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

fix(gemma4): quantize MTP bridge projections (#32440)

Browse Source
This commit is contained in:
Ayushman Singh
2026-08-19 12:41:52 -07:00
committed by GitHub
parent 6f69f927da
commit b1707996e8
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/gemma4_mtp.py
+2 -2
View File
@@ -91,7 +91,7 @@ class Gemma4AssistantForCausalLM(Gemma4ForCausalLM):
2 * self.backbone_hidden_size, 2 * self.backbone_hidden_size,
self.hidden_size, self.hidden_size,
bias=False, bias=False,
quant_config=None, quant_config=quant_config,
prefix=add_prefix("pre_projection", prefix), prefix=add_prefix("pre_projection", prefix),
) )
self.model = Gemma4TextModel( self.model = Gemma4TextModel(
@@ -103,7 +103,7 @@ class Gemma4AssistantForCausalLM(Gemma4ForCausalLM):
self.hidden_size, self.hidden_size,
self.backbone_hidden_size, self.backbone_hidden_size,
bias=False, bias=False,
quant_config=None, quant_config=quant_config,
prefix=add_prefix("post_projection", prefix), prefix=add_prefix("post_projection", prefix),
) )