diff --git a/nemo_rl/models/generation/vllm/vllm_worker.py b/nemo_rl/models/generation/vllm/vllm_worker.py
index 75e3334d4a..d2de455c22 100644
--- a/nemo_rl/models/generation/vllm/vllm_worker.py
+++ b/nemo_rl/models/generation/vllm/vllm_worker.py
@@ -388,6 +388,14 @@ def _patch_vllm_vit_flash_attn_backend():
                 )
                 # disable quantization
                 vllm_kwargs["hf_overrides"]["quantization_config"] = {}
+        elif "Gemma3ForConditionalGeneration" in getattr(
+            hf_config, "architectures", []
+        ):
+            if self.cfg["vllm_cfg"]["skip_tokenizer_init"]:
+                print(
+                    "Gemma3ForConditionalGeneration models may crash when skip_tokenizer_init is True. NeMo-RL is forcing it to False for this architecture. See https://github.com/NVIDIA-NeMo/RL/issues/1681 for more details."
+                )
+            self.cfg["vllm_cfg"]["skip_tokenizer_init"] = False
 
         llm_kwargs = dict(
             model=self.model_name,