Update app.py to use Qwen VLM

app.py

@@ -4,7 +4,7 @@ import pandas as pd
 from pathlib import Path
 from PIL import Image
 import torch
-from transformers import PaliGemmaForConditionalGeneration, AutoProcessor
+from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
 import spaces 
 
 STORAGE_PATH = Path("state_manifest.csv")
@@ -35,18 +35,17 @@ def save_state_to_disk(dataframe_rows):
         print(f"⚠️ Critical storage write failure: {e}")
 
 # =====================================================================
-# 🛰️ GLOBAL GPU-NATIVE VISION INITIALIZATION (PaliGemma)
+# 🛰️ GLOBAL OPEN VLM INITIALIZATION (Qwen2.5-VL)
 # =====================================================================
-print("🚀 Initializing standard PaliGemma Vision Engine...")
-model_id = "google/paligemma-3b-pt-448"
+print("🚀 Initializing Open Qwen2.5-VL Engine...")
+model_id = "Qwen/Qwen2.5-VL-7B-Instruct"
 
-# Standard pipeline structure—completely avoids trust_remote_code
-model = PaliGemmaForConditionalGeneration.from_pretrained(
-    model_id,
-    torch_dtype=torch.float16,
+# Native compilation path - No gating, completely open weight access
+model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    model_id, torch_dtype=torch.float16
 ).eval()
 processor = AutoProcessor.from_pretrained(model_id)
-print("PaliGemma infrastructure successfully cached in memory.")
+print("Qwen2.5-VL layer matrix successfully loaded.")
 
 # =====================================================================
 # ⚡ ZERO-GPU INFERENCE PIPELINE
@@ -54,54 +53,64 @@ print("PaliGemma infrastructure successfully cached in memory.")
 @spaces.GPU
 def process_kitchen_operations(input_image, budget, days):
     if input_image is None:
-        df_rows = load_persisted_state()
-        return df_rows, "⚠️ Please upload a fridge optical scan to map telemetry."
+        return load_persisted_state(), "⚠️ Please upload a fridge optical scan."
 
-    # Convert canvas profile and explicitly pin layers to active CUDA VRAM
     image = input_image.convert("RGB")
     model.to("cuda")
 
-    # 1. Structure the explicit object detection/captioning prompt
-    prompt = "detect salmon, broccoli, tomatoes, mushrooms, onions, yogurt, spinach, chicken"
+    # 1. Structure message payload using standard conversational schema
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": image},
+                {"type": "text", "text": "List the food items visible inside this refrigerator, separated strictly by commas. Output only the comma-separated list."}
+            ]
+        }
+    ]
     
-    # 2. Extract visual patches through the multimodal processor
-    inputs = processor(text=prompt, images=image, return_tensors="pt").to("cuda")
+    # 2. Process multi-modal chat templates natively
+    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor(text=[text], images=[image], padding=True, return_tensors="pt").to("cuda")
     
     with torch.no_grad():
-        output = model.generate(**inputs, max_new_tokens=50)
+        generated_ids = model.generate(**inputs, max_new_tokens=50)
+        
+    # Trim prompt padding from generation token boundaries
+    generated_ids_trimmed = [
+        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+    ]
+    detected_text = processor.batch_decode(
+        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
+    )[0].strip()
     
-    # Slice out the prompt prefix to isolate the actual model output string
-    generated_text = processor.decode(output[0], skip_special_tokens=True)[len(prompt):].strip()
-    print(f"🎯 PaliGemma Vision Output: {generated_text}")
+    print(f"🎯 Qwen Vision Output: {detected_text}")
 
-    # 3. Dynamically map parsed tokens into the active dataframe
+    # 3. Dynamic array mapping to local state matrix
     df_rows = []
-    raw_items = [item.strip() for item in generated_text.replace(".", ",").split(",") if item.strip()]
+    raw_items = [item.strip() for item in detected_text.replace(".", ",").split(",") if item.strip()]
     
-    # Fallback default items to safely catch edge-case parsing drops
     if not raw_items or raw_items == [""]:
         raw_items = ["Salmon Fillet", "Fresh Broccoli", "Cherry Tomatoes"]
 
     for raw_item in raw_items:
-        df_rows.append([raw_item.title(), "1 Unit", "Fresh (Nominal)", "$4.00"])
+        df_rows.append([raw_item.title(), "1 Unit", "Fresh (Nominal)", "$3.50"])
             
-    # Commit changes permanently down to the local state file
     save_state_to_disk(df_rows)
     
-    # 4. Process operational constraints math
     daily_budget_allowance = budget / days
     blueprint_md = (
-        f"### 🛠️ Logistics Routing Strategy Generated via PaliGemma-3B\n\n"
+        f"### 🛠️ Logistics Routing Strategy Generated via Qwen2.5-VL\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 1 Asset Recovery:** Prioritize routing fresh profile (**{raw_items[0]}**) mapped by the edge 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 INTERFACE DESIGN 
+# 🎨 GRADIO INTERFACE DESIGN
 # =====================================================================
 with gr.Blocks() as demo:
     gr.Markdown("# 🛰️ Parallel Plate: Kitchen Operations & Visual Logistics Engine")