Package build & deployment procedure

GPU support

CUDA support is available for Linux and Windows target platforms.

Inference on CUDA GPUs is currently supported only for models with exclusively numerical features.

It is disabled by default and can be enabled by adding --have-cuda to scripts' flags.

CUDA architectures to generate device code for are specified using CMAKE_CUDA_ARCHITECTURES variable, although the default value is non-standard, specified in cuda.cmake. The default value is intended to provide broad GPU compatibility and supported only when building with CUDA 11.8. The most convenient way to override the default value is to use CUDAARCHS environment variable.

Building scripts

Scripts responsible for building and preparing the package for publishing are located in /build_scripts directory.

These scripts are built once on the first execution of any script. To compile build scripts explicitly, run

npm run compile_build_scripts

Scripts

npm run install - main script for installing the package. Behaves differently if called from catboost directory or from the standalone package distribution:
- If called from catboost repository, then builds the local version of catboostmodel library from source, compiles and links package against it. After that the package can be added locally via
```
  npm install $PATH_TO_CATBOOST_REPO/catboost/node-package
```
- If called outside of the repository, then the package is considered to be installed from npm and it will be built and linked against the binaries specified in config.json file. This file is created during package preparation for publishing, see below.
npm run build - build package locally in the catboost repository, link against library built from source.
npm run ci - [requires Linux with docker installed] a single script for CI which runs the procedure described in "Release procedure" section. Add --have-cuda flag to enable tests with CUDA evaluation.
npm run compile - compile Typescript source files only.
npm run test - run local unit tests. Add --have-cuda flag to enable tests with CUDA evaluation.
npm run compile_build_scripts - compile build scripts from Typescript sources.
npm run clean - delete local artifacts.
npm run package_prepublish <version> - prepare the package for publishing, does the following:
1. Copies C headers required for building package on the client's side from repository to a ./inc directory.
2. Compiles Typescript source files.
3. Fetches the binaries from the github.com releases for a given version (for ex. v0.25.1), prepopulates config.json file. This config file includes link to the binaries and sha256 checksums that will be verified on the client side. The binaries have to be downloaded during the package installation as they are too big to be part of the distributed package.

Release procedure

Verify package locally (optional)
1. Checkout the released branch.
2. Run
```
npm run build -- --have-cuda
```
  to build package from source.
3. Run
```
npm run test -- --have-cuda
```
  to verify that the tests are passing.
Run
```
npm run package_prepublish <version>
```
Check generated config.json file, verify that the links and file checksums are correct.
Update package version in package.json.
Run integration deployment test [requires Linux with docker installed]:
```
npm run e2e
```
Publish package with
```
npm publish
```

CI

Requires Linux with Internet access and docker (with permissions to tag images and run containers) installed.

For setting up continuous integration the following had to be done:

Checkout the source code (at the release tag commit).

From catboost/catboost/node-package subdirectory, execute ci script and check that it is executed correctly:

npm run ci -- <catboost-release version> <catboost-node-package version> --have-cuda [build_native arguments]

See build_native documentation about possible arguments. Don't specify already defined --target or --build-root-dir arguments.

Assuming that npm registry credentials are set up on the host, publish the package via
```
npm publish
```