diff --git a/vllm/config/__init__.py b/vllm/config/__init__.py
index 700d29f956a8..fab7866a2cda 100644
--- a/vllm/config/__init__.py
+++ b/vllm/config/__init__.py
@@ -1165,8 +1165,18 @@ def _verify_quantization(self) -> None:
                     "non-quantized models.", self.quantization)
 
     def _verify_cuda_graph(self) -> None:
+        # The `max_seq_len_to_capture` was incorrectly
+        # based on the encoder's input length (448)
+        # but not the decoder's larger input length (1500).
+        # This change ensures the CUDA Graph captures the correct,
+        # larger sequence length, allowing it to work as intended.
+        effective_max_seq_len = self.max_model_len
+        if self.is_encoder_decoder:
+            effective_max_seq_len = max(
+                effective_max_seq_len,
+                getattr(self.hf_config, "max_source_positions", 0))
         self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
-                                          self.max_model_len)
+                                          effective_max_seq_len)
         # CUDAGraph capture not supported for enc-dec models and mllama on ROCm
         ROCM_UNSUPPORTED_MODELS = ['mllama']
         unsupported_rocm = (self.hf_config.model_type