[SOLVED] Cross Compiling for CUDA without CUDA device

I have following setup:

  • Laptop without CUDA device
  • Jetson TX2

I want to compile a Keras model for the TX2 on my laptop. When I build TVM with set(USE_CUDA OFF) I get the error (on relay.build(..)):

TVMError: Check failed: bf != nullptr: Target cuda -model=tx2 is not enabled

When I compile tvm with set(USE_CUDA ON), I get the error

ValueError: arch(sm_xy) is not passed, and we cannot detect it from env

So my question is: Is it possible to compile a (Keras) model for target="cuda" without a CUDA device on your local machine?

See [SOLVED] Compile error related to autotvm on how to solve this issue.

ping @eqy I cannot edit my original post or the subject (probably because it’s too old?). Can you mark this as [SOLVED]?