From da1daf40bf18e5eaae04f26a80a537c8168a8bc2 Mon Sep 17 00:00:00 2001 From: Luciano Martins <22145370+lucianommartins@users.noreply.github.com> Date: Fri, 5 Jun 2026 00:47:38 -0300 Subject: [PATCH] [Bugfix] Exclude vision embedder from quantization in Gemma4 Unified (#44571) Signed-off-by: Luciano Martins Co-authored-by: Luciano Martins --- vllm/model_executor/models/gemma4_unified.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/gemma4_unified.py b/vllm/model_executor/models/gemma4_unified.py index 66fc914dc7..9cc0710c4d 100644 --- a/vllm/model_executor/models/gemma4_unified.py +++ b/vllm/model_executor/models/gemma4_unified.py @@ -80,7 +80,7 @@ class Gemma4UnifiedVisionEmbedder(nn.Module): Pipeline: raw patches → LN₁ → Dense → LN₂ → +factorized_posemb → LN₃. """ - def __init__(self, config, quant_config=None): + def __init__(self, config, quant_config=None, prefix=""): super().__init__() patch_dim = config.model_patch_size**2 * 3 mm_embed_dim = config.mm_embed_dim @@ -91,6 +91,7 @@ class Gemma4UnifiedVisionEmbedder(nn.Module): mm_embed_dim, bias=True, quant_config=quant_config, + prefix=f"{prefix}.patch_dense", gather_output=True, ) self.patch_ln2 = nn.LayerNorm(mm_embed_dim) @@ -267,6 +268,7 @@ class Gemma4UnifiedForConditionalGeneration(Gemma4ForConditionalGeneration): Gemma4UnifiedVisionEmbedder( config.vision_config, quant_config=quant_config, + prefix=maybe_prefix(prefix, "vision_embedder"), ) if config.vision_config is not None else None