Cagra memory optimizations #1790
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This just gets copied over to the output, and requires multiple copies of the output graph on host memory.
was still making a copy from device->host inside serialize_mdspan, and with the include_dataset changes this branch won't even be called
benfred
added
non-breaking
cjnolet
reviewed
Aug 31, 2023
| template <typename T, typename IdxT> | ||
| class device_matrix_view_from_host { | ||
| public: | ||
| device_matrix_view_from_host(raft::resources const& res, host_matrix_view<T, IdxT> host_view) |
There was a problem hiding this comment.
cc @wphicks this pattern seems a lot like the mdbuffer to me. The goal here is to make a device_mdspan when the pointer can be accessed from device or copy memory to device when it can't.
cjnolet
reviewed
Sep 5, 2023
Comment on lines
155
to
166
| /** | ||
| * Utility to sync memory from a host_matrix_view to a device_matrix_view | ||
| * | ||
| * In certain situations (UVM/HMM/ATS) host memory might be directly accessible on the | ||
| * device, and no extra allocations need to be performed. This class checks | ||
| * if the host_matrix_view is already accessible on the device, and only creates device | ||
| * memory and copies over if necessary. In memory limited situations this is preferable | ||
| * to having both a host and device copy | ||
| */ | ||
| template <typename T, typename IdxT> | ||
| class device_matrix_view_from_host { | ||
| public: |
There was a problem hiding this comment.
I'm definitely okay keeping this as an internal utility for now. Could you add a todo to the docs here (and for the host->device conversion function) to use mdbuffer for this once it's available?
There was a problem hiding this comment.
added a TODO here 6552c66
cjnolet
approved these changes
Sep 8, 2023
|
/merge |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
When trying to build a CAGRA index with 500M embeddings, we were running out of memory - even when using managed memory.
This PR contains some changes to reduce the memory usage:
optimizecall, and just use host memory passed in by caller