Bradley Dice

See also #627 and https://github.com/rapidsai/cucim/issues/305#issuecomment-1670273320.

I believe the failures on CUDA 11 are indicative of a problem in include paths. The failures look like this: ``` │ │ sccache $BUILD_PREFIX/bin/nvcc -forward-unknown-to-host-compiler -DCUB_DISABLE_NAMESPACE_MAGIC -DCUB_IGNORE_NAMESPACE_MAGIC_ERROR -DLIBCUDACXX_ENABLE_EXPERIMENTAL_MEMORY_RESOURCE -DRMM_LOG_ACTIVE_LEVEL=RAPIDS_LOGGER_LOG_LEVEL_INFO...

@vyasr Yes, I think scoping to just NVTX is a good start. The problem is that the failure on CUDA 11 is actually just an indication that we have the...

As I think more about this, I see some tension. We would prefer a hard pinning on `nvtx-c` so that we match exactly what is in `rapids-cmake`'s `versions.json`. However, with...

An unpinned CCCL would make building and testing too difficult. We need CCCL pinned at build time. But it’s hard to have pinned dependencies for packages that are mostly header-only....

I think that’s worth exploring. There are sizable parts that are headers because they’re templated, so I’d want to explore the device_uvector and other templated types’ usage of NVTX first.

We already have a `cccl` conda package. What relationship will exist between `cccl` and `cuda-cooperative`? Does one depend on another at build time/run time? Do we ship `cuda-cooperative` features as...

/ok to test

See here: https://github.com/rapidsai/pre-commit-hooks/issues/2 I haven't had bandwidth to do this yet, but it's still on my radar and will probably be one of my projects for 23.08 or so.

Typically a `BinningMemoryResource` is used for making small allocations and large allocations with different MRs. For example, some applications might want to make small allocations in pinned memory that can...