The TFLite model trained with LiteRT cannot run on the GPU.

[luvvien]

Have I written custom code (as opposed to using a stock example script provided in MediaPipe)

None

OS Platform and Distribution

iPhone12

Python Version

3.8

MediaPipe Model Maker version

None

Task name (e.g. Image classification, Gesture recognition etc.)

Object detect

Describe the actual behavior

Switch to GPU，throw error

Describe the expected behaviour

Run on GPU

Standalone code/steps you may have used to try to get what you need

The official MediaPipe demo deployed on iPhone 12 can switch to GPU for execution, but the TFLite model trained with LiteRT can only run on the CPU, and switching to GPU results in an error.

Other info / Complete Logs

E0000 00:00:1741774904.900263 5036181 calculator_graph.cc:887] INTERNAL: CalculatorGraph::Run() failed: 
Calculator::Open() for node "mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__inferencecalculator__mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__InferenceCalculator" failed: ; RET_CHECK failure (mediapipe/calculators/tensor/inference_calculator_metal.cc:310) (TFLGpuDelegateBindMetalBufferToTensor( delegate_.get(), output_indices[i], MtlBufferView::GetWriteView(*gpu_buffers_out_[i], gpu_helper_.mtlDevice) .buffer()))==(true)
Error Domain=com.google.mediapipe.tasks Code=13 "INTERNAL: CalculatorGraph::Run() failed: 
Calculator::Open() for node "mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__inferencecalculator__mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__InferenceCalculator" failed: ; RET_CHECK failure (mediapipe/calculators/tensor/inference_calculator_metal.cc:310) (TFLGpuDelegateBindMetalBufferToTensor( delegate_.get(), output_indices[i], MtlBufferView::GetWriteView(*gpu_buffers_out_[i], gpu_helper_.mtlDevice) .buffer()))==(true)" UserInfo={NSLocalizedDescription=INTERNAL: CalculatorGraph::Run() failed: 
Calculator::Open() for node "mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__inferencecalculator__mediapipe_tasks_vision_objectdetectorgraph__mediapipe_tasks_core_inferencesubgraph__InferenceCalculator" failed: ; RET_CHECK failure (mediapipe/calculators/tensor/inference_calculator_metal.cc:310) (TFLGpuDelegateBindMetalBufferToTensor( delegate_.get(), output_indices[i], MtlBufferView::GetWriteView(*gpu_buffers_out_[i], gpu_helper_.mtlDevice) .buffer()))==(true)}

[kuaashish]

Hi @luvvien,

This issue appears to be more related to LiteRT rather than MediaPipe, as the TFLite model trained with LiteRT cannot run on the GPU. It would be more effective to raise the issue in the TensorFlow repository here and follow up with them for further updates.

Thank you!!