 |
288 | logits_last_BxCxV = logits_Bx1xCxV[:, -1, :, :]
289 | uncond_logits_CxV = logits_last_BxCxV[0, :, :]
290 | cond_logits_CxV = logits_last_BxCxV[1, :, :]
291 |
292 | logits_CxV = cond_logits_CxV + cfg_scale * (cond_logits_CxV - uncond_logits_CxV)
293 | logits_CxV[:, audio_eos_value + 1 :] = -torch.inf
294 | logits_CxV[1:, audio_eos_value:] = -torch.inf
295 |
296 | pred_C = _sample_next_token(
297 | logits_CxV.float(),
298 | temperature=temperature,
299 | top_p=top_p,
300 | cfg_filter_top_k=cfg_filter_top_k,
301 | )
302 | return pred_C
303 |
304 | def _generate_output(self, generated_codes: torch.Tensor) -> np.ndarray:
305 | num_channels = self.config.data.channels
306 | seq_length = generated_codes.shape[0]
307 | delay_pattern = self.config.data.delay_pattern
308 | audio_pad_value = self.config.data.audio_pad_value
309 | max_delay_pattern = max(delay_pattern)
310 |
311 | revert_precomp = build_revert_indices(
312 | B=1,
313 | T=seq_length,
314 | C=num_channels,
315 | delay_pattern=delay_pattern,
316 | )
317 |
318 | codebook = revert_audio_delay(
319 | audio_BxTxC=generated_codes.unsqueeze(0),
320 | pad_value=audio_pad_value,
321 | precomp=revert_precomp,
322 | T=seq_length,
323 | )[:, :-max_delay_pattern, :]
324 |
325 | min_valid_index = 0
326 | max_valid_index = 1023
327 | invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
328 | codebook[invalid_mask] = 0
329 |
330 | audio = decode(self.dac_model, codebook.transpose(1, 2))
331 |
332 | return audio.squeeze().cpu().numpy()
333 |
334 | def load_audio(self, audio_path: str) -> torch.Tensor:
335 | audio, sr = torchaudio.load(audio_path, channels_first=True) # C, T
336 | if sr != DEFAULT_SAMPLE_RATE:
337 | audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
338 | audio = audio.to(self.device).unsqueeze(0) # 1, C, T
339 | audio_data = self.dac_model.preprocess(audio, DEFAULT_SAMPLE_RATE)
340 | _, encoded_frame, _, _, _ = self.dac_model.encode(audio_data) # 1, C, T
341 | return encoded_frame.squeeze(0).transpose(0, 1)
342 |
343 | def save_audio(self, path: str, audio: np.ndarray):
344 | import soundfile as sf
345 |
346 | sf.write(path, audio, DEFAULT_SAMPLE_RATE)
347 |
348 | @torch.inference_mode()
349 | def generate(
350 | self,
351 | text: str,
352 | max_tokens: int | None = None,
353 | cfg_scale: float = 3.0,
354 | temperature: float = 1.3,
355 | top_p: float = 0.95,
356 | use_torch_compile: bool = False,
357 | cfg_filter_top_k: int = 35,
358 | audio_prompt: str | torch.Tensor | None = None,
359 | audio_prompt_path: str | None = None,
360 | use_cfg_filter: bool | None = None,
361 | verbose: bool = False,
362 | ) -> np.ndarray:
363 | audio_eos_value = self.config.data.audio_eos_value
364 | audio_pad_value = self.config.data.audio_pad_value
365 | delay_pattern = self.config.data.delay_pattern
366 | max_tokens = self.config.data.audio_length if max_tokens is None else max_tokens
367 | max_delay_pattern = max(delay_pattern)
368 | self.model.eval()
369 |
370 | if audio_prompt_path:
371 | print("Warning: audio_prompt_path is deprecated. Use audio_prompt instead.")
372 | audio_prompt = audio_prompt_path
373 | if use_cfg_filter is not None:
374 | print("Warning: use_cfg_filter is deprecated.")
375 |
376 | if verbose:
377 | total_start_time = time.time()
378 |
379 | dec_state, dec_output = self._prepare_generation(text, audio_prompt, verbose)
380 | dec_step = dec_output.prefill_step - 1
381 |
382 | bos_countdown = max_delay_pattern
383 | eos_detected = False
384 | eos_countdown = -1
385 |
386 | if use_torch_compile:
387 | step_fn = torch.compile(self._decoder_step, mode="default")
388 | else:
389 | step_fn = self._decoder_step
390 |
391 | if verbose:
392 | print("generate: starting generation loop")
393 | if use_torch_compile:
394 | print("generate: by using use_torch_compile=True, the first step would take long")
395 | start_time = time.time()
396 |
397 | while dec_step < max_tokens:
398 | dec_state.prepare_step(dec_step)
399 | tokens_Bx1xC = dec_output.get_tokens_at(dec_step).unsqueeze(0).expand(2, -1, -1)
400 | pred_C = step_fn(
401 | tokens_Bx1xC,
402 | dec_state,
403 | cfg_scale,
404 | temperature,
405 | top_p,
406 | cfg_filter_top_k,
407 | )
408 |
409 | if (not eos_detected and pred_C[0] == audio_eos_value) or dec_step == max_tokens - max_delay_pattern - 1:
410 | eos_detected = True
411 | eos_countdown = max_delay_pattern
412 |
413 | if eos_countdown > 0:
414 | step_after_eos = max_delay_pattern - eos_countdown
415 | for i, d in enumerate(delay_pattern):
416 | if step_after_eos == d:
417 | pred_C[i] = audio_eos_value
418 | elif step_after_eos > d:
419 | pred_C[i] = audio_pad_value
420 | eos_countdown -= 1
421 |
422 | bos_countdown = max(0, bos_countdown - 1)
423 | dec_output.update_one(pred_C, dec_step + 1, bos_countdown > 0)
424 |
425 | if eos_countdown == 0:
426 | break
427 |
428 | dec_step += 1
429 | if verbose and dec_step % 86 == 0:
430 | duration = time.time() - start_time
431 | print(
432 | f"generate step {dec_step}: speed={86 / duration:.3f} tokens/s, realtime factor={1 / duration:.3f}x"
433 | )
434 | start_time = time.time()
435 |
436 | if dec_output.prefill_step >= dec_step + 1:
437 | print("Warning: Nothing generated")
438 | return None
439 |
440 | generated_codes = dec_output.generated_tokens[dec_output.prefill_step : dec_step + 1, :]
441 |
442 | if verbose:
443 | total_step = dec_step + 1 - dec_output.prefill_step
444 | total_duration = time.time() - total_start_time
445 | print(f"generate: total step={total_step}, total duration={total_duration:.3f}s")
446 |
447 | return self._generate_output(generated_codes)
448 |
--------------------------------------------------------------------------------
/dia/state.py:
--------------------------------------------------------------------------------
1 | from dataclasses import dataclass
2 |
3 | import torch
4 |
5 | from .config import DiaConfig
6 |
7 |
8 | def create_attn_mask(
9 | q_padding_mask_1d: torch.Tensor,
10 | k_padding_mask_1d: torch.Tensor,
11 | device: torch.device,
12 | is_causal: bool = False,
13 | ) -> torch.Tensor:
14 | """
15 | Creates the attention mask (self or cross) mimicking JAX segment ID logic.
16 | """
17 | B1, Tq = q_padding_mask_1d.shape
18 | B2, Tk = k_padding_mask_1d.shape
19 | assert B1 == B2, "Query and key batch dimensions must match"
20 |
21 | p_mask_q = q_padding_mask_1d.unsqueeze(2) # Shape [B, Tq, 1]
22 | p_mask_k = k_padding_mask_1d.unsqueeze(1) # Shape [B, 1, Tk]
23 |
24 | # Condition A: Non-padding query attends to non-padding key
25 | non_pad_attends_non_pad = p_mask_q & p_mask_k # Shape [B, Tq, Tk]
26 |
27 | # Condition B: Padding query attends to padding key
28 | pad_attends_pad = (~p_mask_q) & (~p_mask_k) # Shape [B, Tq, Tk]
29 |
30 | # Combine: True if padding status is compatible (both non-pad OR both pad)
31 | mask = non_pad_attends_non_pad | pad_attends_pad # Shape [B, Tq, Tk]
32 |
33 | if is_causal:
34 | assert Tq == Tk, "Causal mask requires query and key sequence lengths to be equal"
35 | causal_mask_2d = torch.tril(torch.ones((Tq, Tk), dtype=torch.bool, device=device)) # Shape [Tq, Tk]
36 | causal_mask = mask & causal_mask_2d # Shape [B, Tq, Tk]
37 | return causal_mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
38 | else:
39 | return mask.unsqueeze(1) # Shape [B, 1, Tq, Tk]
40 |
41 |
42 | @dataclass
43 | class EncoderInferenceState:
44 | """Parameters specifically for encoder inference."""
45 |
46 | max_seq_len: int
47 | device: torch.device
48 | positions: torch.Tensor
49 | padding_mask: torch.Tensor
50 | attn_mask: torch.Tensor
51 |
52 | @classmethod
53 | def new(cls, config: DiaConfig, cond_src: torch.Tensor) -> "EncoderInferenceState":
54 | """Creates EtorchrInferenceParams from DiaConfig and a device."""
55 | device = cond_src.device
56 |
57 | positions = torch.arange(config.data.text_length, device=device).to(torch.long).unsqueeze(0).expand(2, -1)
58 | padding_mask = (cond_src != config.data.text_pad_value).to(device).expand(2, -1)
59 | attn_mask = create_attn_mask(padding_mask, padding_mask, device, is_causal=False)
60 |
61 | return cls(
62 | max_seq_len=config.data.text_length,
63 | device=device,
64 | positions=positions,
65 | padding_mask=padding_mask,
66 | attn_mask=attn_mask,
67 | )
68 |
69 |
70 | class KVCache:
71 | def __init__(
72 | self,
73 | num_heads: int,
74 | max_len: int,
75 | head_dim: int,
76 | dtype: torch.dtype,
77 | device: torch.device,
78 | k: torch.Tensor | None = None,
79 | v: torch.Tensor | None = None,
80 | ):
81 | self.k = torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device) if k is None else k
82 | self.v = torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device) if v is None else v
83 | self.current_idx = torch.tensor(0)
84 |
85 | @classmethod
86 | def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
87 | return cls(
88 | num_heads=k.shape[1],
89 | max_len=k.shape[2],
90 | head_dim=k.shape[3],
91 | dtype=k.dtype,
92 | device=k.device,
93 | k=k,
94 | v=v,
95 | )
96 |
97 | def update(self, k: torch.Tensor, v: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
98 | self.k[:, :, self.current_idx : self.current_idx + 1, :] = k
99 | self.v[:, :, self.current_idx : self.current_idx + 1, :] = v
100 | self.current_idx += 1
101 | return self.k[:, :, : self.current_idx, :], self.v[:, :, : self.current_idx, :]
102 |
103 | def prefill(self, k: torch.Tensor, v: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
104 | prefill_len = k.shape[2]
105 | self.k[:, :, :prefill_len, :] = k
106 | self.v[:, :, :prefill_len, :] = v
107 | self.current_idx = prefill_len - 1
108 |
109 |
110 | @dataclass
111 | class DecoderInferenceState:
112 | """Parameters specifically for decoder inference."""
113 |
114 | device: torch.device
115 | dtype: torch.dtype
116 | enc_out: torch.Tensor
117 | enc_positions: torch.Tensor
118 | dec_positions: torch.Tensor
119 | dec_cross_attn_mask: torch.Tensor
120 | self_attn_cache: list[KVCache]
121 | cross_attn_cache: list[KVCache]
122 |
123 | @classmethod
124 | def new(
125 | cls,
126 | config: DiaConfig,
127 | enc_state: EncoderInferenceState,
128 | enc_out: torch.Tensor,
129 | dec_cross_attn_cache: list[KVCache],
130 | compute_dtype: torch.dtype,
131 | ) -> "DecoderInferenceState":
132 | """Creates DecoderInferenceParams from DiaConfig and a device."""
133 | device = enc_out.device
134 | max_audio_len = config.data.audio_length
135 |
136 | dec_positions = torch.full((2, 1), fill_value=0, dtype=torch.long, device=device)
137 | tgt_padding_mask = torch.ones((2, 1), dtype=torch.bool, device=device)
138 | dec_cross_attn_mask = create_attn_mask(tgt_padding_mask, enc_state.padding_mask, device, is_causal=False)
139 |
140 | self_attn_cache = [
141 | KVCache(
142 | config.model.decoder.kv_heads,
143 | max_audio_len,
144 | config.model.decoder.gqa_head_dim,
145 | compute_dtype,
146 | device,
147 | )
148 | for _ in range(config.model.decoder.n_layer)
149 | ]
150 |
151 | return cls(
152 | device=device,
153 | dtype=compute_dtype,
154 | enc_out=enc_out,
155 | enc_positions=enc_state.positions,
156 | dec_positions=dec_positions,
157 | dec_cross_attn_mask=dec_cross_attn_mask,
158 | self_attn_cache=self_attn_cache,
159 | cross_attn_cache=dec_cross_attn_cache,
160 | )
161 |
162 | def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
163 | if step_to is None:
164 | step_to = step_from + 1
165 | self.dec_positions = torch.arange(step_from, step_to, device=self.device).unsqueeze(0).expand(2, -1)
166 |
167 |
168 | @dataclass
169 | class DecoderOutput:
170 | generated_tokens: torch.Tensor
171 | prefill_step: int
172 |
173 | @classmethod
174 | def new(cls, config: DiaConfig, device: torch.device) -> "DecoderOutput":
175 | max_audio_len = config.data.audio_length
176 | return cls(
177 | generated_tokens=torch.full(
178 | (max_audio_len, config.data.channels),
179 | fill_value=-1,
180 | dtype=torch.int,
181 | device=device,
182 | ),
183 | prefill_step=0,
184 | )
185 |
186 | def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
187 | if step_to is None:
188 | step_to = step_from + 1
189 | return self.generated_tokens[step_from:step_to, :]
190 |
191 | def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
192 | if apply_mask:
193 | mask = self.generated_tokens[step : step + 1, :] == -1
194 | self.generated_tokens[step : step + 1, :] = torch.where(
195 | mask, dec_out, self.generated_tokens[step : step + 1, :]
196 | )
197 | else:
198 | self.generated_tokens[step : step + 1, :] = dec_out
199 |
200 | def prefill(self, dec_out: torch.Tensor, prefill_step: int):
201 | length = dec_out.shape[0]
202 | self.generated_tokens[0:length, :] = dec_out
203 | self.prefill_step = prefill_step
204 |
--------------------------------------------------------------------------------
/dia/static/images/banner.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nari-labs/dia/355a2a849d468630e88895196a6fd648849611f6/dia/static/images/banner.png
--------------------------------------------------------------------------------
/docker/Dockerfile.cpu:
--------------------------------------------------------------------------------
1 | # Dockerfile.cpu - CPU-only deployment for DIA
2 | # --------------------------------------------------
3 | # Build: docker build . -f docker/Dockerfile.cpu -t dia-cpu
4 | # Run: docker run --rm -p 7860:7860 dia-cpu
5 |
6 | FROM python:3.10-slim
7 |
8 | # Set non-interactive frontend
9 | ENV DEBIAN_FRONTEND=noninteractive
10 |
11 | # Install venv, and system dependencies
12 | RUN apt-get update && apt-get install -y \
13 | python3-venv \
14 | libsndfile1 \
15 | ffmpeg \
16 | curl \
17 | && apt-get clean && rm -rf /var/lib/apt/lists/*
18 |
19 | # Create non-root user and set up directories
20 | RUN useradd -m -u 1001 appuser && \
21 | mkdir -p /app/outputs /app && \
22 | chown -R appuser:appuser /app
23 |
24 | USER appuser
25 | WORKDIR /app
26 |
27 | # Copy all code (including pyproject.toml)
28 | COPY --chown=appuser:appuser . .
29 |
30 | # Create and activate virtual environment
31 | RUN python3 -m venv /app/venv
32 | ENV PATH="/app/venv/bin:$PATH"
33 |
34 | # Install all project dependencies (CPU-only PyTorch)
35 | RUN pip install --upgrade pip && \
36 | pip install torch torchaudio --index-url https://download.pytorch.org/whl/cpu && \
37 | pip install --no-cache-dir -e .[dev]
38 |
39 | # Set environment variables
40 | ENV PYTHONUNBUFFERED=1 \
41 | PYTHONPATH=/app
42 |
43 | # Expose Gradio default port
44 | ENV GRADIO_SERVER_NAME="0.0.0.0"
45 | EXPOSE 7860
46 |
47 | # Entrypoint
48 | CMD ["python3", "app.py"]
49 |
--------------------------------------------------------------------------------
/docker/Dockerfile.gpu:
--------------------------------------------------------------------------------
1 | # Dockerfile.gpu - GPU deployment for DIA
2 | # --------------------------------------------------
3 | # Build: docker build . -f docker/Dockerfile.gpu -t dia-gpu
4 | # Run: docker run --rm --gpus all -p 7860:7860 dia-gpu
5 | # Requires NVIDIA Container Toolkit on host.
6 |
7 | FROM pytorch/pytorch:2.1.2-cuda12.1-cudnn8-runtime
8 |
9 | # Set non-interactive frontend
10 | ENV DEBIAN_FRONTEND=noninteractive
11 |
12 | # Install venv, and system dependencies
13 | RUN apt-get update && apt-get install -y \
14 | python3-venv \
15 | libsndfile1 \
16 | ffmpeg \
17 | curl \
18 | && apt-get clean && rm -rf /var/lib/apt/lists/*
19 |
20 | # Create non-root user and set up directories
21 | RUN useradd -m -u 1001 appuser && \
22 | mkdir -p /app/outputs /app && \
23 | chown -R appuser:appuser /app
24 |
25 | USER appuser
26 | WORKDIR /app
27 |
28 | # Copy all code (including pyproject.toml)
29 | COPY --chown=appuser:appuser . .
30 |
31 | # Create and activate virtual environment
32 | RUN python3 -m venv /app/venv
33 | ENV PATH="/app/venv/bin:$PATH"
34 |
35 | # Install all project dependencies
36 | RUN pip install --upgrade pip && pip install --no-cache-dir .
37 |
38 | # Set environment variables
39 | ENV PYTHONUNBUFFERED=1 \
40 | PYTHONPATH=/app \
41 | USE_GPU=true \
42 | LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda-12.1/lib64:${LD_LIBRARY_PATH}
43 |
44 | # Expose Gradio default port
45 | ENV GRADIO_SERVER_NAME="0.0.0.0"
46 | EXPOSE 7860
47 |
48 | # Entrypoint
49 | CMD ["python3", "app.py"]
50 |
--------------------------------------------------------------------------------
/example/simple.py:
--------------------------------------------------------------------------------
1 | from dia.model import Dia
2 |
3 |
4 | model = Dia.from_pretrained("nari-labs/Dia-1.6B", compute_dtype="float16")
5 |
6 | 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."
7 |
8 | output = model.generate(text, use_torch_compile=True, verbose=True)
9 |
10 | model.save_audio("simple.mp3", output)
11 |
--------------------------------------------------------------------------------
/example/voice_clone.py:
--------------------------------------------------------------------------------
1 | from dia.model import Dia
2 |
3 |
4 | model = Dia.from_pretrained("nari-labs/Dia-1.6B", compute_dtype="float16")
5 |
6 | # You should put the transcript of the voice you want to clone
7 | # We will use the audio created by running simple.py as an example.
8 | # Note that you will be REQUIRED TO RUN simple.py for the script to work as-is.
9 | 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."
10 | clone_from_audio = "simple.mp3"
11 |
12 | # For your custom needs, replace above with below and add your audio file to this directory:
13 | # clone_from_text = "[S1] ... [S2] ... [S1] ... corresponding to your_audio_name.mp3"
14 | # clone_from_audio = "your_audio_name.mp3"
15 |
16 | # Text to generate
17 | 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."
18 |
19 | # It will only return the audio from the text_to_generate
20 | output = model.generate(
21 | clone_from_text + text_to_generate, audio_prompt=clone_from_audio, use_torch_compile=True, verbose=True
22 | )
23 |
24 | model.save_audio("voice_clone.mp3", output)
25 |
--------------------------------------------------------------------------------
/example_prompt.mp3:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/nari-labs/dia/355a2a849d468630e88895196a6fd648849611f6/example_prompt.mp3
--------------------------------------------------------------------------------
/pyproject.toml:
--------------------------------------------------------------------------------
1 | [project]
2 | name = "nari-tts"
3 | version = "0.1.0"
4 | description = "Dia - A text-to-speech model for dialogue generation"
5 | readme = "README.md"
6 | requires-python = ">=3.10"
7 | license = {file = "LICENSE"}
8 | authors = [
9 | {name = "Nari Labs", email = "contact@narilabs.ai"}
10 | ]
11 | dependencies = [
12 | "descript-audio-codec>=1.0.0",
13 | "gradio>=5.25.2",
14 | "huggingface-hub>=0.30.2",
15 | "numpy>=2.2.4",
16 | "pydantic>=2.11.3",
17 | "soundfile>=0.13.1",
18 | "torch>=2.6.0",
19 | "torchaudio>=2.6.0",
20 | "triton>=3.2.0 ; sys_platform == 'linux'",
21 | "triton-windows>=3.2.0.post18 ; sys_platform == 'win32'",
22 | ]
23 |
24 | [build-system]
25 | requires = ["hatchling"]
26 | build-backend = "hatchling.build"
27 |
28 | [project.urls]
29 | "Homepage" = "https://github.com/nari-labs/dia"
30 | "Bug Tracker" = "https://github.com/nari-labs/dia/issues"
31 |
32 | [tool.hatch.build.targets.wheel]
33 | packages = ["dia"]
34 |
35 | [tool.ruff]
36 | # Never enforce `E501` (line length violations).
37 | lint.ignore = ["C901", "E501", "E741", "W605"]
38 | lint.select = ["C", "E", "F", "I", "W"]
39 | line-length = 119
40 |
41 | # Ignore import violations in all `__init__.py` files.
42 | [tool.ruff.lint.per-file-ignores]
43 | "__init__.py" = ["E402", "F401", "F403", "F811"]
44 |
45 | [tool.ruff.lint.isort]
46 | lines-after-imports = 2
47 |
48 | [tool.uv.sources]
49 | torch = [
50 | { index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
51 | ]
52 | torchaudio = [
53 | { index = "pytorch-cu126", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
54 | ]
55 |
56 | [[tool.uv.index]]
57 | name = "pytorch-cu126"
58 | url = "https://download.pytorch.org/whl/cu126"
59 | explicit = true
60 |
--------------------------------------------------------------------------------
4 |
5 | 
10 | 
13 | 
14 |