Update app.py

app.py

@@ -6,9 +6,11 @@ os.system("ulimit -l unlimited")
 
 app = FastAPI()
 
-hf_hub_download("TheBloke/deepseek-coder-6.7B-base-GGUF", "deepseek-coder-6.7b-base.Q5_K_M.gguf", local_dir="./")
+hf_hub_download("TheBloke/deepseek-coder-1.3b-base-GGUF", "deepseek-coder-1.3b-base.Q5_K_M.gguf", local_dir="./")
+hf_hub_download("TheBloke/deepseek-coder-6.7B-base-GGUF", "deepseek-coder-6.7b-base.Q4_K_M.gguf", local_dir="./")
 
-model_l = Llama(model_path="./deepseek-coder-6.7b-base.Q5_K_M.gguf", n_ctx=16000, n_gpu_layers=0, n_threads=2, use_mlock=True)
+model_l = Llama(model_path="./deepseek-coder-1.3b-base.Q5_K_M.gguf", n_ctx=16000, n_gpu_layers=0, n_threads=2, use_mlock=True)
+model_f = Llama(model_path="./deepseek-coder-6.7b-base.Q5_K_M.gguf", n_ctx=16000, n_gpu_layers=0, n_threads=2, use_mlock=True)
 
 @app.get("/check")
 async def index():
@@ -18,12 +20,21 @@ async def index():
 async def completion(request: Request):
     data = await request.json()
     prompt = data["prompt"]
-    res = model_l(
-        prompt,
-        temperature=0.6,
-        echo=False,
-        max_tokens=41,
-    )
+    mode = data['mode']
+    if mode == 'f':
+        res = model_f(
+            prompt,
+            temperature=0.6,
+            echo=False,
+            max_tokens=41,
+        )
+    else:
+        res = model_l(
+            prompt,
+            temperature=0.6,
+            echo=False,
+            max_tokens=41,
+        )
     return {"responses": res["choices"]}
 
 if __name__ == "__main__":