tqchen opened a new pull request, #20395: URL: https://github.com/apache/tvm/pull/20395
Preserve kernel launch metadata through `tirx.call_ffi_kernel`. Ordinary host targets lower these calls through packed functions, while `cuda_host` emits direct CUDA launches with explicit diagnostics for unsupported launch modes. Add `tvm.backend.cuda.bundle_cuda_host(mod)` to concatenate retained CUDA device source before its host wrappers into one translation unit. Existing CUDA device code generation stays unchanged; source generation, compilation and loading remain separate steps. CUDA-host wrappers use tvm-ffi and CUDA libraries for streams, scoped device selection, tensor-map encoding and error handling. Tensor-map descriptors use aligned host storage and the device's by-value launch ABI. Unsupported TVM runtime services are diagnosed during code generation. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
