diff --git a/docling/models/rapid_ocr_model.py b/docling/models/rapid_ocr_model.py
index 24c3ff73..2dd5feb4 100644
--- a/docling/models/rapid_ocr_model.py
+++ b/docling/models/rapid_ocr_model.py
@@ -110,6 +110,9 @@ class RapidOcrModel(BaseOcrModel):
             use_cuda = str(AcceleratorDevice.CUDA.value).lower() in device
             use_dml = accelerator_options.device == AcceleratorDevice.AUTO
             intra_op_num_threads = accelerator_options.num_threads
+            gpu_id = 0
+            if use_cuda and ":" in device:
+                gpu_id = int(device.split(":")[1])
             _ALIASES = {
                 "onnxruntime": EngineType.ONNXRUNTIME,
                 "openvino": EngineType.OPENVINO,
@@ -184,6 +187,10 @@ class RapidOcrModel(BaseOcrModel):
                 "Det.engine_type": backend_enum,
                 "Cls.engine_type": backend_enum,
                 "Rec.engine_type": backend_enum,
+                "EngineConfig.paddle.use_cuda": use_cuda,
+                "EngineConfig.paddle.gpu_id": gpu_id,
+                "EngineConfig.torch.use_cuda": use_cuda,
+                "EngineConfig.torch.gpu_id": gpu_id,
             }
 
             if self.options.rec_font_path is not None:
diff --git a/docs/usage/gpu.md b/docs/usage/gpu.md
index 7ca1576e..012f6568 100644
--- a/docs/usage/gpu.md
+++ b/docs/usage/gpu.md
@@ -43,6 +43,21 @@ Setting a higher `page_batch_size` will run the Docling models (in particular th
 
 For a complete example see [gpu_standard_pipeline.py](../examples/gpu_standard_pipeline.py).
 
+#### OCR engines
+
+The current Docling OCR engines rely on third-party libraries, hence GPU support depends on the availability in the respective engines.
+
+The only setup which is known to work at the moment is RapidOCR with the torch backend, which can be enabled via
+
+```py
+pipeline_options = PdfPipelineOptions()
+pipeline_options.ocr_options = RapidOcrOptions(
+    backend="torch",
+)
+```
+
+More details in the GitHub discussion [#2451](https://github.com/docling-project/docling/discussions/2451).
+
 
 ### VLM Pipeline