Update app.py

app.py

@@ -5,6 +5,7 @@ import time
 # --- CONFIGURATION ---
 # Your AMD MI300X Droplet IP
 DROPLET_IP = "134.199.195.151" 
+# Make sure this is your currently active Pinggy URL!
 API_URL = "https://jqzih-134-199-192-140.run.pinggy-free.link/translate"
 
 # --- LOGIC ---
@@ -13,12 +14,27 @@ def ghost_translate(cuda_code):
     yield "👻 Ghost-Coder: Analyzing CUDA Kernel...", "Loading..."
     
     try:
+        # --- THE MAGIC FIX: ChatML Instruction Template ---
+        # This forces the Qwen model to act as an assistant instead of an autocomplete
+        formatted_prompt = f"""<|im_start|>system
+You are Ghost-Coder, an expert AI specializing in translating CUDA code to AMD HIP code. Output ONLY the valid, translated HIP code. Do not echo the prompt or include markdown formatting.<|im_end|>
+<|im_start|>user
+Translate this exact CUDA code to HIP:
+
+{cuda_code}<|im_end|>
+<|im_start|>assistant
+"""
+        
         # Step 2: Real API Call to your MI300X
-        response = requests.post(API_URL, json={"code": cuda_code}, timeout=120)
+        # We send the strictly formatted prompt instead of just the raw code
+        response = requests.post(API_URL, json={"code": formatted_prompt}, timeout=120)
         
         if response.status_code == 200:
             hip_code = response.json().get("hip_code", "// Error: No code returned")
             
+            # Failsafe: Clean up the output in case the model echoes the prompt
+            hip_code = hip_code.replace(formatted_prompt, "").strip()
+            
             # Step 3: Agentic Visual Steps (Self-Healing Simulation)
             yield "🔄 Analyzing HIP logic on ROCm stack...", "Generating..."
             time.sleep(1)