masahi commented on issue #9939: URL: https://github.com/apache/tvm/issues/9939#issuecomment-1013754272
The error is ` Check failed: (pval != nullptr) is false: Cannot allocate memory symbolic tensor shape [?, 1280, 9, 16]`. When you do `relay.build`, we use something called graph codegen / runtime, where all shapes need to be statically known. The `?` indicates your model has a dynamic batch dimension somewhere, so you need to use a compiler and runtime that supports dynamic shape. That's what the VM compiler and runtime for. > @masahi does VM compiler support metal? Yes, there is no coupling between targets and the VM runtime, all targets work. > I actually need to consume the model in c++, so I'm assuming all of the above is possible with c++ API as well In principle yes. Deployment in C++ is common, but less so for VM. There is an example in https://github.com/apache/tvm/blob/e7024fb39ea27494fa5618102dae42e7e5551986/src/contrib/torch/pt_call_tvm/tvm_class.cc#L143. On the other hand, I've never seen someone doing VM compilation in C++ (equivalent of doing `vm.compile(...)`). By `consume` I hope you meant the former use case. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
