app.py

import spaces

import gradio as gr
import io
from urllib.request import urlopen

import soundfile as sf
import torch
from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig




MODEL_ID = "microsoft/Phi-4-multimodal-instruct"

processor = AutoProcessor.from_pretrained(MODEL_ID, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="cuda" if torch.cuda.is_available() else "cpu",
    torch_dtype="auto",
    trust_remote_code=True
)

model.load_adapter(MODEL_ID, adapter_name="speech", device_map="cuda" if torch.cuda.is_available() else "cpu", adapter_kwargs={"subfolder": 'speech-lora'})
model.set_adapter("speech")

generation_config = GenerationConfig.from_pretrained(MODEL_ID)
generation_config.num_logits_to_keep = 1

@spaces.GPU
def run_phi4(audio_path: str, instruction: str) -> str:
    if not audio_path:
        return "Please upload an audio file."

    audio, samplerate = sf.read(audio_path)

    user_prompt = "<|user|>"
    assistant_prompt = "<|assistant|>"
    prompt_suffix = "<|end|>"

    prompt = f"{user_prompt}<|audio_1|>{instruction}{prompt_suffix}{assistant_prompt}"

    inputs = processor(text=prompt, audios=[(audio, samplerate)], return_tensors="pt").to(model.device)

    output_ids = model.generate(
        **inputs,
        max_new_tokens=4096,
        generation_config=generation_config,
    )
    output_ids = output_ids[:, inputs["input_ids"].shape[1]:]
    response = processor.batch_decode(output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    return response


with gr.Blocks(title="Phi-4 Multimodal Audio Demo") as demo:
    gr.Markdown("# Phi-4 Multimodal (Audio) Demo")
    gr.Markdown("Upload an audio file and run instructions with Phi-4.")

    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(type="filepath", label="Upload Audio")
            instruction = gr.Textbox(
                label="Instruction",
                value=(
                    "Transcribe the audio to text, and then translate the audio to French. "
                    "Use <sep> as a separator between the original transcript and the translation."
                ),
            )
            submit_btn = gr.Button("Run", variant="primary")
        with gr.Column():
            output_text = gr.Textbox(label="Model Response", lines=14)

    submit_btn.click(run_phi4, [audio_input, instruction], output_text)


if __name__ == "__main__":
    demo.queue().launch(share=False, ssr_mode=False)