LudovicoYIN commented on PR #19633: URL: https://github.com/apache/tvm/pull/19633#issuecomment-4570990114
Thanks for the detailed review. I updated both converters to address the semantic issues you pointed out. 1. For LSTM, the fused activation is now applied on the cell-state path before the output-gate multiply, and unsupported peephole/projection/layer-norm variants now raise OpNotImplemented instead of being ignored. 2. For SVDF, I fixed the modeling to separate rank from memory_size, produce [batch, num_units] with num_units = num_filters / rank, and update the FIFO state correctly before subsequent uses. I also strengthened the tests with LSTM NONE/TANH structural checks, an unsupported-mode rejection test, an SVDF shape/modeling test with rank > 1 and memory_size != rank, and a shared-state SVDF regression test covering two consecutive ops. -- 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]
