CUDA Errors
A quick index of CUDA-related errors. The full explanations, with causes and fixes, are in Troubleshooting CUDA.
| Error text | Where to look |
|---|---|
cuda: cudaMemcpyToSymbol (matrix) failed: the provided PTX was compiled with an unsupported toolchain. (cudaError_t 222) | The package was built without CMAKE_CUDA_ARCHITECTURES. See the fix |
cuda: cudaMalloc failed: out of memory | Device memory was claimed after MENOS measured it. See the fix |
| The job runs on the CPU although a GPU is present | See why |
CMake picks the wrong nvcc | See the fix |
| Device properties look wrong | See the fix |
How CUDA errors surface
- A request that does not fit the device budget never errors: the planner keeps the region on the CPU.
- A CUDA failure that happens despite the budget check surfaces as
InternalErrorwith the CUDA error text, because it means something MENOS could not predict went wrong. - A build with no CUDA support never attempts GPU work.