use src module

[frederik-uni]

having a src module did not work as expected, so I opted to use a helper crate. I didnt use all the features I implemented.

features disabled are

• shared not tested
• crt-dynamic on static builds
• export-vendor
• download releases(src code) instead of using the submodule

Compiling works on macOS, linux, windows. I also tested if vendor & using it works on macOS. I will test the vendor feature on linux tomorrow, but I expect it to work

[frederik-uni]

I may add rocm builds in the future, but only as a vendor release(the rocm patches I found are forks of the original code/break cuda[cuda is the used backend, but the logic is replaced with rocm])

[frederik-uni]

I tested it manually for macOS & linux(cuda). compiling works for all OSes. it can be assumed that the vendor feature works on every operating system & building from source too

[jkawamoto]

Thank you for this PR.

I really appreciate your contribution, but I don’t think we should merge it as-is. I’m currently working on the system feature (#108), which already makes use of the pre-compiled libraries you uploaded in Release v4.6.0 · frederik-uni/ctranslate2-src.

For example, on macOS I can pass all tests on the system branch with the following commands:

mkdir vendor
cd vendor
curl -LO https://github.com/frederik-uni/ctranslate2-src/releases/download/v4.6.0/ctranslate2-static-macos-arm64.tar.gz
tar -zxvf ctranslate2-static-macos-arm64.tar.gz

cd ..
export LIBRARY_PATH=$(pwd)/vendor/lib
cargo build --no-default-features -F hub,whisper,system -p ct2rs

And on Linux, after copying libdnnl.a and libiomp5.a into vendor/lib, the following works for me:

mkdir vendor
cd vendor
curl -LO https://github.com/frederik-uni/ctranslate2-src/releases/download/v4.6.0/ctranslate2-static-linux-x86_64.tar.gz
tar -zxvf ctranslate2-static-linux-x86_64.tar.gz
# copy libdnnl.a, libiomp5.a into lib

cd ..
export LIBRARY_PATH=$(pwd)/vendor/lib
export LD_LIBRARY_PATH=$(pwd)/vendor/dyn
export MKLROOT=/path/to/mklroot

cargo build --no-default-features -F whisper,hub,system,cuda,cudnn,dnnl,openmp-runtime-intel,cuda-dynamic-loading,mkl -p ct2rs

So instead of merging this PR, I think the better approach would be to provide pre-compiled libraries (using your ctranslate2-src-build-support) and distribute them (in https://github.com/frederik-uni/ctranslate2-src?).

Would that work for you?