Update app.py

app.py

@@ -17,7 +17,7 @@ def log_status(msg):
 # --- CONFIGURATION ---
 SOURCE_REPO = "metanthropic/metanthropic-phi3-encrypted"
 SOURCE_FILE = "metanthropic-phi3-v1.mguf"
-TEMP_DECRYPTED = "/tmp/model_sovereign.gguf"
+TEMP_DECRYPTED = "/tmp/model_sovereign_v2.gguf"
 HF_TOKEN = os.environ.get("HF_TOKEN")
 SECRET_KEY_HEX = os.environ.get("DECRYPTION_KEY")
 
@@ -25,20 +25,17 @@ SECRET_KEY_HEX = os.environ.get("DECRYPTION_KEY")
 def initialize_weights():
     try:
         if os.path.exists(TEMP_DECRYPTED):
-            log_status("⚡ [CACHE] Resuming from existing sovereign weights.")
+            log_status("⚡ [CACHE] Resuming from sovereign weights.")
             return True
 
         if not HF_TOKEN or not SECRET_KEY_HEX:
-            log_status("❌ [SECURITY] Credentials missing. Verify HF_TOKEN and DECRYPTION_KEY.")
+            log_status("❌ [SECURITY] Credentials missing.")
             return False
 
-        log_status("🔐 [AUTH] Establishing secure link to Hugging Face...")
         login(token=HF_TOKEN)
-
-        log_status(f"⬇️ [NETWORK] Fetching {SOURCE_FILE}...")
         path = hf_hub_download(repo_id=SOURCE_REPO, filename=SOURCE_FILE, local_dir=".")
         
-        log_status("🔓 [DECRYPT] Unlocking GGUF weights...")
+        log_status("🔓 [DECRYPT] Unlocking weights...")
         key = bytes.fromhex(SECRET_KEY_HEX)
         aes = AESGCM(key)
         
@@ -50,60 +47,50 @@ def initialize_weights():
                 f_out.write(chunk)
         
         os.remove(path)
-        log_status("✅ [SYSTEM] Weight integrity verified.")
+        log_status("✅ [SYSTEM] Integrity verified.")
         return True
     except Exception as e:
-        log_status(f"❌ [CRITICAL] Boot failure: {str(e)}")
-        log_status(traceback.format_exc())
+        log_status(f"❌ [BOOT ERROR] {str(e)}")
         return False
 
-# --- ENGINE INITIALIZATION (PERFORMANCE TUNED) ---
+# --- ENGINE ---
 llm = None
 if initialize_weights():
     try:
-        log_status("🧠 [ENGINE] Initializing Neural Infrastructure...")
+        log_status("🧠 [ENGINE] Initializing Llama...")
         llm = Llama(
             model_path=TEMP_DECRYPTED,
             n_ctx=2048,
-            n_threads=2,     # Locked to HF Free Tier vCPU limit
-            n_batch=512,     # Optimized for prompt ingestion speed
-            use_mlock=True,  # Pin model to RAM
+            n_threads=2,
+            n_batch=512,
+            use_mlock=True,
             verbose=False
         )
-        log_status("🚀 [SYSTEM] Sovereign Node Online.")
+        log_status("🚀 [SYSTEM] Node Online.")
     except Exception as e:
-        log_status(f"❌ [ENGINE ERROR] Neural load failed: {e}")
+        log_status(f"❌ [ENGINE ERROR] {e}")
 
-# --- SECURED API CORE (CONVEX BRIDGE) ---
+# --- API CORE ---
 app = FastAPI()
 
 @app.post("/run_inference")
 async def run_inference(request: Request):
-    if not llm:
-        return {"error": "System Offline", "logs": DIAGNOSTIC_LOG[-5:]}
-    
+    if not llm: return {"error": "Offline"}
     data = await request.json()
-    
-    # 🔐 SECURITY HANDSHAKE
-    # We use the SECRET_KEY_HEX as a shared secret for API authorization.
     if data.get("secretKey") != SECRET_KEY_HEX:
-        return {"error": "Unauthorized API Access. Use the Gradio UI instead."}
+        return {"error": "Unauthorized Access"}
     
     prompt = data.get("prompt", "")
-    
-    output = llm(
-        f"<|user|>\n{prompt}<|end|>\n<|assistant|>", 
-        max_tokens=512, 
-        stop=["<|end|>", "<|endoftext|>"]
-    )
+    output = llm(f"<|user|>\n{prompt}<|end|>\n<|assistant|>", max_tokens=512, stop=["<|end|>"])
     return {"response": output['choices'][0]['text'].strip()}
 
-# --- PREMIUM UI LOGIC (STREAMING) ---
+# --- UI LOGIC (FIXED DELTA ERROR) ---
 def ui_chat(msg, hist):
     if not llm:
-        yield f"🚨 **SYSTEM ARCHITECTURE FAILURE**\n\nLatest Diagnostics:\n```\n" + "\n".join(DIAGNOSTIC_LOG[-3:]) + "\n```"
+        yield "🚨 System Offline."
         return
 
+    # High-level completion stream
     stream = llm(
         f"<|user|>\n{msg}<|end|>\n<|assistant|>",
         max_tokens=512,
@@ -113,48 +100,20 @@ def ui_chat(msg, hist):
     
     partial_text = ""
     for chunk in stream:
-        delta = chunk['choices'][0]['delta']
-        if 'content' in delta:
-            partial_text += delta['content']
+        # FIX: Access 'text' directly from choices[0]
+        token = chunk['choices'][0]['text']
+        if token:
+            partial_text += token
             yield partial_text
 
-# --- METANTHROPIC BRANDED INTERFACE ---
-custom_css = """
-footer {visibility: hidden}
-.gradio-container {background-color: #050505 !important}
-#title-container {text-align: center; margin-bottom: 30px}
-#title-container h1 {color: #ffffff; font-family: 'Inter', sans-serif; font-weight: 800; letter-spacing: -1.5px}
-.message.user {background-color: #1a1a1a !important; border: 1px solid #333 !important; border-radius: 12px !important}
-.message.assistant {background-color: #0f0f0f !important; border: 1px solid #222 !important; border-radius: 12px !important}
-"""
+# --- BRANDED UI ---
+custom_css = "footer {visibility: hidden} .gradio-container {background-color: #050505 !important}"
 
 demo = gr.ChatInterface(
     ui_chat,
     title="METANTHROPIC · PHI-3 SOVEREIGN",
-    description="""
-    <div id="title-container">
-        <p style="color: #a3a3a3; font-size: 1.1em; max-width: 600px; margin: 0 auto;">
-            Accessing <b>Node-01</b> of the Metanthropic Neural Infrastructure. 
-            Secure inference via localized sovereign weights.
-        </p>
-        <div style="margin-top: 15px; display: flex; justify-content: center; gap: 20px;">
-            <span style="color: #22c55e; font-size: 0.85em; font-family: monospace;">● ENGINE: READY</span>
-            <span style="color: #3b82f6; font-size: 0.85em; font-family: monospace;">● ENCRYPTION: AES-GCM</span>
-            <span style="color: #a855f7; font-size: 0.85em; font-family: monospace;">● TYPE: STREAMING</span>
-        </div>
-    </div>
-    """,
-    theme=gr.themes.Soft(
-        primary_hue="slate",
-        neutral_hue="zinc",
-        font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui"],
-    ).set(
-        body_background_fill="#050505",
-        block_background_fill="#0a0a0a",
-        block_border_width="1px",
-        button_primary_background_fill="#ffffff",
-        button_primary_text_color="#000000",
-    ),
+    description="Secure inference via decentralized sovereign weights.",
+    theme=gr.themes.Soft(primary_hue="slate", neutral_hue="zinc"),
     css=custom_css
 )