[Bug] TVM compilation Segfaults or just fails on NVIDIA Xavier Nano

[PFernandes7]

After installing TVM correctly on this machine when trying to compile a model from Keras using relay.frontend.from_keras(model, shape_dict) defining target either as target = tvm.target.cuda() I get [17:27:26] ../src/target/target_kind.cc:163: Warning: Unable to detect CUDA version, default to "-arch=sm_20" instead or target = tvm.target.cuda(arch="sm_20") and then I get the same warning when calling lib = relay.build(mod, target=target, params=params) it again prints the same warning and exits with

Compilation error: nvcc fatal : Unsupported gpu architecture 'compute_20'

After building it from source during the cmake or make command it seems to find the correct CUDA what might be the issue here?

[santiweide]

I have met the same problem

[LeiWang1999]

I got the same issue, any update please?

[LeiWang1999]

Ok fine, I forgot to enable cuda when build tvm.