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_v2.gguf"
+TEMP_DECRYPTED = "/tmp/model_stable_v3.gguf"
 HF_TOKEN = os.environ.get("HF_TOKEN")
 SECRET_KEY_HEX = os.environ.get("DECRYPTION_KEY")
 
@@ -25,7 +25,6 @@ SECRET_KEY_HEX = os.environ.get("DECRYPTION_KEY")
 def initialize_weights():
     try:
         if os.path.exists(TEMP_DECRYPTED):
-            log_status("⚡ [CACHE] Resuming from sovereign weights.")
             return True
 
         if not HF_TOKEN or not SECRET_KEY_HEX:
@@ -35,7 +34,7 @@ def initialize_weights():
         login(token=HF_TOKEN)
         path = hf_hub_download(repo_id=SOURCE_REPO, filename=SOURCE_FILE, local_dir=".")
         
-        log_status("🔓 [DECRYPT] Unlocking weights...")
+        log_status("🔓 [DECRYPT] Unlocking GGUF weights...")
         key = bytes.fromhex(SECRET_KEY_HEX)
         aes = AESGCM(key)
         
@@ -47,13 +46,13 @@ def initialize_weights():
                 f_out.write(chunk)
         
         os.remove(path)
-        log_status("✅ [SYSTEM] Integrity verified.")
+        log_status("✅ [SYSTEM] Weight integrity verified.")
         return True
     except Exception as e:
         log_status(f"❌ [BOOT ERROR] {str(e)}")
         return False
 
-# --- ENGINE ---
+# --- ENGINE INITIALIZATION ---
 llm = None
 if initialize_weights():
     try:
@@ -61,21 +60,22 @@ if initialize_weights():
         llm = Llama(
             model_path=TEMP_DECRYPTED,
             n_ctx=2048,
-            n_threads=2,
+            n_threads=2,     # Locked to 2-vCPU Free Tier limit
             n_batch=512,
-            use_mlock=True,
+            use_mlock=True,  # Pin model to RAM
             verbose=False
         )
         log_status("🚀 [SYSTEM] Node Online.")
     except Exception as e:
-        log_status(f"❌ [ENGINE ERROR] {e}")
+        log_status(f"❌ [ENGINE ERROR] Neural load failed: {e}")
 
-# --- API CORE ---
+# --- API CORE (CONVEX BRIDGE) ---
 app = FastAPI()
 
 @app.post("/run_inference")
 async def run_inference(request: Request):
-    if not llm: return {"error": "Offline"}
+    if not llm: return {"error": "System Offline"}
+    
     data = await request.json()
     if data.get("secretKey") != SECRET_KEY_HEX:
         return {"error": "Unauthorized Access"}
@@ -84,14 +84,14 @@ async def run_inference(request: Request):
     output = llm(f"<|user|>\n{prompt}<|end|>\n<|assistant|>", max_tokens=512, stop=["<|end|>"])
     return {"response": output['choices'][0]['text'].strip()}
 
-# --- UI LOGIC (FIXED DELTA ERROR) ---
+# --- PREMIUM UI LOGIC (STREAMING FIX) ---
 def ui_chat(msg, hist):
     if not llm:
-        yield "🚨 System Offline."
+        yield "🚨 SYSTEM OFFLINE. CHECK LOGS."
         return
 
-    # High-level completion stream
-    stream = llm(
+    # Use the High-Level __call__ API with stream=True
+    stream_iterator = llm(
         f"<|user|>\n{msg}<|end|>\n<|assistant|>",
         max_tokens=512,
         stop=["<|end|>", "<|endoftext|>"],
@@ -99,20 +99,23 @@ def ui_chat(msg, hist):
     )
     
     partial_text = ""
-    for chunk in stream:
-        # FIX: Access 'text' directly from choices[0]
-        token = chunk['choices'][0]['text']
-        if token:
-            partial_text += token
-            yield partial_text
+    try:
+        for chunk in stream_iterator:
+            # FIX: CompletionChunks store text in ['choices'][0]['text']
+            # There is NO 'delta' or 'content' in this API mode.
+            token = chunk['choices'][0].get('text', "")
+            if token:
+                partial_text += token
+                yield partial_text
+    except Exception as e:
+        yield f"⚠️ Stream Interrupted: {str(e)}"
 
-# --- BRANDED UI ---
+# --- BRANDED INTERFACE ---
 custom_css = "footer {visibility: hidden} .gradio-container {background-color: #050505 !important}"
 
 demo = gr.ChatInterface(
     ui_chat,
     title="METANTHROPIC · PHI-3 SOVEREIGN",
-    description="Secure inference via decentralized sovereign weights.",
     theme=gr.themes.Soft(primary_hue="slate", neutral_hue="zinc"),
     css=custom_css
 )