Deploying with Flex AI

Steps

1. Add Source

Add your model source to Flex AI:

flexai source add <model_source_name> https://github.com/Paul-HenriBJT/train-parler-tts.git

Replace <model_source_name> with your actual source name.

2. Choose Dataset

The dataset is loaded directly from the Hugging Face Hub during the training phase. Choose a small dataset from FCS to reduce build time. The Hugging Face dataset name is specified in the config JSON file.

3. Launch Training

Incorrect Way (Will Fail)

flexai training run <name_of_the_training> \
    --source-name <model_source_name> \
    --source-revision master \
    --dataset <your_dataset_name> \
    -- ./training/run_parler_tts_training.py ./training/config/preprocess.json

This method is expected to fail without an error message (it will get stuck).

Correct Way

flexai training run <name_of_the_training> \
    --source-name <model_source_name> \
    --source-revision <revision> \
    --dataset <your_dataset_name> \
    -- -m accelerate.commands.launch ./training/run_parler_tts_training.py ./training/config/preprocess.json

This method should succeed.

Fetch the output

You can fetch the output with

flexai training fetch <name_of_the_training>

Notes

• Ensure all placeholders (e.g., <name_of_the_training>, <model_source_name>, <your_dataset_name>) are replaced with your actual values.
• The key difference in the correct method is the use of accelerate.commands.launch, which properly initializes the training process.
• This does not requires a Hugging Face account.