Update app.py

app.py

@@ -56,8 +56,8 @@ async def lifespan(app: FastAPI):
     
     print("Downloading MoE model GGUF...")
     model_path = hf_hub_download(
-        repo_id='Jackrong/Qwopus3.6-35B-A3B-Coder-MTP-GGUF',
-        filename="Qwopus3.6-35B-A3B-Coder-MTP-Q3_K_M.gguf",
+        repo_id='unsloth/GLM-4.7-Flash-REAP-23B-A3B-GGUF',
+        filename="GLM-4.7-Flash-REAP-23B-A3B-Q3_K_S.gguf",
     )
     print("Model downloaded successfully!")
 
@@ -80,7 +80,7 @@ async def lifespan(app: FastAPI):
         
         n_batch=512,    
         n_ubatch=512,   
-        n_draft=3,
+        #n_draft=3,
         
         # ⚡ A3B MoE OPTIMIZATION: 
         # use_mmap=True ensures the OS only pages the active 3B parameters