init

app.py

@@ -1,4 +1,76 @@
 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,
+    _attn_implementation="flash_attention_2",
+)
+
+generation_config = GenerationConfig.from_pretrained(MODEL_ID)
+
+
+def run_phi4(audio_path: str, instruction: str, max_tokens: int = 512) -> 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=int(max_tokens),
+        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."
+                ),
+            )
+            max_tokens = gr.Slider(128, 2000, value=1000, step=64, label="Max Output Tokens")
+            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, max_tokens], output_text)
+
+
+if __name__ == "__main__":
+    demo.queue().launch(share=False, ssr_mode=False)
+
+import gradio as gr
 
 def greet(name):
     return "Hello " + name + "!!"

requirements.txt

@@ -0,0 +1,7 @@
+gradio>=4.0.0
+torch>=2.1.0
+transformers>=4.43.0
+accelerate>=0.30.0
+soundfile>=0.12.1
+sentencepiece>=0.1.99
+