chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Clean up swa handling in fa3 backend (#15877)

Browse Source
This commit is contained in:
Ke Bao
2025-12-30 08:51:40 +08:00
committed by GitHub
parent f4ec6f8e17
commit 3de23274ee
2 changed files with 35 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
python/sglang/srt/model_executor/model_runner.py
View File

@@ -440,16 +440,6 @@ class ModelRunner(ModelRunnerKVCacheMixin):
self.model, self.remote_instance_transfer_engine
)
# Check if the model is using hybrid SWA
if (
not self.server_args.disable_hybrid_swa_memory
and self.sliding_window_size is not None
and self.sliding_window_size > 0
):
architectures = self.model_config.hf_config.architectures
if architectures and not any("Llama4" in arch for arch in architectures):
self.is_hybrid_swa = self.model_config.is_hybrid_swa = True
# For MTP models like DeepSeek-V3 or GLM-4.5, the MTP layer(s) are used separately as draft
# models for speculative decoding. In those cases, `num_nextn_predict_layers` is used to
# determine the number of layers.
@@ -941,6 +931,12 @@ class ModelRunner(ModelRunnerKVCacheMixin):
self.sliding_window_size = None
if hasattr(self.model, "get_attention_sliding_window_size"):
self.sliding_window_size = self.model.get_attention_sliding_window_size()
elif (
self.model_config.is_hybrid_swa
and self.model_config.sliding_window_size is not None
):
# sliding window field in model config may have different meaning for different kinds of models (e.g., dllm), here we only consider the sliding window in SWA model
self.sliding_window_size = self.model_config.sliding_window_size
elif self.model_config.attention_chunk_size is not None:
self.sliding_window_size = self.model_config.attention_chunk_size
logger.info(