From b1707996e8c50b921a91c815d28fe44fe54d8dea Mon Sep 17 00:00:00 2001 From: Ayushman Singh <40520701+ayush1399@users.noreply.github.com> Date: Wed, 19 Aug 2026 15:41:52 -0400 Subject: [PATCH] fix(gemma4): quantize MTP bridge projections (#32440) --- python/sglang/srt/models/gemma4_mtp.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/models/gemma4_mtp.py b/python/sglang/srt/models/gemma4_mtp.py index f652661fc..3978a1be1 100644 --- a/python/sglang/srt/models/gemma4_mtp.py +++ b/python/sglang/srt/models/gemma4_mtp.py @@ -91,7 +91,7 @@ class Gemma4AssistantForCausalLM(Gemma4ForCausalLM): 2 * self.backbone_hidden_size, self.hidden_size, bias=False, - quant_config=None, + quant_config=quant_config, prefix=add_prefix("pre_projection", prefix), ) self.model = Gemma4TextModel( @@ -103,7 +103,7 @@ class Gemma4AssistantForCausalLM(Gemma4ForCausalLM): self.hidden_size, self.backbone_hidden_size, bias=False, - quant_config=None, + quant_config=quant_config, prefix=add_prefix("post_projection", prefix), )