Merge OpenAI Triton commit e9db186
#2793
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 improves a warm-cache macOS build from ~25 mins to 2 mins.
Follow up to #5202 It's currently failing with the error ``` du: /Users/runner/.triton/**: No such file or directory Error: Process completed with exit code 1. ``` which happens because even though the `.triton` directory exists, it is empty. This instead uses du on `.triton` with a depth of 1.
#### Commits in this PR
1. [CI] Fix cache not saving
Re-using the output of the cache restore step was recommended by the
`actons/cache` docs, but it doesn't work here because we actually start
from a clean cache when we run save so there is no output available to
read.
The annoyances of testing in the PR but main being a different
environment.
2. Bump macOS timeout
We also exercise this in scale_dot, where we enable support for warps of arbitrary shape (before we just allowed `[num_warps, 1]`). With this infra in place, it should be rather easy to move from the legacy layouts to using LLs to represent all of our layouts. Something I'm concerned about is the amount of recomputation that happens when calling methods like `getSizePerThread` and the like, where we keep recomputing the result. There might be an optimisation opportunity here where we cache the result of all these functions. We choose the IR representation of an LL via its canonical form + a `repOrder` for several reasons: - It's generally more compact - It's easier to CSE, so it's easier to see when two layouts are in fact the same. - A technical reason: the `toLinearLayout` function returns a tensor with dimensions `dim0, ..., dim<rank-1>`, in other words, it "forgets" the repetition order. Without the repetition order, we cannot recover the tile size of the argument. In particular, we cannot recover `getSizePerThread`. There is an argument to be made about whether `getSizePerThread` is useful on its own, or whether it is `getElemsPerThread` the real useful abstraction here, but for now, we keep both for BC.
Currently you can manually call a workflow dispatch, but it won't actually run the tests because the variable enable_integration isn't set.
…ra-kernel perf tooling (#5119) This PR introduces the `Proton Dialect` to enable intra kernel profiling and tooling for Triton. As a third-party dialect, it serves as the building blocks to create 3rd-party perf tools (e.g., profilers, analysis, modeling) for Triton compiler developers in a compiler-centric way, such as an intra-kernel latency profiler to understand software pipelining, warp specialization, and CTA fine-grained orchestration (e.g., cuda core, tensor core, TMA). Future developments would integrate this dialect with the existing Proton backend profiling infrastructure to make it a powerful and general perf tool utility. As a first step, this PR adds some basic boilerplate code and mechanics, and the `proton.record` op for the `Proton Dialect`. --------- Co-authored-by: Yuanwei Fang <[email protected]> Co-authored-by: Keren Zhou <[email protected]>
pbchekin
approved these changes
Nov 21, 2024
whitneywhtsang
changed the title
d5ba6ace9db186
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.
This PR change the Triton base from 251ec88 to e9db186 (Nov 21).
Pass rate: 93.23%->93.24%
Please do not squash and merge this PR.