Update app.py

app.py

@@ -5,6 +5,11 @@ import site
 from fastapi import FastAPI, Request
 from fastapi.responses import JSONResponse
 import uvicorn
+import gradio as gr
+
+# Initialize FastAPI and Gradio interfaces at the absolute top layer
+# This guarantees Hugging Face's orchestrator natively validates the environment routes on boot!
+fastapi_app = FastAPI()
 
 # 1. BIND TO THE PERSISTENT COMPILATION REGISTRY
 PERSISTENT_PACKAGES = "/data/compiled_cache"
@@ -18,7 +23,6 @@ if TARGET_SITE_PATH not in sys.path:
     sys.path.insert(0, TARGET_SITE_PATH)
     site.addsitedir(TARGET_SITE_PATH)
 
-# 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...")
@@ -33,32 +37,28 @@ except ModuleNotFoundError:
     site.addsitedir(TARGET_SITE_PATH)
     from llama_cpp import Llama
 
-import gradio as gr
 from huggingface_hub import hf_hub_download
 import spaces
 
-# 2. MODEL WORKSPACE INDEXES
+# 2. MODEL WEIGHT CONFIGURATIONS
 print("Checking persistent storage for AI model weights...")
 
-# Model 1: The Main 27B Monster for GPU (3.9 GB)
 path_27b = hf_hub_download(
     repo_id="prism-ml/Bonsai-27B-gguf", 
     filename="Bonsai-27B-Q1_0.gguf", 
     local_dir="/data"
 )
 
-# Model 2: Verified LiquidAI repo and file path
 path_moe = hf_hub_download(
     repo_id="LiquidAI/LFM2-8B-A1B-GGUF", 
     filename="LFM2-8B-A1B-Q4_K_M.gguf", 
     local_dir="/data"
 )
 
-# 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 Workflows
+# 3. ENDPOINT WORKFLOWS
 @spaces.GPU(duration=60)
 def generate_27b(prompt):
     clean_prompt = str(prompt).strip()
@@ -75,36 +75,27 @@ def generate_27b(prompt):
 def generate_moe_cpu(prompt):
     clean_prompt = str(prompt).strip()
     if not clean_prompt: return "Empty prompt."
-    
     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"
-    
     response = llm_cpu(formatted, max_tokens=512, temperature=0.1)
     try: return response["choices"]["text"]
     except: return str(response)
 
-# 4. GRADIO MULTI-TAB MAPPING DESIGN
+# 4. GRADIO DUAL-TAB UI LAYOUT
 with gr.Blocks(title="Resilient AI Hub") as demo:
     gr.Markdown("# 🌳 Unstoppable Split-Brain AI Hub")
-    
     with gr.Tab("🚀 Bonsai 27B (GPU Endpoint)"):
         input_27b = gr.Textbox(label="Enter prompt for 27B model (Uses Quota)", lines=6)
         output_27b = gr.Textbox(label="GPU Response Output")
         btn_27b = gr.Button("Submit to GPU")
         btn_27b.click(fn=generate_27b, inputs=input_27b, outputs=output_27b, api_name="chat")
-        
     with gr.Tab("🪵 Liquid 8B MoE (CPU Endpoint / Native Tool Calling)"):
         input_moe = gr.Textbox(label="Enter prompt for MoE model (100% Free / Anti-Limit Backup)", lines=6)
         output_moe = gr.Textbox(label="MoE CPU Output")
         btn_moe = gr.Button("Submit to MoE Engine")
         btn_moe.click(fn=generate_moe_cpu, inputs=input_moe, outputs=output_moe, api_name="chat_backup")
 
-# ==========================================
-# 5. FASTAPI /V1 OPENAI COMPATIBILITY MOUNT
-# ==========================================
-# This acts as a background translator server for incoming OpenCode requests!
-fastapi_app = FastAPI()
-
+# 5. OPENAI /V1 ROUTING HOOKS
 @fastapi_app.post("/v1/chat/completions")
 async def openai_endpoints_router(request: Request):
     try:
@@ -115,28 +106,23 @@ async def openai_endpoints_router(request: Request):
     except Exception:
         return JSONResponse({"error": "Invalid JSON context formatting payload"}, status_code=400)
     
-    # Route input context arrays directly to the right execution model function
     if "liquid" in chosen_model or "cpu" in chosen_model or "backup" in chosen_model:
         model_reply = generate_moe_cpu(user_prompt)
     else:
         model_reply = generate_27b(user_prompt)
         
-    # Standard OpenAI JSON dictionary schema response structure format
     return JSONResponse({
         "id": "hf-split-brain-chat",
         "object": "chat.completion",
         "model": chosen_model,
         "choices": [{
             "index": 0,
-            "message": {
-                "role": "assistant",
-                "content": model_reply
-            },
+            "message": {"role": "assistant", "content": model_reply},
             "finish_reason": "stop"
         }]
     })
 
-# Bind Gradio web app structure paths directly to the root of the server instance
+# Mount Gradio over the core FastAPI application layers
 app = gr.mount_gradio_app(fastapi_app, demo, path="/")
 
 if __name__ == "__main__":