Update app.py moondream VLM

app.py

@@ -1,27 +1,22 @@
-import spaces
+import os
 import gradio as gr
 import pandas as pd
 from pathlib import Path
-from transformers import AutoModelForCausalLM, AutoTokenizer
 from PIL import Image
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import spaces  # <-- Crucial ZeroGPU orchestration engine
 
-# =====================================================================
-# 💾 STATE PERSISTENCE ENGINE (Local File System Checkpoints)
-# =====================================================================
 STORAGE_PATH = Path("state_manifest.csv")
 
 def load_persisted_state():
-    """Reads the asset manifest from the local disk partition upon boot."""
     if STORAGE_PATH.exists():
         try:
-            print("💾 Persistent storage record detected. Loading digital twin...")
             df = pd.read_csv(STORAGE_PATH)
             if list(df.columns) == ["Ingredient Asset", "Current Volume / Qty", "Telemetry Status", "Book Value (Est)"]:
                 return df.values.tolist()
         except Exception as e:
             print(f"⚠️ State reconstruction exception: {e}")
-            
-    # Default baseline parameters if no checkpoint exists
     return [
         ["Greek Yogurt", "0.5 Liters", "Fresh", "$2.50"],
         ["Baby Spinach", "1 Bag", "Wilting (Critical)", "$3.00"],
@@ -30,86 +25,93 @@ def load_persisted_state():
     ]
 
 def save_state_to_disk(dataframe_rows):
-    """Commits the current state machine rows permanently to the disk."""
     try:
         df = pd.DataFrame(
             dataframe_rows, 
             columns=["Ingredient Asset", "Current Volume / Qty", "Telemetry Status", "Book Value (Est)"]
         )
         df.to_csv(STORAGE_PATH, index=False)
-        print("💾 Digital twin state successfully synced to disk partition.")
     except Exception as e:
         print(f"⚠️ Critical storage write failure: {e}")
 
 # =====================================================================
-# 🤖 FIXED COMPACT VISION-LANGUAGE INFRASTRUCTURE (CPU Optimized)
+# 🛰️ GLOBAL VISION-LANGUAGE INITIALIZATION
 # =====================================================================
-print("🤖 Initializing Edge-native Vision-Telemetry Model...")
+print("🤖 Initializing Moondream VLM weights into memory pool...")
 model_id = "vikhyatk/moondream2"
 
-# Removing the old revision tag forces transformers to fetch their updated, stable script
-model = AutoModelForCausalLM.from_pretrained(
-    model_id, 
-    trust_remote_code=True
-)
+# Keep the model load global. On ZeroGPU, this registers it in container RAM.
+model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
 tokenizer = AutoTokenizer.from_pretrained(model_id)
+print("Model layers successfully loaded.")
 
-@spaces.GPU
+# =====================================================================
+# ⚡ ZERO-GPU INTERFERENCE PIPELINE
+# =====================================================================
+@spaces.GPU  # <-- This moves the actual image-math onto the A10G GPU cluster dynamically!
 def process_kitchen_operations(input_image, budget, days):
     if input_image is None:
-        # Fallback to your baseline if no image is uploaded
         df_rows = load_persisted_state()
-        blueprint_md = "No telemetry image provided."
-        return df_rows, blueprint_md
+        return df_rows, "⚠️ Please upload a fridge optical scan to map telemetry."
 
-    # Ensure image is in RGB format
+    # 1. Capture spatial metadata and shift data to CUDA tensor space
     image = input_image.convert("RGB")
+    width, height = image.size
     
-    # 1. Ask the Vision Model to identify the items
-    prompt = "List the main fresh ingredients visible in this refrigerator with approximate quantities, separated by commas."
-    detected_text = model.answer_question(image, prompt, tokenizer)
-    
-    # Example raw output: "Salmon fillets (2 pieces), Broccoli (1 head), Mushrooms, Lettuce"
-    print(f"Telemetried Inventory: {detected_text}")
+    # Move the model to GPU *inside* the decorated ZeroGPU environment
+    model.to("cuda")
+
+    # 2. Extract raw visual assets directly from your image text prompt
+    prompt = "List the specific fresh food items visible inside this refrigerator, separated by commas. Be accurate."
     
-    # 2. Parse the text dynamically into DataFrame Rows
-    # (You can split the string and map defaults for status and cost)
+    # Moondream's native GPU answer method
+    detected_text = model.answer_question(image, prompt, tokenizer)
+    print(f"🎯 Moondream Vision Output: {detected_text}")
+
+    # 3. Parse VLM output into the digital twin dataframe matrix
     df_rows = []
-    items = detected_text.split(",")
-    for item in items:
-        clean_item = item.strip()
-        if clean_item:
-            # Dynamically push parsed edge data into the twin matrix
-            df_rows.append([clean_item, "1 Unit", "Nominal", "$3.50"])
+    # Splitting the string output into dynamic lines or comma elements
+    raw_items = [item.strip() for item in detected_text.replace(".", ",").split(",") if item.strip()]
+    
+    if not raw_items:
+        raw_items = ["Detected Asset Cluster"]
+
+    # Loop through what Moondream ACTUALLY sees (Salmon, Broccoli, etc.)
+    for raw_item in raw_items:
+        # Dynamically append discovered assets to the dashboard UI table
+        df_rows.append([raw_item.title(), "1 Unit", "Fresh (Nominal)", "$3.50"])
             
-    # Save the real detected state to disk
+    # Sync this newly discovered dataset permanently to disk
     save_state_to_disk(df_rows)
     
-    # 3. Pass this data to your recipe generation string...
+    # 4. Generate Strategy Blueprint Text
     daily_budget_allowance = budget / days
-    blueprint_md = f"Plan optimized for identified assets: {detected_text}."
+    blueprint_md = (
+        f"### 🛠️ Logistics Routing Strategy Generated via Moondream VLM\n\n"
+        f"**Operational Horizon:** {days} Days | **Daily Financial Runway Ceiling:** ${daily_budget_allowance:.2f}/day\n\n"
+        f"📋 **Dynamic Cook Execution Pathway:**\n"
+        f"- **Day 1 Asset Recovery:** Prioritize routing the fresh profile (**{raw_items[0]}**) discovered by the optical scan sensor.\n"
+        f"- **Day 2-3 Runway Balance:** Allocate remaining items under your strict financial boundary of ${daily_budget_allowance:.2f}/day."
+    )
     
     return df_rows, blueprint_md
 
 # =====================================================================
-# 🎨 GRADIO DASHBOARD INTERFACE
+# 🎨 GRADIO INTERFACE DESIGN
 # =====================================================================
 with gr.Blocks() as demo:
     gr.Markdown("# 🛰️ Parallel Plate: Kitchen Operations & Visual Logistics Engine")
-    gr.Markdown("### *Physical-to-Digital Twin Asset Management Pipeline (Hardware Tier: 💻 CPU Free Tier Optimized)*")
-    
+    gr.Markdown("### *Physical-to-Digital Twin Asset Management Pipeline (Hardware Tier: ⚡ ZeroGPU Accelerated)*")
+        
     with gr.Row():
         with gr.Column(scale=1):
             gr.Markdown("### 📸 1. Physical Edge Ingestion")
             image_input = gr.Image(label="Upload Fridge Optical Survey Scan", type="pil")
-            
-            with gr.Row():
+                        with gr.Row():
                 budget_slider = gr.Slider(minimum=5, maximum=100, value=25, step=5, label="Runway Budget ($ USD)")
                 days_slider = gr.Slider(minimum=1, maximum=7, value=3, step=1, label="Target Horizon (Days)")
-                
-            scan_btn = gr.Button("🚀 Initialize Asset Optimization Scan", variant="primary")
-            
-        with gr.Column(scale=1):
+                            scan_btn = gr.Button("🚀 Initialize Asset Optimization Scan", variant="primary")
+                    with gr.Column(scale=1):
             gr.Markdown("### 📊 2. Active Refrigerator Digital Twin")
             inventory_df = gr.Dataframe(
                 headers=["Ingredient Asset", "Current Volume / Qty", "Telemetry Status", "Book Value (Est)"],
@@ -118,10 +120,9 @@ with gr.Blocks() as demo:
                 interactive=True,
                 wrap=True,
                 column_widths=["30%", "25%", "25%", "20%"],
-                value=load_persisted_state()  # Auto-injects last known state at container boot
+                value=load_persisted_state()
             )
-            
-            gr.Markdown("### 🔀 3. Resource Routing & Recipe Output")
+                        gr.Markdown("### 🔀 3. Resource Routing & Recipe Output")
             output_text = gr.Markdown("*Awaiting asset telemetry mapping to generate optimal cooking execution paths...*")
 
     scan_btn.click(