{"id":19285,"date":"2026-05-22T09:19:14","date_gmt":"2026-05-22T13:19:14","guid":{"rendered":"https:\/\/blogs.mathworks.com\/deep-learning\/?p=19285"},"modified":"2026-05-22T09:19:14","modified_gmt":"2026-05-22T13:19:14","slug":"from-pytorch-litert-to-c-c-and-cuda-source-code","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/deep-learning\/2026\/05\/22\/from-pytorch-litert-to-c-c-and-cuda-source-code\/","title":{"rendered":"From PyTorch & LiteRT to C, C++, and CUDA source code"},"content":{"rendered":"
\r\n
\r\n
<\/div>\r\n
\r\n
<\/h6>\r\n\r\n\r\n\r\n
Guest writer: Christoph Stockhammer<\/a><\/strong>\r\n
<\/h6>\r\nChristoph Stockhammer <\/span>is an application engineer at MathWorks, focusing on AI use cases. Christoph holds a Master's in Mathematics from the Technical University of Munich.<\/td>\r\n<\/tr>\r\n<\/tbody>\r\n<\/table>\r\n
<\/h6>\r\n<\/div>\r\n
<\/div>\r\n<\/div>\r\n
<\/div>\r\n
\r\n
If you have ever tried to deploy a PyTorch model on embedded hardware, you know the story: the model itself is only half the battle. The real challenge starts when you need a runtime, shared libraries, hardware-specific builds, and a toolchain that somehow still fits on your target device.<\/div>\r\n
Starting with MATLAB R2026a, there is an alternative route. Instead of shipping a runtime to the device, you can now generate standalone C\/C++ (and CUDA®) source code<\/strong> directly from PyTorch ExportedProgram<\/strong> and LiteRT<\/strong> models. No interpreter. No inference engine. No import process into MATLAB. Just readable and portable source code that you compile with your existing toolchain.<\/div>\r\n

Why standalone code generation?<\/h2>\r\n
Embedded targets--from microcontrollers to devices like Raspberry Pi or NVIDIA® Jetson™--care deeply about predictability. Memory usage, startup time, and binary size often matter more than flexibility. Runtime-based solutions such as LiteRT (for Microcontrollers) or ONNX Runtime do a great job, but they still bring dependencies and abstraction layers that require more processing power and memory usage above and beyond pure source code.<\/div>\r\n
Standalone code generation removes that layer entirely. The generated code contains only what your model needs: loops, math, and data. That makes it easier to analyze, debug, certify, and integrate with existing embedded software.<\/div>\r\n

Under the hood: MLIR as the secret sauce<\/h2>\r\n
The key technology behind this workflow is MLIR (Multi-Level Intermediate Representation)<\/strong>. MATLAB Coder lowers PyTorch ExportedProgram and LiteRT models into MLIR<\/a>, applies a series of graph- and hardware-aware optimizations, then emits C\/C++ or CUDA source code.<\/div>\r\n
Because the optimizations happen at the IR level, the generated code can take advantage of:<\/div>\r\n