From ff0c8854946b75dc0fe160756e8adff10a8da053 Mon Sep 17 00:00:00 2001 From: Matt Mastracci Date: Sun, 20 Sep 2026 22:05:24 -0600 Subject: [PATCH] [Bugfix] DiffusionGemma: fix multimodal support (#57589) Signed-off-by: Matt Mastracci Co-authored-by: Claude Fable 5.1 Co-authored-by: Lucas Wilkinson Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/model_executor/models/diffusion_gemma.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/model_executor/models/diffusion_gemma.py b/vllm/model_executor/models/diffusion_gemma.py index 10e0d8a2c2..aef9d81625 100644 --- a/vllm/model_executor/models/diffusion_gemma.py +++ b/vllm/model_executor/models/diffusion_gemma.py @@ -197,6 +197,11 @@ class DiffusionGemmaForConditionalGeneration( text_config.attention_k_eq_v = True # ---- Vision tower ---- + # Gemma4's image path, borrowed below, reads this flag. + lora_config = vllm_config.lora_config + self._enable_mm_lora = bool( + lora_config is not None and lora_config.enable_tower_connector_lora + ) vision_config = getattr(config, "vision_config", None) self.embed_vision: Gemma4MultimodalEmbedder | None if vision_config is not None: