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

handling incomplete rope_scaling config ci after transformers upgrade (#14784)

Browse Source
This commit is contained in:
Yuhao Yang
2025-12-09 22:56:16 -08:00
committed by GitHub
parent b1cbfce612
commit 793c98afaf
1 changed files with 6 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_v2.py
+6
View File
@@ -2704,6 +2704,12 @@ class DeepseekV2DecoderLayer(nn.Module):
self.config = config self.config = config
rope_theta = getattr(config, "rope_theta", 10000) rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None) rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None:
# In transformers 5.0.0rc0+, rope_theta and rope_type are also included in rope_scaling.
# Therefore, if rope_scaling contains only these two keys,
# it effectively means there are no special rope_scaling parameters.
if set(rope_scaling.keys()) <= {"rope_theta", "rope_type"}:
rope_scaling = None
max_position_embeddings = getattr(config, "max_position_embeddings", 8192) max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
self.speculative_algorithm = SpeculativeAlgorithm.from_string( self.speculative_algorithm = SpeculativeAlgorithm.from_string(
get_global_server_args().speculative_algorithm get_global_server_args().speculative_algorithm