Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[RELEASE] rmm v24.02 #1440

Merged
merged 31 commits into from
Feb 12, 2024
Merged

[RELEASE] rmm v24.02 #1440

merged 31 commits into from
Feb 12, 2024

Conversation

raydouglass
Copy link
Member

❄️ Code freeze for branch-24.02 and v24.02 release

What does this mean?

Only critical/hotfix level issues should be merged into branch-24.02 until release (merging of this PR).

What is the purpose of this PR?

  • Update documentation
  • Allow testing for the new release
  • Enable a means to merge branch-24.02 into main for the release

raydouglass and others added 29 commits November 9, 2023 16:27
@raydouglass
v24.02 Updates [skip ci]
f2e4e53
@GPUtester
Merge pull request #1378 from rapidsai/branch-23.12
9900522 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1379 from rapidsai/branch-23.12
e2e6cf2 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1382 from rapidsai/branch-23.12
57faaf8 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1383 from rapidsai/branch-23.12
6da978e 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1387 from rapidsai/branch-23.12
64d2148 
Forward-merge branch-23.12 to branch-24.02
@tgravescs
Fix Arena MR to support simultaneous access by PTDS and other streams (…
da793c5 
…#1395)

Replaces #1394, this is targeted for 24.02.

fixes #1393

In Spark with the Spark Rapids accelerator using cudf 23.12 snapshot we have an application that is reading ORC files, doing some light processing and then writing ORC files. It consistently fails while doing the ORC write with:

```
terminate called after throwing an instance of 'rmm::logic_error'
  what():  RMM failure at:/home/jenkins/agent/workspace/jenkins-spark-rapids-jni_nightly-dev-594-cuda11/thirdparty/cudf/cpp/build/_deps/rmm-src/include/rmm/mr/device/arena_memory_resource.hpp:238: allocation not found
```

The underlying issue is brought about because Spark with the Rapids accelerate is using ARENA allocator with per default streams enabled.  CUDF recently added its own stream pool that is used in addition to when per default streams are used.  
It's now possible to use per thread default streams along with another pool of streams. This means that it's possible for an arena to move from a thread or stream arena back  into the global arena during a defragmentation and then move down into another arena type. For instance, thread arena -> global arena -> stream arena. If this happens and  there was an allocation from it while it was a thread arena, we now have to check to see if the allocation is part of a stream arena.

I added a test here. I was trying to make sure that all the allocations were now in stream arenas, if there is a better way to do this please let me know.

Authors:
  - Thomas Graves (https://github.com/tgravescs)

Approvers:
  - Lawrence Mitchell (https://github.com/wence-)
  - Bradley Dice (https://github.com/bdice)
  - Rong Ou (https://github.com/rongou)
  - Mark Harris (https://github.com/harrism)

URL: #1395
@harrism
Fix else-after-throw clang tidy error (#1391)
ee3817c 
Updates code for `RMM_CUDA_TRY_ALLOC` macro in `detail/error.hpp` to eliminate a clang-tidy error.

Authors:
  - Mark Harris (https://github.com/harrism)

Approvers:
  - Rong Ou (https://github.com/rongou)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #1391
@bdice
Update to fmt 10.1.1 and spdlog 1.12.0. (#1374)
d8579b7 
This PR updates to fmt 10.1.1 and spdlog 1.12.

Depends on rapidsai/rapids-cmake#473.

Closes #1356

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - Jake Awe (https://github.com/AyodeAwe)

URL: #1374
@GPUtester
Merge pull request #1398 from rapidsai/branch-23.12
e9d5007 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1399 from rapidsai/branch-23.12
423c0d4 
Forward-merge branch-23.12 to branch-24.02
@GPUtester
Merge pull request #1403 from rapidsai/branch-23.12
fe0e340 
Forward-merge branch-23.12 to branch-24.02
@vyasr
Remove RMM_BUILD_WHEELS and standardize Python builds (#1401)
cd63a08 
Some minor simplification in advance of the scikit-build-core migration to better align wheel and non-wheel Python builds.

Authors:
  - Vyas Ramasubramani (https://github.com/vyasr)

Approvers:
  - Robert Maynard (https://github.com/robertmaynard)
  - Ray Douglass (https://github.com/raydouglass)

URL: #1401
@bdice
Use latest cuda-python within CUDA major version. (#1406)
53c8043 
This PR updates cuda-python. The CUDA 11 build was locked to an outdated version (11.7.1). This matches the specifications in dependencies.yaml and also cudf recipes.

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - Ray Douglass (https://github.com/raydouglass)

URL: #1406
@vyasr
Switch to scikit-build-core (#1287)
57ed533 
Contributes to rapidsai/build-planning#2

Authors:
  - Vyas Ramasubramani (https://github.com/vyasr)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - AJ Schmidt (https://github.com/ajschmidt8)

URL: #1287
@bdice
Define python dependency range as a matrix fallback. (#1409)
362b935 
This PR moves the definition of `python>=3.9,<3.11` into the `py_version` dependency list, under the empty (fallback) matrix. This change aligns RMM's `dependencies.yaml` with other RAPIDS repositories.

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - Vyas Ramasubramani (https://github.com/vyasr)
  - Ray Douglass (https://github.com/raydouglass)

URL: #1409
@wence-
Correct signatures for torch allocator plug in (#1407)
0b931f6 
Since pytorch/pytorch#91398, the signature of the pluggable allocate and deallocate functions must accept the device id. The current version only accepts a device id for allocate, which means that when using a stream ordered allocator with devices other than device zero, we pass an invalid stream into the deallocation function. To fix this, adapt the signature to match the one pytorch expects.

Now, since we have the device available during allocation and deallocation, we would like to use that device to obtain the appropriate memory resource.

Unfortunately, since RMM's cuda_device_id does not have a nullary constructor, we can't use it in Cython without some hacky workarounds.

However, since we don't actually need to build a Python module, but rather just a single shared library that offers two extern "C" functions, let's just write our allocator hooks directly in C++.

- Closes #1405

Authors:
  - Lawrence Mitchell (https://github.com/wence-)

Approvers:
  - Mark Harris (https://github.com/harrism)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #1407
@raydouglass
Update CODEOWNERS (#1410)
ec7b164
@harrism
Update GPU support docs to drop Pascal (#1413)
cf14617 
We are dropping Pascal support in 24.02 (see rapidsai/rapids-cmake#482) 

This PR changes the way we document GPU support in RMM to explain what is tested and supported rather than what is required (since it may work on earlier hardware than we test/support).

Authors:
  - Mark Harris (https://github.com/harrism)

Approvers:
  - Bradley Dice (https://github.com/bdice)

URL: #1413
@vyasr
Remove HTML builds of librmm (#1415)
77b5500 
We no longer require separate librmm doc builds since they are incorporated into the Sphinx build now.

Authors:
  - Vyas Ramasubramani (https://github.com/vyasr)

Approvers:
  - Ray Douglass (https://github.com/raydouglass)

URL: #1415
@bdice
Update to CCCL 2.2.0. (#1404)
a4dd4f5 
This PR updates RMM to CCCL 2.2.0. Do not merge until all of RAPIDS is ready to update.

Depends on rapidsai/rapids-cmake#495.

Replaces #1247.

Authors:
  - Bradley Dice (https://github.com/bdice)
  - Vyas Ramasubramani (https://github.com/vyasr)

Approvers:
  - Mark Harris (https://github.com/harrism)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #1404
@bdice
Update dependencies.yaml to support CUDA 12.*. (#1414)
e2dc727 
This PR updates `dependencies.yaml` so that generic CUDA 12.* dependencies can be specified with a glob, like `cuda: "12.*"`. This feature requires `rapids-dependency-file-generator>=1.8.0`, so the pre-commit hook has been updated.

I have not yet added support for a specific CUDA version like 12.1 or 12.2. That can be done separately.

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - Mark Harris (https://github.com/harrism)
  - AJ Schmidt (https://github.com/ajschmidt8)

URL: #1414
@jameslamb
remove references to setup.py in docs (#1420)
9265976 
Removes remaining references to `setup.py` in documentation.

This project no longer has a `setup.py` as of its switch to `pyproject.toml` + `scikit-build-core` (see #1287, #1300).

Authors:
  - James Lamb (https://github.com/jameslamb)

Approvers:
  - Lawrence Mitchell (https://github.com/wence-)
  - Mark Harris (https://github.com/harrism)

URL: #1420
@bdice
Refactor CUDA versions in dependencies.yaml. (#1422)
5c30e87 
This is a follow-up PR to #1414. I thought some more about how to separate `cuda-version` pinnings (which control the CUDA version we use to build and test in conda) from actual CUDA Toolkit package dependencies (which we can handle according to only the major version 11/12). I discussed this PR on a call with @jameslamb in the context of upgrading to CUDA 12.2 (rapidsai/build-planning#6). This set of changes is mostly important for conda builds/tests, since `cuda-version` only controls conda. The pip wheel build/test process is unchanged, since its CUDA versions are controlled by the `shared-workflows` CI images.

Authors:
  - Bradley Dice (https://github.com/bdice)

Approvers:
  - https://github.com/jakirkham
  - Vyas Ramasubramani (https://github.com/vyasr)
  - Ray Douglass (https://github.com/raydouglass)

URL: #1422
@KyleFromNVIDIA
Remove usages of rapids-env-update (#1423)
40ce295 
Reference: rapidsai/ops#2766

Replace rapids-env-update with rapids-configure-conda-channels,
rapids-configure-sccache, and rapids-date-string.

Authors:
  - Kyle Edwards (https://github.com/KyleFromNVIDIA)

Approvers:
  - AJ Schmidt (https://github.com/ajschmidt8)

URL: #1423
@harrism
Require explicit pool size in pool_memory_resource and move some th…
64aa941 
…ings out of detail namespace (#1417)

Fixes #1416. 

 - ~Deprecates existing ctors of `pool_memory_resource` that provide optional parameter for the initial pool size.~
 - Adds new ctors that require an explicit initial pool size.
 - We don't yet deprecate anything in this PR because that would break builds of some RAPIDS libraries. We will follow up with PRs to cuDF, cuGraph and anything else needed to remove deprecated usages after this PR is merged.
 - Adds a new utility `fraction_of_available_device_memory` that calculates the specified fraction of free memory on the current CUDA device. This is now used in tests to provide an explicit pool size and can be used to produce the previous behavior of `pool_memory_resource` for consumers of the library.
 - Moves `available_device_memory` from a detail header to `cuda_device.hpp` so it is now publicly usable, along with the above utility.
 - Temporarily adds `detail::available_device_memory` as an alias of the above in order to keep cudf and cugraph building until we can update them.
 - Duplicates commonly externally used alignment functions that are currently in `rmm::detail` to the public `rmm` namespace.  The detail versions will be removed after cuDF and cuGraph are updated to not use them.

Authors:
  - Mark Harris (https://github.com/harrism)
  - Lawrence Mitchell (https://github.com/wence-)

Approvers:
  - Michael Schellenberger Costa (https://github.com/miscco)
  - Lawrence Mitchell (https://github.com/wence-)
  - Jake Hemstad (https://github.com/jrhemstad)

URL: #1417
@harrism
Deprecate detail::available_device_memory, most detail/aligned.hpp ut…
bb8fdf1 
…ilities, and optional pool_memory_resource initial size (#1424)

Follow-on to #1417, this PR deprecates the following:

 - `rmm::detail::available_device_memory` in favor of rmm::available_device_memory
 - `rmm::detail::is_aligned`, `rmm::detail::align_up` and related alignment utility functions in favor of the `rmm::` top level namespace versions.
 - The `rmm::pool_memory_resource` constructors that take an optional initial size parameter.

Should be merged after the following:
 - rapidsai/cugraph#4086
 - rapidsai/cudf#14741
 - rapidsai/raft#2088

Authors:
  - Mark Harris (https://github.com/harrism)

Approvers:
  - Michael Schellenberger Costa (https://github.com/miscco)
  - Rong Ou (https://github.com/rongou)

URL: #1424
@harrism
Add a host-pinned memory resource that can be used as upstream for `p…
12f8de3 
…ool_memory_resource`. (#1392)

Depends on #1417

Adds a new `host_pinned_memory_resource` that implements the new `cuda::mr::memory_resource` and `cuda::mr::async_memory_resource` concepts which makes it usable as an upstream MR for `rmm::mr::device_memory_resource`. 

Also tests a pool made with this new MR as the upstream.

Note that the tests explicitly set the initial and maximum pool sizes as using the defaults does not currently work. See #1388 .

Closes #618

Authors:
  - Mark Harris (https://github.com/harrism)
  - Lawrence Mitchell (https://github.com/wence-)

Approvers:
  - Michael Schellenberger Costa (https://github.com/miscco)
  - Alessandro Bellina (https://github.com/abellina)
  - Lawrence Mitchell (https://github.com/wence-)
  - Jake Hemstad (https://github.com/jrhemstad)
  - Bradley Dice (https://github.com/bdice)

URL: #1392
@harrism
Make device_memory_resource::do_get_mem_info() and supports_get_mem_i…
6c904f7 
…nfo() nonvirtual. Remove derived implementations and calls in RMM (#1430)

Closes #1426

As part of #1388, this PR contributes to deprecating and removing all `get_mem_info` functionality from memory resources.  This first PR makes these methods optional without deprecating them. 

 - Makes `rmm::mr::device_memory_resource::supports_get_mem_info()` nonvirtual (and always return false) 
 - Makes `rmm::mr::device_memory_resource::do_get_mem_info()` nonvirtual (and always return `{0, 0}`).
 - Removes all derived implementations of the above.
 - Removes all calls to the above.

Authors:
  - Mark Harris (https://github.com/harrism)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Vyas Ramasubramani (https://github.com/vyasr)

URL: #1430
@raydouglass raydouglass requested a review from a team as a code owner January 26, 2024 19:18
@raydouglass raydouglass requested review from a team as code owners January 26, 2024 19:18
@raydouglass raydouglass requested review from cwharris and vyasr January 26, 2024 19:18
@github-actions github-actions bot added CMake Python Related to RMM Python API conda cpp Pertains to C++ code ci labels Jan 26, 2024
@raydouglass raydouglass merged commit f9f1bee into main Feb 12, 2024
4 of 5 checks passed
@AyodeAwe AyodeAwe mentioned this pull request May 22, 2024
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@harrism harrism harrism approved these changes

@cwharris cwharris Awaiting requested review from cwharris cwharris is a code owner automatically assigned from rapidsai/rmm-cpp-codeowners

@vyasr vyasr Awaiting requested review from vyasr vyasr is a code owner automatically assigned from rapidsai/rmm-cpp-codeowners

Assignees
No one assigned
Labels
ci CMake conda cpp Pertains to C++ code Python Related to RMM Python API
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
9 participants
@raydouglass @harrism @GPUtester @tgravescs @bdice @vyasr @wence- @jameslamb @KyleFromNVIDIA