amalia-llm/amalia-fala-loader

AMALIA-FALA Loader

Lightweight loading and inference package for the AMALIA-FALA ASR models.

This repository contains only the Python code required to load and run AMALIA-FALA models.
The model weights are stored separately in the corresponding Hugging Face model repositories and are downloaded automatically on first use.

Supported Models

• amalia-llm/AMALIA-SFT-FALA
• amalia-llm/AMALIA-DPO-FALA

Installation

Install the loader directly from Hugging Face:

pip install git+https://huggingface.co/amalia-llm/amalia-fala-loader

Usage

Run inference on an audio file

from amalia_fala import pipeline

asr = pipeline(
    "amalia-llm/AMALIA-DPO-FALA",
    device="cuda:0",
)

result = asr("/path/to/audio.wav")
print(result["text"])

The checkpoint is downloaded automatically from the selected model repository and cached locally by Hugging Face. Future runs reuse the cached checkpoint.

Use the SFT model

To use the AMALIA-SFT-FALA model, change the repository name:

from amalia_fala import pipeline

asr = pipeline(
    "amalia-llm/AMALIA-SFT-FALA",
    device="cuda:0",
)

result = asr("/path/to/audio.wav")
print(result["text"])

Direct model loading

If you need direct access to the model, tokenizer, and configuration objects:

from amalia_fala import load_model

model, tokenizer, model_config, dataset_config = load_model(
    "amalia-llm/AMALIA-DPO-FALA",
    device="cuda:0",
)

CPU usage

GPU inference is strongly recommended. CPU loading may work, but it will be very slow because the checkpoint contains the speech encoder, projector, and AMALIA language model weights.

from amalia_fala import pipeline

asr = pipeline(
    "amalia-llm/AMALIA-DPO-FALA",
    device="cpu",
)

result = asr("/path/to/audio.wav")
print(result["text"])

Repository structure

amalia-fala-loader/
├── README.md
├── pyproject.toml
└── amalia_fala/
    ├── __init__.py
    └── load_amalia_fala.py

Notes

• This repository does not contain model weights.
• The model checkpoints are hosted in the corresponding AMALIA-FALA model repositories.
• The checkpoint is downloaded automatically on first use.
• The Hugging Face cache is reused after the first download.
• A GPU is recommended for practical inference speed.
• The AMALIA-FALA checkpoints contain the speech encoder, projector, and AMALIA LLM weights.