Update app.py

app.py

@@ -1,54 +1,54 @@
 import os
+import io
 import gradio as gr
-import torch
 from PIL import Image
-from transformers import AutoProcessor, AutoModelForImageTextToText
+from huggingface_hub import InferenceClient
 
 # 1. Safely extract your secure secret token
 hf_token = os.environ.get("HF_TOKEN")
 
-# 2. Define target repository path
+# 2. Initialize the optimized serverless client targeting LEGATO
 MODEL_ID = "guangyangmusic/legato-small"
+client = InferenceClient(model=MODEL_ID, token=hf_token)
 
-print("Loading model architecture and weights...")
-
-# FIX: Use the updated AutoModelForImageTextToText class for modern transformers compatibility
-processor = AutoProcessor.from_pretrained(MODEL_ID, token=hf_token, trust_remote_code=True)
-model = AutoModelForImageTextToText.from_pretrained(
-    MODEL_ID, 
-    token=hf_token,
-    trust_remote_code=True,
-    torch_dtype=torch.bfloat16,   # Lowers operational RAM allocation to under 16GB
-    low_cpu_mem_usage=True         # Prevents transient configuration-phase crashes
-)
-
-model.eval()
-print("🎉 LEGATO architecture successfully instantiated!")
+print("Hugging Face API Pipeline active. Forwarding inference tasks...")
 
 def run_legato_omr(image):
     if image is None:
         return "Please upload a sheet music image first!"
     
     try:
-        # Pass visual assets into the tokenizer pipeline
-        inputs = processor(images=image, return_tensors="pt")
+        # Convert PIL Image to raw bytes for standard network payload streaming
+        buffered = io.BytesIO()
+        image.save(buffered, format="JPEG")
+        image_bytes = buffered.getvalue()
         
-        # CPU-isolated matrix multiplication
-        with torch.no_grad():
-            generated_ids = model.generate(**inputs, max_new_tokens=512)
-            generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        # Stream the image asset straight to the dedicated model architecture host
+        print("Sending payload to Hugging Face Inference clusters...")
+        response = client.image_to_text(image_bytes)
         
-        return generated_text if isinstance(generated_text, list) else generated_text
+        # Handle different inference response formats
+        if hasattr(response, 'generated_text'):
+            return response.generated_text
+        elif isinstance(response, dict) and "generated_text" in response:
+            return response["generated_text"]
+        else:
+            return str(response)
+            
     except Exception as e:
-        return f"Inference pipeline execution error: {str(e)}"
+        return (
+            f"API Inference Call Error: {str(e)}\n\n"
+            "💡 Tip: Ensure your HF_TOKEN has 'Read' access, and that "
+            "the model page doesn't require separate terms acceptance."
+        )
 
 # 3. Present UI layout blocks
 demo = gr.Interface(
     fn=run_legato_omr,
     inputs=gr.Image(type="pil", label="1. Upload Sheet Music Snippet"),
     outputs=gr.Textbox(label="2. LEGATO Output (ABC Notation Text)", show_copy_button=True),
-    title="🎼 LEGATO End-to-End OMR Engine",
-    description="Running entirely inside cloud-native CPU constraints. Music recognition cycles take 10-15 seconds."
+    title="🎼 LEGATO End-to-End OMR Engine (API Optimized)",
+    description="Proxied through Hugging Face's serverless pipeline to bypass local 16GB CPU limits. Transcriptions finish in seconds."
 )
 
 if __name__ == "__main__":
@@ -56,4 +56,3 @@ if __name__ == "__main__":
 
 
 
-