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

Bug fix for llama eagle3 (#21397)

Browse Source
This commit is contained in:
khalilzhk
2026-04-01 15:01:53 +08:00
committed by GitHub
parent 912494f596
commit 835e19656f
2 changed files with 13 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/llama.py
+7 -2
View File
@@ -252,8 +252,13 @@ class LlamaDecoderLayer(nn.Module):
) -> None: ) -> None:
super().__init__() super().__init__()
self.hidden_size = config.hidden_size self.hidden_size = config.hidden_size
rope_theta = config.rope_parameters["rope_theta"] rope_parameters = getattr(config, "rope_parameters", None)
rope_scaling = config.rope_parameters if rope_parameters is not None:
rope_theta = rope_parameters.get("rope_theta", 10000)
rope_scaling = rope_parameters
else:
rope_theta = getattr(config, "rope_theta", 10000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr( if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None config, "original_max_position_embeddings", None
): ):
python/sglang/srt/models/llama_eagle3.py
+6 -2
View File
@@ -111,13 +111,17 @@ class LlamaModel(nn.Module):
super().__init__() super().__init__()
self.config = config self.config = config
rope_scaling = config.rope_parameters rope_parameters = getattr(config, "rope_parameters", None)
if rope_parameters is not None:
rope_scaling = rope_parameters
else:
rope_scaling = getattr(config, "rope_scaling", None)
self.is_mrope_enabled = ( self.is_mrope_enabled = (
rope_scaling is not None and "mrope_section" in rope_scaling rope_scaling is not None and "mrope_section" in rope_scaling
) )
# fix rope_scaling for qwen2.5-vl # fix rope_scaling for qwen2.5-vl
if self.is_mrope_enabled: if self.is_mrope_enabled:
config.rope_parameters["rope_type"] = "default" rope_scaling["rope_type"] = "default"
self.vocab_size = config.vocab_size self.vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding( self.embed_tokens = VocabParallelEmbedding(