fix qwen2 sliding_window logic bug

server/text_generation_server/models/__init__.py

@@ -527,6 +527,8 @@ def get_model(
 
     if model_type == "qwen2":
         sliding_window = config_dict.get("sliding_window", -1)
+        use_sliding_window = config_dict.get("use_sliding_window", False)
+        sliding_window = sliding_window if use_sliding_window else None
         if (
             (sliding_window is None or sliding_window == -1) and FLASH_ATTENTION
         ) or HAS_FLASH_ATTN_V2_CUDA:

server/text_generation_server/models/flash_qwen2.py

@@ -54,6 +54,7 @@ class FlashQwen2(BaseFlashMistral):
         )
         config.quantize = quantize
         config.use_medusa = use_medusa
+        config.sliding_window = config.sliding_window if config.use_sliding_window else None
 
         # Set context windows
         if config.sliding_window is not None: