Update app.py

app.py

@@ -2,6 +2,9 @@ import os
 import sys
 import subprocess
 import site
+from fastapi import FastAPI, Request
+from fastapi.responses import JSONResponse
+import uvicorn
 
 # 1. BIND TO THE PERSISTENT COMPILATION REGISTRY
 PERSISTENT_PACKAGES = "/data/compiled_cache"
@@ -34,7 +37,7 @@ import gradio as gr
 from huggingface_hub import hf_hub_download
 import spaces
 
-# 2. UPDATED MODEL WORKSPACE INDEXES (Uses your exact link maps!)
+# 2. MODEL WORKSPACE INDEXES
 print("Checking persistent storage for AI model weights...")
 
 # Model 1: The Main 27B Monster for GPU (3.9 GB)
@@ -44,8 +47,7 @@ path_27b = hf_hub_download(
     local_dir="/data"
 )
 
-# Model 2: YOUR EXACT WORKING LINK SWAP 
-# Points precisely to the repo and file schema you provided to eliminate the 404 crash!
+# 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", 
@@ -74,7 +76,6 @@ def generate_moe_cpu(prompt):
     clean_prompt = str(prompt).strip()
     if not clean_prompt: return "Empty prompt."
     
-    # 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"
     
@@ -98,5 +99,45 @@ with gr.Blocks(title="Resilient AI Hub") as demo:
         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()
+
+@fastapi_app.post("/v1/chat/completions")
+async def openai_endpoints_router(request: Request):
+    try:
+        json_data = await request.json()
+        messages = json_data.get("messages", [])
+        user_prompt = messages[-1]["content"] if messages else ""
+        chosen_model = json_data.get("model", "bonsai-27b")
+    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
+            },
+            "finish_reason": "stop"
+        }]
+    })
+
+# Bind Gradio web app structure paths directly to the root of the server instance
+app = gr.mount_gradio_app(fastapi_app, demo, path="/")
+
 if __name__ == "__main__":
-    demo.launch()
+    uvicorn.run(app, host="0.0.0.0", port=7860)