Update app.py

app.py

@@ -3,52 +3,51 @@ import requests
 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 ---
 def ghost_translate(cuda_code):
-    # Step 1: Initial UI Feedback
     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:
+        # We use a pure-text marker that the backend cannot delete or mangle
+        split_marker = "// --- GHOST-CODER HIP OUTPUT ---"
+        
+        # A standard completion prompt that forces the model to finish the file
+        formatted_prompt = f"""Task: Translate the following CUDA code to AMD HIP. 
+Output ONLY valid C++ code. Do not include markdown blocks or explanations.
+
+// --- ORIGINAL CUDA CODE ---
+{cuda_code}
 
-{cuda_code}<|im_end|>
-<|im_start|>assistant
+{split_marker}
 """
         
-        # Step 2: Real API Call to your MI300X
-        # 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")
+            raw_response = response.json().get("hip_code", "")
             
-            # Failsafe: Clean up the output in case the model echoes the prompt
-            hip_code = hip_code.replace(formatted_prompt, "").strip()
+            # THE HARD SLICE: Cut away the entire input prompt
+            if split_marker in raw_response:
+                hip_code = raw_response.split(split_marker)[-1].strip()
+            else:
+                # Failsafe if the model acts weird
+                hip_code = raw_response.strip()
             
-            # Step 3: Agentic Visual Steps (Self-Healing Simulation)
+            # Step 3: Agentic Visual Steps
             yield "🔄 Analyzing HIP logic on ROCm stack...", "Generating..."
             time.sleep(1)
             yield "🛠️ Verifying syntax and memory offsets...", "Verifying..."
             time.sleep(1)
             yield "✅ Self-Healing successful! HIP Code generated.", hip_code
         else:
-            yield f"❌ Droplet Error: {response.status_code}", "// Check bridge logs on MI300X terminal"
+            yield f"❌ Droplet Error: {response.status_code}", "// Check logs"
             
     except Exception as e:
-        yield f"❌ Connection Error: Ensure bridge is running on {DROPLET_IP}", str(e)
+        yield f"❌ Connection Error: Ensure bridge is running", str(e)
 
-# --- UI DESIGN ---
-# Using gr.Blocks() for a professional, agentic look
 with gr.Blocks() as demo:
     gr.Markdown("# 👻 Ghost-Coder: Autonomous CUDA-to-HIP Agent")
     gr.Markdown("### Powered by AMD Instinct™ MI300X | Qwen2.5-Coder-32B")
@@ -67,8 +66,6 @@ with gr.Blocks() as demo:
             output_code = gr.Code(label="Generated HIP Code", language="cpp", lines=15)
             logs = gr.Textbox(label="Agent Status & Self-Healing Logs", interactive=False)
 
-    # Linking the button to our translation function
     run_btn.click(ghost_translate, inputs=[input_code], outputs=[logs, output_code])
 
-# Launching the app with the 'Soft' theme moved to the launch method
 demo.queue().launch(theme=gr.themes.Soft())