EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Model] Adding Qwen3 and Qwen3MoE (#4693)

Browse Source
This commit is contained in:
yhyang201
2025-04-19 00:51:29 +08:00
committed by GitHub
parent bfa3922451
commit 4db463b1ad
5 changed files with 780 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/models/qwen2.py
View File

@@ -239,6 +239,7 @@ class Qwen2Model(nn.Module):
config: Qwen2Config,
quant_config: Optional[QuantizationConfig] = None,
prefix: str = "",
decoder_layer_type: type[nn.Module] = Qwen2DecoderLayer,
) -> None:
super().__init__()
self.config = config
@@ -250,9 +251,11 @@ class Qwen2Model(nn.Module):
quant_config=quant_config,
prefix=add_prefix("embed_tokens", prefix),
)
# Use the provided decoder layer type or default to Qwen2DecoderLayer
decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
self.layers = make_layers(
config.num_hidden_layers,
lambda idx, prefix: Qwen2DecoderLayer(
lambda idx, prefix: decoder_layer_type(
layer_id=idx,
config=config,
quant_config=quant_config,