[TLE] Add device intra node barrier and change the logic of pulling the third-party tileir#739
Merged
Conversation
Galaxy1458
marked this pull request as ready for review
July 9, 2026 03:08
Galaxy1458
requested review from
i3wanna2,
menchunlei and
zhzhcookie
as code owners
July 9, 2026 03:08
zhzhcookie
reviewed
Jul 16, 2026
zhzhcookie
reviewed
Jul 16, 2026
zhzhcookie
reviewed
Jul 16, 2026
Galaxy1458
changed the title
zhzhcookie
reviewed
Jul 17, 2026
zhzhcookie
reviewed
Jul 17, 2026
zhzhcookie
reviewed
Jul 17, 2026
zhzhcookie
reviewed
Jul 17, 2026
Removed commented line related to flagtree in distributed context.
zhzhcookie
force-pushed
the
add_intra_node_barrier
branch
from
July 17, 2026 16:54
a196e55 to
3ec900f
Compare
zhzhcookie
added a commit
that referenced
this pull request
Jul 19, 2026
…he third-party tileir (#739) * init device intra node barrier * add conversions * fix conversion bug * add python api args * add python api args * fix conversion bug * add test case * adjust dir * fix test case bug * solve conflict * polish code * fix third_party/nvidia/backend/distributed.py * fix bugs * add arg for GroupKind * add flag to control tileir * add flag to control tileir * polish code * enable distribued params pass * fix bug * update * update * fix comment * polist code * fix backend bugs * fix comment * Remove commented line in jit.py Removed commented line related to flagtree in distributed context. * Update jit.py * Update compiler.py * Apply code-format changes * Initialize params list in JIT class * Apply code-format changes * [Build] Fix default_backends without tileir --------- Co-authored-by: flagtree-bot <flagtree_ai@163.com> Co-authored-by: zhzhcookie <zhengyang@baai.ac.cn>
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Summary
This PR adds support for GPU-to-GPU (device-to-device) barriers within a node in the TLE backend, enabling synchronization across multiple GPUs through the FlagCX runtime.
Also fix build tilrir backend.
Changes
Python frontend
check_and_handle_device_intra_barrier()to recognize device barriers.distributed_barrier(..., space="device", comm_ptr=...)into thetle.device_intra_barrieroperation.MLIR
Introduce a new operation:
tle.device_intra_barrierfor FlagCxOpThe operation represents an intra-node GPU-to-GPU synchronization barrier with configurable:
arrive,wait,sync)LLVM Lowering
Lower
tle.device_intra_barrierto the corresponding FlagCX runtime APIs:flagcxIntraBarrierArriveSflagcxIntraBarrierWaitSflagcxIntraBarrierSyncSThe runtime function is selected based on the requested barrier type.
Motivation
This PR introduces a dedicated IR representation for intra-node GPU-to-GPU synchronization, instead of representing it as a generic distributed barrier. It also establishes the infrastructure for future multi-GPU synchronization primitives built on the FlagCX runtime.
Example: