Fast Neural Machine Translation in C++
Convert a model in the .npz format and normal memory layout to a mmap-able binary model which could be in normal memory layout or packed memory layout or convert a text lexical shortlist to a binary shortlist with {–shortlist,-s} option
Version: v1.12.0 65bf82f 2023-02-21 09:56:29 -0800
Usage: ./marian-conv [OPTIONS]
-h,--help Print this help message and exit
--version Print the version number and exit
-f,--from TEXT=model.npz Input model
-t,--to TEXT=model.bin Output model
--export-as TEXT=marian-bin Kind of conversion: marian-bin or
onnx-{encode,decoder-step,decoder-init,decoder-stop}
-g,--gemm-type TEXT=float32 GEMM Type to be used: float32, packed16, packed8avx2,
packed8avx512, intgemm8, intgemm8ssse3, intgemm8avx2,
intgemm8avx512, intgemm16, intgemm16sse2, intgemm16avx2,
intgemm16avx512
--add-lsh VECTOR ... Encode output matrix and optional rotation matrix into model
file. arg1: number of bits in LSH encoding, arg2: name of
output weights matrix
-V,--vocabs VECTOR ... Vocabulary file, required for ONNX export
-s,--shortlist VECTOR ... Shortlist conversion: filePath firstNum bestNum threshold
-d,--dump-shortlist TEXT=lex.bin Binary shortlist dump path
./marian-conv -f model.npz -t model.bin --gemm-type packed16