Mark Harris

Part of #1596

[BUG] Segfault in PyTorch after initializing RMM pytorch allocator with managed pool.

2

**Describe the bug** Originally posted by @alexbarghi-nv: Hi rmm team, I'm running into issues on Grace Hopper (more recently reproduced on A100) with the rmm pytorch allocator. When using a...

[EXPERIMENT] Use resource_ref for get/set_per_device_resource

2

## Description Components of this PR. 1. C++: Use `rmm::device_async_resource_ref` in `get/set_per_device_resource()`. (per_device_resource.hpp) 2. C++: Use the new API above in tests. 3. C++: Modify resource adaptor classes to store...

Is your feature request related to a problem? Please describe. I'm about to undertake yet another change to every function that takes an MR parameter. To prevent doing this too...

**Is your feature request related to a problem? Please describe.** To reduce thrash due to RMM refactoring, RAFT is adding wrappers around RMM `rmm::get_current_device_resource()`, `rmm::set_current_device_resource()` and aliases for RMM `resource_ref`...

**Describe the bug** cuVS .clang-format does not include the header ordering code that was added for all other RAPIDS C++ libraries. E.g. from RAFT: ``` IncludeCategories: - Regex: '^"' #...

[BUG] Memory Leak in SVM

3

**Describe the bug** There are allocations with no matching deallocation in svm/sparse_util.cuh. https://github.com/rapidsai/cuml/blob/973a65ff703a14d9dd8f355af373e2053444b27a/cpp/src/svm/sparse_util.cuh#L766-L778 **Steps/Code to reproduce bug** Inspect the linked code. **Expected behavior** See #5852. There should be no raw...

### Is this a new feature, an improvement, or a change to existing functionality? New Feature ### Please provide a clear description of problem this feature solves To reduce thrash...

### Is this a new feature, an improvement, or a change to existing functionality? New Feature ### How would you describe the priority of this feature request High ### Please...

**Is your feature request related to a problem? Please describe.** We can remove the optimizations around CUDA_STATIC_RUNTIME and instead see if the function is already in the process space so...