example.py

# Example usage script for AudioLLM
import torch
import torchaudio
from transformers import pipeline

# Load the pipeline directly (recommended)
audio_llm = pipeline(
    "text-generation",
    model="cdreetz/audio-llama-hf",
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Process audio file
result = audio_llm("path/to/audio.wav")
print(result[0]["generated_text"])

# Process audio with custom prompt
result = audio_llm(("path/to/audio.wav", "Describe the music in this audio:"))
print(result[0]["generated_text"])

# Text-only generation
result = audio_llm("Write a poem about sound:")
print(result[0]["generated_text"])

# Advanced usage: load model components manually
from transformers import AutoTokenizer, AutoModelForCausalLM, WhisperProcessor

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("cdreetz/audio-llama-hf")
model = AutoModelForCausalLM.from_pretrained("cdreetz/audio-llama-hf")

# Process inputs manually
waveform, sample_rate = torchaudio.load("path/to/audio.wav")
if waveform.shape[0] > 1:  # Convert stereo to mono
    waveform = torch.mean(waveform, dim=0, keepdim=True)

# Preprocess audio
whisper_processor = WhisperProcessor.from_pretrained(model.config.whisper_model_id)
audio_features = whisper_processor(
    waveform.squeeze().numpy(),
    sampling_rate=16000,
    return_tensors="pt"
).input_features

# Tokenize text prompt
inputs = tokenizer("Describe the audio:", return_tensors="pt")

# Generate
with torch.no_grad():
    outputs = model.generate(
        input_ids=inputs.input_ids.to(model.device),
        attention_mask=inputs.attention_mask.to(model.device),
        audio_features=audio_features.to(model.device),
        max_new_tokens=256
    )

# Decode
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)