A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 09:45 UTC
[Model] Allow the use of sliding window in Qwen2 (#17772)
Signed-off-by: inkcherry <mingzhi.liu@intel.com>
inkcherry committed
May 15, 2025 at 13:29 UTC
dd2a94596abe43eb0e556cde306fc80d442183c0
1 file changed
+8
-8
vllm/model_executor/models/qwen2.py
+8
-8
@@ -291,14 +291,14 @@ class Qwen2Model(nn.Module):
291
# TODO (@robertgshaw2): see if this can be moved out
292
if (cache_config.sliding_window is not None
293
and hasattr(config, "max_window_layers")):
294
- raise ValueError("Sliding window for some but all layers is not "
295
- "supported. This model uses sliding window "
296
- "but `max_window_layers` = {} is less than "
297
- "`num_hidden_layers` = {}. Please open an issue "
298
- "to discuss this feature.".format(
299
- config.max_window_layers,
300
- config.num_hidden_layers,
301
- ))
294
+ assert config.max_window_layers == config.num_hidden_layers, (
295
+ "Sliding window for some but all layers is not supported. "
296
+ "This model uses sliding window but `max_window_layers` = {} "
297
+ "is less than `num_hidden_layers` = {}. Please open an issue "
298
+ "to discuss this feature.".format(
299
+ config.max_window_layers,
300
+ config.num_hidden_layers,
301
+ ))
302
303
self.config = config
304
self.quant_config = quant_config