What's the difference with https://github.com/bazel-contrib/rules_cuda?

[fuhailin]

I see https://github.com/bazel-contrib/rules_cuda also provide the CUDA compiling tool in Bazel

[ybaturina]

According to the documentation, rules_cuda provides Starlark wrappers around local CUDA tools and sources.

rules_ml_toolchain provides hermetic CUDA toolkit support which doesn't require the user to install CUDA locally. Instead, the user can just specify CUDA/CUDNN/NCCL/NVSHMEM versions, and rules_ml_toolchain repositories will download all of that directly in Bazel cache. We also have LOCAL_CUDA_PATH support for some scenarios (see here).

Apart from that, rules_ml_toolchain offers hermetic C++ support for a number of platforms, e.g. the user doesn't have to install Clang locally - it is managed by rules_ml_toolchain repository rules.

We are planning to add more useful rules and wrappers here, for example, those that add more features to rules_python rules.

[cerisier]

From using both rulesets extensively,

rules_ml_toolchain is developed by Google
rules_cuda is developed by NVIDIA

Both provide hermetic CUDA toolkit from NVIDIA redist.

The main different lies in the implementation:

1. rules_cuda uses a cc toolchain (cuda_library is a rule that invokes cc_common.compile).
2. rules_ml_toolchain is the cc toolchain (cuda_library is a macro over cc_library, provides custom cc_toolchain_config)

Both can achieve total hermeticity:

1. Registering a hermetic CC toolchain will make rules_cuda fully hermetic.
2. rules_ml_toolchain provides that by default (by choosing for you).

Both have advantages and inconveniences.

[medaminezghal]

@ybaturina Could rules_cuda fix this problem?