Upload 5 files

main.py

@@ -311,17 +311,30 @@ if not os.path.exists(model_path):
     print(f"[Gateway] Đang tải {model_path} (khoảng 533MB)...")
     model_path = hf_hub_download(repo_id="unsloth/Qwen3.5-0.8B-GGUF", filename="Qwen3.5-0.8B-Q4_K_M.gguf", local_dir=".")
 
-try:
-    from llama_cpp import Llama
-    llm = Llama(
-        model_path=model_path,
-        n_ctx=2048,
-        n_threads=2, # Tối ưu hóa cho máy 2 vCPU
-        verbose=False
-    )
-except ImportError:
-    print("[Gateway] Cảnh báo: Chưa cài đặt llama-cpp-python. Fallback to mock.")
-    llm = None
+llm = None
+
+def load_llm(lora_path=None):
+    global llm
+    try:
+        from llama_cpp import Llama
+        print(f"[Gateway] Đang nạp mô hình vào RAM (LoRA: {lora_path})...")
+        llm = Llama(
+            model_path=model_path,
+            lora_path=lora_path,
+            n_ctx=2048,
+            n_threads=2, # Tối ưu hóa cho máy 2 vCPU
+            verbose=False
+        )
+        print("[Gateway] Nạp mô hình thành công!")
+    except ImportError:
+        print("[Gateway] Cảnh báo: Chưa cài đặt llama-cpp-python. Fallback to mock.")
+        llm = None
+    except Exception as e:
+        print(f"[Gateway] Lỗi nạp mô hình: {e}")
+        llm = None
+
+# Khởi tạo lần đầu
+load_llm()
 
 import re
 
@@ -418,9 +431,33 @@ def read_root():
     return {
         "status": "online", 
         "service": "EvoNet B2B API Gateway",
-        "message": "Gateway is running. Send POST requests to /v1/chat/completions"
+        "message": "Gateway is running. Send POST requests to /v1/chat/completions",
+        "lora_supported": True
     }
 
+class LoadLoraRequest(BaseModel):
+    repo_id: Optional[str] = None
+    filename: str
+
+@app.post("/v1/admin/load-lora")
+async def api_load_lora(req: LoadLoraRequest, token: str = Depends(verify_b2b_token)):
+    """API để nạp nóng Adapter LoRA từ Hugging Face (Dành cho Universal Agent)"""
+    try:
+        hf_token = os.environ.get("HF_ACCESS_TOKEN", None)
+        target_repo = req.repo_id if req.repo_id else os.environ.get("HF_LORA_REPO")
+        
+        if not target_repo:
+            raise HTTPException(status_code=400, detail="Không có repo_id. Vui lòng truyền repo_id hoặc cấu hình HF_LORA_REPO")
+            
+        print(f"[Gateway Admin] Yêu cầu nạp LoRA từ {target_repo}/{req.filename} (Private Mode: {bool(hf_token)})...")
+        lora_local_path = hf_hub_download(repo_id=target_repo, filename=req.filename, local_dir=".", token=hf_token)
+        # Chạy trong luồng riêng để không block API
+        await asyncio.to_thread(load_llm, lora_local_path)
+        return {"success": True, "message": f"Đã nạp thành công LoRA: {req.filename}"}
+    except Exception as e:
+        print(f"[Gateway Admin] Lỗi tải LoRA: {e}")
+        raise HTTPException(status_code=500, detail=str(e))
+
 @app.post("/v1/chat/completions")
 async def chat_completions(req: ChatCompletionRequest, request: Request, token: str = Depends(verify_b2b_token)):
     if not req.messages:
@@ -471,11 +508,11 @@ async def chat_completions(req: ChatCompletionRequest, request: Request, token:
                 raise Exception("Swarm returned non-200 after retries")
         except Exception as e:
             print(f"[Gateway Routing] Lỗi từ DePIN Swarm ({e}), chuyển sang Fallback Local.")
-            result_text = f"[Extreme JIT] {await asyncio.to_thread(fallback_infer, messages_dicts)}"
+            result_text = f"{await asyncio.to_thread(fallback_infer, messages_dicts)}"
     else:
         # Step 2b: Fallback to Local Extreme Batching
         print(f"[Gateway Routing] ➡️ DePIN bận/thiếu Node. Chuyển hướng xử lý tại Cụm Server Nội bộ.")
-        result_text = f"[Extreme JIT] {await asyncio.to_thread(fallback_infer, messages_dicts)}"
+        result_text = f"{await asyncio.to_thread(fallback_infer, messages_dicts)}"
 
     # Step 3: Format OpenAI response
     prompt_tokens = len(last_message.split())