[Bugfix] DiffusionGemma: fix multimodal support (#57589)
Signed-off-by: Matt Mastracci <matthew@mastracci.com> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Matt Mastracci committed
Sep 20, 2026 at 22:05 UTC
ff0c8854946b75dc0fe160756e8adff10a8da053
1 file changed
+5
vllm/model_executor/models/diffusion_gemma.py
+5
@@ -197,6 +197,11 @@ class DiffusionGemmaForConditionalGeneration(
197
text_config.attention_k_eq_v = True
198
199
# ---- Vision tower ----
200
+ # Gemma4's image path, borrowed below, reads this flag.
201
+ lora_config = vllm_config.lora_config
202
+ self._enable_mm_lora = bool(
203
+ lora_config is not None and lora_config.enable_tower_connector_lora
204
+ )
205
vision_config = getattr(config, "vision_config", None)
206
self.embed_vision: Gemma4MultimodalEmbedder | None
207
if vision_config is not None: