The TT-Forge FE is a graph compiler designed to optimize and transform computational graphs for deep learning models, enhancing their performance and efficiency.
Running llama backward pass generates select op from concat op from forward pass. Concat op is then lowered to select op in autograd pass. Next, in post-autograd stage, select op is decomposed in sequence of few ops consisting of pad_tile, narrow, sparse_matmul which are not supported in TT-MLIR. This is legacy code anyway.
Solution for select op can be to lower it to ttir.slice op instead.
rpavlovicTT
commented
3 days ago
Running llama backward pass generates select op from concat op from forward pass. Concat op is then lowered to select op in autograd pass. Next, in post-autograd stage, select op is decomposed in sequence of few ops consisting of
pad_tile,narrow,sparse_matmulwhich are not supported in TT-MLIR. This is legacy code anyway.Solution for select op can be to lower it to
ttir.sliceop instead.