### Usage example for DPO training

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Shows how to load a JSON dataset, create a DPOText2AudioDataset, and use it with a DataLoader. The example unpacks the batch into text, chosen paths, rejected paths, durations, and indices.

```python
from datasets import load_dataset
from torch.utils.data import DataLoader
from tangoflux.utils import DPOText2AudioDataset

raw = load_dataset("json", data_files={"train": "data/train_dpo.json"})
train_dataset = DPOText2AudioDataset(
    raw["train"],
    prefix="",
    text_column="captions",
    audio_w_column="chosen",
    audio_l_column="reject",
    duration="duration",
    num_examples=-1,
)
loader = DataLoader(train_dataset, shuffle=True, batch_size=4, collate_fn=train_dataset.collate_fn)
text, chosen_paths, rejected_paths, durations, indices = next(iter(loader))
```

--------------------------------

### Example: decode and save audio

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-nodes.md

Example usage of the play method to decode and save audio to the output directory. It prints the UI audios result.

```python
node = TangoFluxVAEDecodeAndPlay()
result = node.play(vae, latents, filename_prefix="my_audio", format="wav", save_output=True)
print(result["ui"]["audios"])
```

--------------------------------

### Installation: Install requirements

Source: https://github.com/declare-lab/tangoflux/blob/main/comfyui/README.md

Changes into the comfyui subdirectory and runs the install.py script to install dependencies.

```bash
cd ComfyUI-TangoFlux/comfyui
python install.py
```

--------------------------------

### Installation: Navigate to custom_nodes directory

Source: https://github.com/declare-lab/tangoflux/blob/main/comfyui/README.md

First step of the installation process. Change to the ComfyUI custom_nodes directory before cloning the repository.

```bash
cd ComfyUI/custom_nodes
```

--------------------------------

### HTTP request example for GET /tangoflux/playaudio

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/endpoints.md

Shows the URL query parameters for requesting an audio file. The filename must be a valid output file name; type and subfolder are optional.

```http
GET /tangoflux/playaudio?filename=TangoFlux_00001_.wav&type=output&subfolder=
```

--------------------------------

### Install TangoFlux from GitHub

Source: https://github.com/declare-lab/tangoflux/blob/main/Demo.ipynb

Installs the TangoFlux package directly from the GitHub repository.

```python
!pip install git+https://github.com/declare-lab/TangoFlux.git
```

--------------------------------

### TangoFluxSampler usage example

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-nodes.md

Shows how to call the sample method with a prompt, steps, guidance_scale, duration, and seed, and prints the duration from the returned latents dict.

```python
latents, = sampler.sample(model, "Rain on a tin roof",
                          steps=50, guidance_scale=4.5, duration=10, seed=7)
print(latents["duration"])  # 10
```

--------------------------------

### DPOText2AudioDataset.__len__

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Returns the number of examples in the dataset.

```APIDOC
## DPOText2AudioDataset.__len__

### Description
Returns the number of examples in the dataset.

### Method
`__len__`

### Endpoint
`DPOText2AudioDataset.__len__`

### Parameters
None

### Request Example
```python
len(dataset)
```

### Response
#### Success Response
- **int** - `len(self.inputs)`.

#### Response Example
```python
1000
```
```

--------------------------------

### Usage example for SFT training with Text2AudioDataset

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Shows how to load a JSON dataset, create a Text2AudioDataset with column names 'captions', 'location', and 'duration', and build a DataLoader with the dataset's collate_fn. The example is from tangoflux/train.py.

```python
from datasets import load_dataset
from torch.utils.data import DataLoader
from tangoflux.utils import Text2AudioDataset

raw = load_dataset("json", data_files={"train": "data/train.json"})
train_dataset = Text2AudioDataset(
    raw["train"], "", "captions", "location", "duration", num_examples=-1
)
train_dataloader = DataLoader(
    train_dataset,
    shuffle=True,
    batch_size=4,
    collate_fn=train_dataset.collate_fn,
)
text, audio_paths, durations, _ = next(iter(train_dataloader))
```

--------------------------------

### Direct usage example of TangoFluxLoader

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-nodes.md

Shows how to instantiate TangoFluxLoader and call load_tangoflux with TeaCache enabled and a custom threshold, returning the model and VAE.

```python
from comfyui.nodes import TangoFluxLoader

loader = TangoFluxLoader()
model, vae = loader.load_tangoflux(
    enable_teacache=True, rel_l1_thresh=0.3, device="cuda"
)
```

--------------------------------

### Launch TangoFlux DPO training with accelerate

Source: https://github.com/declare-lab/tangoflux/blob/main/README.md

Use this command to start DPO training. The training files must contain 'chosen', 'reject', 'caption', and 'duration' fields. The command uses the same accelerate and config files as standard training.

```bash
CUDA_VISIBLE_DEVICES=0,1 accelerate launch --config_file='configs/accelerator_config.yaml' tangoflux/train_dpo.py   --checkpointing_steps="best" --save_every=5 --config='configs/tangoflux_config.yaml'
```

--------------------------------

### DPOText2AudioDataset __len__ method

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Returns the number of examples in the dataset, which is the length of the internal inputs list.

```python
    def __len__(self):
```

--------------------------------

### Launch TangoFlux training with accelerate

Source: https://github.com/declare-lab/tangoflux/blob/main/README.md

Use this command to start standard training of TangoFlux. It requires an accelerate configuration file and a TangoFlux config file specifying training data paths. The command sets CUDA_VISIBLE_DEVICES to 0,1 and uses the provided config files.

```bash
CUDA_VISIBLE_DEVICES=0,1 accelerate launch --config_file='configs/accelerator_config.yaml' tangoflux/train.py   --checkpointing_steps="best" --save_every=5 --config='configs/tangoflux_config.yaml'
```

--------------------------------

### Installation: Clone the repository

Source: https://github.com/declare-lab/tangoflux/blob/main/comfyui/README.md

Clones the TangoFlux repository into the ComfyUI-TangoFlux directory. Note the space in the command between the repository URL and the target directory name.

```bash
git clone https://github.com/declare-lab/TangoFlux  ComfyUI-TangoFlux
```

--------------------------------

### Read WAV file and convert to fake stereo

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Example usage from the training scripts: reads a WAV file with a 30-second duration and repeats mono waveforms to create fake stereo.

```python
wav = read_wav_file("data/test.wav", duration_sec=30)
if wav.shape[0] == 1:
    wav = wav.repeat(2, 1)   # "fake stereo"
```

--------------------------------

### SFT dataset row example (JSON)

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/types.md

Each element in data/train.json must contain captions, location, and duration fields. The captions field is the text prompt, location is the audio file path, and duration is the target duration in seconds. The column names can be configured via --text_column and --audio_column.

```json
{"captions": "Rhythmic wooden table tapping...", "location": "data/test.wav", "duration": 10.0}
```

--------------------------------

### Python client fetch for GET /tangoflux/playaudio

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/endpoints.md

Equivalent client-side fetch using requests. Expects a 200 status and audio/wav content type, then saves the response body to a local file.

```python
# Equivalent client-side fetch (aiohttp / requests)
import requests

r = requests.get(
    "http://127.0.0.1:8188/tangoflux/playaudio",
    params={"filename": "TangoFlux_00001_.wav", "type": "output"},
)
assert r.status_code == 200
assert r.headers["Content-Type"] == "audio/wav"
open("TangoFlux_00001_.wav", "wb").write(r.content)
```

--------------------------------

### Run tangoflux-demo basic

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Basic usage of the `tangoflux-demo` command to launch the Gradio web interface with default settings.

```bash
tangoflux-demo
```

--------------------------------

### Run tangoflux-demo with sharing

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Usage of `tangoflux-demo` with custom host, port, and sharing enabled to create a public share link.

```bash
tangoflux-demo --host 0.0.0.0 --port 7860 --share
```

--------------------------------

### DurationEmbedder usage example

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-model.md

Shows how to instantiate DurationEmbedder and embed a tensor of duration values. The example asserts the output shape is (2, 1, 1024).

```python
import torch
from tangoflux.model import DurationEmbedder

embedder = DurationEmbedder(number_embedding_dim=1024, min_value=0, max_value=30)
emb = embedder(torch.tensor([10.0, 27.5]))    # shape (2, 1, 1024)
assert emb.shape == (2, 1, 1024)
```

--------------------------------

### Define tangoflux-demo CLI command

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Defines the CLI entry point for `tangoflux-demo` with options `--host` (default '127.0.0.1'), `--port` (default None), and `--share` flag. Launches `demo.queue().launch(server_name=host, server_port=port, share=share)`.

```python
@click.command()
@click.option('--host', default='127.0.0.1', help='Host to bind to')
@click.option('--port', default=None, help='Port to bind to')
@click.option('--share', is_flag=True, help='Enable sharing via Gradio')
def main(host, port, share):
```

--------------------------------

### DPOText2AudioDataset.__init__

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Initializes the dataset with the given parameters, preparing internal lists for captions, chosen and rejected audio paths, durations, and indices.

```APIDOC
## DPOText2AudioDataset.__init__

### Description
Initializes the dataset with the given parameters, preparing internal lists for captions, chosen and rejected audio paths, durations, and indices.

### Method
Constructor

### Endpoint
`DPOText2AudioDataset.__init__`

### Parameters
#### Constructor Parameters
- **dataset** (datasets.Dataset or IterableDataset) - Required - HF dataset (from `tangoflux/train_dpo.py`: `load_dataset("json", ...)["train"]`).
- **prefix** (str) - Required - String prepended to every caption (`prefix + inp`).
- **text_column** (str) - Required - Column name holding captions.
- **audio_w_column** (str) - Required - Column name holding "chosen" (`w` = winner) audio paths.
- **audio_l_column** (str) - Required - Column name holding "reject" (`l` = loser) audio paths.
- **duration** (str) - Required - Column name holding durations.
- **num_examples** (int) - Optional - Default `-1` - Truncates all internal lists to this length when `!= -1`; number of usable examples otherwise.

### Request Example
```python
DPOText2AudioDataset(
    dataset=raw["train"],
    prefix="",
    text_column="captions",
    audio_w_column="chosen",
    audio_l_column="reject",
    duration="duration",
    num_examples=-1,
)
```

### Response
#### Success Response
No return value.

#### Response Example
No response body.
```

--------------------------------

### Basic inference_flow example

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-model.md

Basic usage of inference_flow with a single prompt. Returns latents of shape (1, audio_seq_len, 64).

```python
import torch
from tangoflux import TangoFluxInference

model = TangoFluxInference()
latents = model.model.inference_flow(
    "Hammer slowly hitting the wooden table",
    num_inference_steps=50,
    guidance_scale=4.5,
    duration=10,
)
assert latents.shape == (1, model.model.audio_seq_len, 64)
```

--------------------------------

### Register console scripts in setup.py

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Registers the two console scripts `tangoflux` and `tangoflux-demo` in `setup.py`, mapping them to the `main` functions in `tangoflux/cli.py` and `tangoflux/demo.py` respectively.

```python
entry_points={
    "console_scripts": [
        "tangoflux=tangoflux.cli:main",
        "tangoflux-demo=tangoflux.demo:main",
    ],
},
```

--------------------------------

### GET /tangoflux/playaudio

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/endpoints.md

Serves an audio file produced by the TangoFluxVAEDecodeAndPlay node. The filename, type, and subfolder query parameters determine which file is returned. No authentication is required.

```APIDOC
## GET /tangoflux/playaudio

### Description
Serves an audio file produced by the TangoFluxVAEDecodeAndPlay node. The filename, type, and subfolder query parameters determine which file is returned. No authentication is required.

### Method
GET

### Endpoint
/tangoflux/playaudio

### Parameters
#### Query Parameters
- **filename** (str) - Required - Output file name; resolved via `folder_paths.annotated_filepath(filename)`. Leading-`/` values and any value containing `..` are rejected (403). After resolution, `os.path.basename` is applied again before joining.
- **type** (str) - Optional - Directory type when the annotated path has no output dir: `folder_paths.get_directory_by_type(file_type)` — typically `"output"` or `"temp"` (the `TangoFluxVAEDecodeAndPlay` node writes `type="output"` or `type="temp"` into its `ui.audios` entries). Default: `"output"`.
- **subfolder** (str) - Optional - Optional subfolder under the output dir; combined via `os.path.join(output_dir, subfolder)`; the result is checked with `os.path.commonpath` to stay inside `output_dir` (otherwise 403). Default: `None`.

### Request Example
GET /tangoflux/playaudio?filename=my_audio_00001_.wav&type=output&subfolder=

### Response
#### Success Response (200)
- **Content-Type** - Derived from file extension (see table below).
- **Body** - Raw bytes of the audio file.

#### Response Example
HTTP/1.1 200 OK
Content-Type: audio/wav

<binary audio data>

#### Content-Type Map
| Extension | Content-Type |
|-----------|--------------|
| .wav | `audio/wav` |
| .mp3 | `audio/mpeg` |
| .flac | `audio/flac` |
| .aac | `audio/aac` |
| .wma | `audio/x-ms-wma` |

#### Error Responses
| Status | Condition |
|--------|-----------|
| 404 | `filename` query param missing or resolved `file_path` is not a file. |
| 403 | `filename[0] == "/"` or `".." in filename`; subfolder escape — `os.path.commonpath((os.path.abspath(full_output_dir), output_dir)) != output_dir`. |
| 400 | no output dir resolved — `output_dir is None` after `annotated_filepath`/`get_directory_by_type`; unsupported extension — `content_type is None`. |
| 500 | `open(file_path, "rb").read()` raises anything. |

All error responses are empty-bodied `web.Response(status=N)`.

### Usage Example
GET /tangoflux/playaudio?filename=TangoFlux_00001_.wav&type=output

```python
import requests

r = requests.get(
    "http://127.0.0.1:8188/tangoflux/playaudio",
    params={"filename": "TangoFlux_00001_.wav", "type": "output"},
)
assert r.status_code == 200
assert r.headers["Content-Type"] == "audio/wav"
open("TangoFlux_00001_.wav", "wb").write(r.content)
```
```

--------------------------------

### Load TangoFlux model and generate audio

Source: https://github.com/declare-lab/tangoflux/blob/main/Inference.ipynb

Imports torchaudio, TangoFluxInference, and IPython Audio. Creates a TangoFluxInference instance with the 'declare-lab/TangoFlux' model, generates audio from the prompt 'Hammer slowly hitting the wooden table' with 50 steps and 10 seconds duration, and displays the audio in the notebook.

```python
import torchaudio
from tangoflux import TangoFluxInference
from IPython.display import Audio

model = TangoFluxInference(name="declare-lab/TangoFlux")


audio = model.generate("Hammer slowly hitting the wooden table", steps=50, duration=10)

Audio(data=audio, rate=44100)
```

--------------------------------

### Route definition for GET /tangoflux/playaudio

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/endpoints.md

Registers the route on ComfyUI's PromptServer instance. No authentication is required. The handler parses query parameters from the request URL.

```python
@server.PromptServer.instance.routes.get("/tangoflux/playaudio")
async def play_audio(request):
```

--------------------------------

### Launch tangoflux/train.py with accelerate

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-training.md

Command to launch the training script tangoflux/train.py using accelerate with a specified accelerator config and training config. Uses --checkpointing_steps="best" and --save_every=5.

```bash
accelerate launch --config_file='configs/accelerator_config.yaml' tangoflux/train.py \
  --checkpointing_steps="best" --save_every=5 --config='configs/tangoflux_config.yaml'
```

--------------------------------

### DPO usage in tangoflux/train_dpo.py

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-model.md

Usage of the forward method for DPO objective (sft=False). Requires explicit reference-model setup: deep copy of transformer, freezing, and setting requires_grad=False. Returns loss, raw_model_loss, raw_ref_loss, implicit_acc.

```python
import copy
model.ref_transformer = copy.deepcopy(model.transformer)
model.ref_transformer.requires_grad_ = False
model.ref_transformer.eval()
for param in model.ref_transformer.parameters():
    param.requires_grad = False

loss, raw_model_loss, raw_ref_loss, implicit_acc = model(
    audio_latent, text, duration=duration, sft=False
)
```

--------------------------------

### Export nodes and server from comfyui/__init__.py

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-nodes.md

Re-exports the node class mappings and server module, and sets the web directory for the ComfyUI extension.

```python
from .nodes import NODE_CLASS_MAPPINGS
from .server import *
WEB_DIRECTORY = "./comfyui/web"
__all__ = ["NODE_CLASS_MAPPINGS", "WEB_DIRECTORY"]
```

--------------------------------

### tangoflux_config.yaml sample configuration

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/configuration.md

Sample YAML configuration for training TangoFlux. Contains paths, training, and model sections. Loaded by train.py and train_dpo.py via --config (default tangoflux_config.yaml).

```yaml
paths:
  train_file: "data/train.json"
  val_file: "data/val.json"
  test_file: "data/val.json"
  resume_from_checkpoint: ""
  output_dir: "outputs/"

training:
  per_device_batch_size: 4
  learning_rate: 5e-4
  gradient_accumulation_steps: 1
  num_train_epochs: 80
  num_warmup_steps: 1000
  max_audio_duration: 30

model:
  num_layers: 6
  num_single_layers: 18
  in_channels: 64
  attention_head_dim: 128
  joint_attention_dim: 1024
  num_attention_heads: 8
  audio_seq_len: 645
  max_duration: 30
  uncondition: false
  text_encoder_name: "google/flan-t5-large"
```

--------------------------------

### Manual enablement of TeaCache (comfyui.teacache)

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-teacache.md

Shows how to enable TeaCache without the loader: import teacache_forward, assign it to FluxTransformer2DModel.forward, and set the required class attributes (enable_teacache, cnt, num_steps, rel_l1_thresh, accumulated_rel_l1_distance, previous_modulated_input, previous_residual). After setup, call the model normally; the function runs the cached forward when enable_teacache is true.

```python
from diffusers import FluxTransformer2DModel
from comfyui.teacache import teacache_forward

model = FluxTransformer2DModel(...)

FluxTransformer2DModel.forward = teacache_forward
model.__class__.enable_teacache = True
model.__class__.cnt = 0
model.__class__.num_steps = 50
model.__class__.rel_l1_thresh = 0.25
model.__class__.accumulated_rel_l1_distance = 0
model.__class__.previous_modulated_input = None
model.__class__.previous_residual = None

out = model(hidden_states=..., encoder_hidden_states=...,
            timestep=..., img_ids=..., txt_ids=..., return_dict=False)[0]
```

--------------------------------

### Initialize TangoFlux Inference Model

Source: https://github.com/declare-lab/tangoflux/blob/main/Demo.ipynb

Imports required libraries and initializes the TangoFlux inference model using the pretrained 'declare-lab/TangoFlux' checkpoint.

```python
import IPython
import torchaudio
from tangoflux import TangoFluxInference
from IPython.display import Audio

model = TangoFluxInference(name='declare-lab/TangoFlux')
```

--------------------------------

### Run tangoflux basic generation

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Basic usage of the `tangoflux` command to generate audio from a text prompt, saving to `output.wav` with specified duration and steps.

```bash
tangoflux "Hammer slowly hitting the wooden table" output.wav --duration 10 --steps 50
```

--------------------------------

### Import datasets and audio utilities from tangoflux.utils

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Imports the dataset classes and audio helper functions used by the training scripts.

```python
from tangoflux.utils import Text2AudioDataset, DPOText2AudioDataset
from tangoflux.utils import read_wav_file, pad_wav, normalize_wav
```

--------------------------------

### DPOText2AudioDataset class definition

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-utils.md

Constructor signature for the DPOText2AudioDataset class. It takes a Hugging Face dataset, a prefix for captions, column names for text, chosen and rejected audio paths, and a duration column. The num_examples parameter truncates internal lists when not -1.

```python
class DPOText2AudioDataset(Dataset):
    def __init__(
        self,
        dataset,
        prefix,
        text_column,
        audio_w_column,
        audio_l_column,
        duration,
        num_examples=-1,
    ):
```

--------------------------------

### main function signature for generate_crpo_dataset.py

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-crpo.md

Main CLI entry point for the candidate generation script. It parses arguments, validates GPU count, samples prompts, splits across GPUs, and runs generate_audio_chunk processes.

```python
def main():

```

--------------------------------

### Define tangoflux CLI command with click

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-cli.md

Defines the CLI entry point for `tangoflux` with positional arguments `prompt` and `output_file`, and options `--duration` (default 10) and `--steps` (default 50). Validation raises `click.BadParameter` (exit code 2) when duration is outside [1,30] or steps outside [10,100].

```python
@click.command()
@click.argument('prompt')
@click.argument('output_file')
@click.option('--duration', default=10, type=int, help='Duration in seconds (1-30)')
@click.option('--steps', default=50, type=int, help='Number of inference steps (10-100)')
def main(prompt: str, output_file: str, duration: int, steps: int):
```

--------------------------------

### TangoFluxLoader __init__ method

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/comfyui-nodes.md

Initializes the loader with default values for model, vae, enable_teacache, rel_l1_thresh, and stores the original forward method of FluxTransformer2DModel for potential restoration.

```python
    def __init__(self):
        self.model = None
        self.vae = None
        self.enable_teacache = False
        self.rel_l1_thresh = 0.25
        self.original_forward = FluxTransformer2DModel.forward
```

--------------------------------

### TangoFluxInference constructor

Source: https://github.com/declare-lab/tangoflux/blob/main/_autodocs/api-reference/tangoflux-inference.md

Initializes the inference wrapper. Downloads the model snapshot from the Hugging Face Hub, loads the VAE and model weights, and moves them to the specified device. The default device is CUDA if available, otherwise CPU.

```python
    def __init__(
        self,
        name="declare-lab/TangoFlux",
        device="cuda" if torch.cuda.is_available() else "cpu",
    ):
```