alvations

Specifically , maybe throw some errors when: ``` -- Looking for cblas_sgemm -- Looking for cblas_sgemm - not found -- CBLAS library not found. Please specify library location ```

Starring at https://github.com/marian-nmt/marian-dev/blob/master/CMakeLists.txt#L437 =)

Maybe a simple throwing of `FATAL_ERROR` at the MKL/CBLAS checks will fix this quickly. There seems to be no exact fix here. ``` if(CBLAS_FOUND) include_directories(${BLAS_INCLUDE_DIR} ${CBLAS_INCLUDE_DIR}) set(EXT_LIBS ${EXT_LIBS} ${BLAS_LIBRARIES} ${CBLAS_LIBRARIES})...

For marian devs, Not sure what's the resolution in code though, throw a warning to user? Updating the `dim-vocabs` to `max(sp_vocab_size, 8000)`? ----- For marian users, to avoid this feature/bug...

Maybe I should check whether `--max-length-crop` vs `--max-length 100000` makes a big difference. Regardlessly, I guess omitting and returning `\n` is okay but stopping the decoding on the line is...

BTW, what's the default `SIZE_MAX`?

# Data https://drive.google.com/file/d/1hr-RzBz-5zCMhwbi4ogzPulVB1ZLCDUI/view?usp=sharing # Marian version `v1.11.0 f00d0621 2022-02-08 08:39:24 -0800` # Command ``` #!/bin/bash SRC=fr # en TRG=xx # ja RANDSEED=42 # 42 ELAYERS=6 # 6 DLAYERS=6 HEADS=8 #...

This is interesting, - Broke batching: `--max-length 5000 --valid-max-length 5000` - Broke batching: `--max-length 2000 --valid-max-length 2000` - Seems to work: `--max-length 1000 --valid-max-length 1000`

Just to check-in on this, this PR is obsolete as of v1.9.0 (that would materialize soon), is that right? Is the `--fp16` option using something else or @kpu implementation of...

+1 for consistency checks, esp. when annotating text for AMR, it's very much free form =)