Update app.py

app.py

@@ -15,7 +15,7 @@ if TARGET_SITE_PATH not in sys.path:
     sys.path.insert(0, TARGET_SITE_PATH)
     site.addsitedir(TARGET_SITE_PATH)
 
-# Check and boot the persistent engine compiled in your previous run
+# Pull the pre-compiled llama-cpp wheel built in your previous runtime pass
 try:
     from llama_cpp import Llama
     print("🚀 Perfect! Pre-compiled engine found in /data. Loading instantly...")
@@ -34,7 +34,7 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 import spaces
 
-# 2. FIXED MODEL PATH MAPS (Bypasses the 404 Error)
+# 2. UPDATED MODEL WORKSPACE INDEXES (Uses your exact link maps!)
 print("Checking persistent storage for AI model weights...")
 
 # Model 1: The Main 27B Monster for GPU (3.9 GB)
@@ -44,19 +44,19 @@ path_27b = hf_hub_download(
     local_dir="/data"
 )
 
-# Model 2: FIXED REPOSITORY PATH AND FILE LAYOUT
-# Changes repo to LiquidAI and filename to use dot quantization separation to fix the 404 crash!
+# Model 2: YOUR EXACT WORKING LINK SWAP 
+# Points precisely to the repo and file schema you provided to eliminate the 404 crash!
 path_moe = hf_hub_download(
-    repo_id="LiquidAI/LFM2.5-8B-A1B-GGUF", 
-    filename="LFM2.5-8B-A1B.Q4_K_M.gguf", 
+    repo_id="LiquidAI/LFM2-8B-A1B-GGUF", 
+    filename="LFM2-8B-A1B-Q4_K_M.gguf", 
     local_dir="/data"
 )
 
-# Load the MoE model globally on the free CPU thread so it never falls asleep
+# Initialize the 8B MoE model to the unmetered CPU thread
 print("Initializing Liquid 8B MoE on active CPU thread...")
 llm_cpu = Llama(model_path=path_moe, n_ctx=4096, n_gpu_layers=0, verbose=False)
 
-# 3. ENDPOINT FUNCTION WORKFLOWS
+# 3. ENDPOINT Workflows
 @spaces.GPU(duration=60)
 def generate_27b(prompt):
     clean_prompt = str(prompt).strip()
@@ -74,7 +74,7 @@ def generate_moe_cpu(prompt):
     clean_prompt = str(prompt).strip()
     if not clean_prompt: return "Empty prompt."
     
-    # Liquid AI specific chat markers
+    # Liquid AI specific instruction tags
     system_tool_prompt = "You are an advanced AI agent with Tool Calling capabilities."
     formatted = f"<|im_start|>system\n{system_tool_prompt}<|im_end|>\n<|im_start|>user\n{clean_prompt}<|im_end|>\n<|im_start|>assistant\n"
     
@@ -82,7 +82,7 @@ def generate_moe_cpu(prompt):
     try: return response["choices"]["text"]
     except: return str(response)
 
-# 4. GRADIO DUAL TAB REGISTRY
+# 4. GRADIO MULTI-TAB MAPPING DESIGN
 with gr.Blocks(title="Resilient AI Hub") as demo:
     gr.Markdown("# 🌳 Unstoppable Split-Brain AI Hub")