diff --git a/src/inference_server/backends/custom_torch_backend.py b/src/inference_server/backends/custom_torch_backend.py
index 6dc5b17..1cc54ce 100644
--- a/src/inference_server/backends/custom_torch_backend.py
+++ b/src/inference_server/backends/custom_torch_backend.py
@@ -264,7 +264,7 @@ def load_model(self, model_name: str) -> None:
             self.model = model_cls.from_safetensors(model_name, device=self.device)
         else:
             self.model = model_cls.from_hf(model_name, dtype=torch.bfloat16).to(self.device).eval()
-        self.max_positions = int(getattr(AutoConfig.from_pretrained(model_name), "max_position_embeddings", 0) or 0)
+        self.max_positions = int(getattr(AutoConfig.from_pretrained(model_name).get_text_config(), "max_position_embeddings", 0) or 0)   # Gemma 4 keeps it on the text config
         self._eos_ids = stop_token_ids(model_name, self.tokenizer)
 
         # Weight-only int8: store Linear weights as int8 (per-channel), read 2× fewer bytes/step.