Change IFRT and PjRt layout API to return std::shared_ptr<const xla::PjRtLayout> instead of std::unique_ptr<xla::PjRtLayout>
#21012
+114
−115
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
![copybara-service[bot]](https://avatars.githubusercontent.com/in/44061?s=60&v=4){kind=small}
copybara-service
bot
force-pushed
the
test_711892970
branch
2 times, most recently
from
c358bd9 to
0e1e739
Compare
copybara-service
bot
changed the title
std::shared_ptr<const xla::PjRtLayout> from IFRT and PjRt instead of std::unique_ptr<xla::PjRtLayout>std::shared_ptr<const xla::PjRtLayout> instead of std::unique_ptr<xla::PjRtLayout>
copybara-service
bot
force-pushed
the
test_711892970
branch
12 times, most recently
from
c23a050 to
91b6581
Compare
…:PjRtLayout>` instead of `std::unique_ptr<xla::PjRtLayout>` The current API design that uses `std::unique_ptr<xla::PjRtLayout>` has several issues: * The API requires `xla::PjRtLayout` to be copied in some scenarios, e.g., `xla::ifrt::Array` internally stores a layout and returns its copy every time `layout()` is called. This forces implementations to break the abstraction boundary because `xla::PjRtLayout` is an abstract class and `std::unique_ptr<xla::PjRtLayout>` is not copyable. The current implementation either stores `xla::Layout` and creates `xla::PjRtLayout` every time, or downcasts `xla::PjRtLayout` to `xla::PjRtXlaLayout` to perform the copy. * `xla::Layout` is expensive to copy (`sizeof(xla::Layout)` is 248 bytes as of 2025-01-03) and copying `xla::PjRtXlaLayout` requires copying or moving `xla::Layout`. To address these two problems, this CL changes PjRt and IFRT APIs that return `xla::PjRtLayout` to instead use `std::shared_ptr<const xla::PjRtLayout>`, so that PjRt layouts can be cheaply copied. Similar patterns have been used in other places such as `xla::ifrt::Sharding` and `xla::PjRtExecutable::GetHloModules()`. Some implementations have been updated to take advantage of this change. For example, `PjRtCApiBuffer::layout()` no longer performs a layout copy and instead reuses an internally cached instance of `std::shared_ptr<const xla::PjRtLayout>`. PiperOrigin-RevId: 712645581
copybara-service
bot
force-pushed
the
test_711892970
branch
from
91b6581 to
fe28dbd
Compare
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.
Change IFRT and PjRt layout API to return
std::shared_ptr<const xla::PjRtLayout>instead ofstd::unique_ptr<xla::PjRtLayout>The current API design that uses
std::unique_ptr<xla::PjRtLayout>has several issues:xla::PjRtLayoutto be copied in some scenarios, e.g.,xla::ifrt::Arrayinternally stores a layout and returns its copy every timelayout()is called. This forces implementations to break the abstraction boundary becausexla::PjRtLayoutis an abstract class andstd::unique_ptr<xla::PjRtLayout>is not copyable. The current implementation either storesxla::Layoutand createsxla::PjRtLayoutevery time, or downcastsxla::PjRtLayouttoxla::PjRtXlaLayoutto perform the copy.xla::Layoutis expensive to copy (sizeof(xla::Layout)is 248 bytes as of 2025-01-03) and copyingxla::PjRtXlaLayoutrequires copying or movingxla::Layout.To address these two problems, this CL changes PjRt and IFRT APIs that return
xla::PjRtLayoutto instead usestd::shared_ptr<const xla::PjRtLayout>, so that PjRt layouts can be cheaply copied. Similar patterns have been used in other places such asxla::ifrt::Shardingandxla::PjRtExecutable::GetHloModules().Some implementations have been updated to take advantage of this change. For example,
PjRtCApiBuffer::layout()no longer performs a layout copy and instead reuses an internally cached instance ofstd::shared_ptr<const xla::PjRtLayout>.