ci: add micro-tlx build & publish workflow - #100
Open
plotfi wants to merge 1 commit into
Open
Conversation
Adds .github/workflows/triton-utlx.yaml: builds the uTLX extension (extensions/utlx) into a pip wheel and publishes it as the distribution "micro-tlx" to PyPI on a v* tag, or to TestPyPI/PyPI on a manual run (workflow_dispatch defaults to a TestPyPI dry run). Also un-ignores the workflow from .gitignore, which would otherwise match the root-anchored 'triton-*' dependency-dir pattern (same collision the existing '!triton-ext.toml' exception handles). Signed-off-by: Puyan Lotfi <puyan@puyan.org> Assisted-by: Claude Code/claude-opus-4-8
abrown
reviewed
Jun 25, 2026
abrown
left a comment
abrown
left a comment
Collaborator
There was a problem hiding this comment.
Makes sense to me; seems like something worth iterating on. Some high-level feedback:
- if you want to publish as
micro_tlx, why not just change the name of the extension to that? - I'm not a big fan of duplicating the CI steps from the main workflow in this one (e.g., drifting out of sync); isn't there a way to run one after the other or call one from the other?
- feels like something is missing here CI-wise: don't you need credentials to publish these packages? How about a smoke test that the package actually loads from Triton?
abrown
reviewed
Jun 25, 2026
| llvm-* | ||
| triton-* | ||
| !triton-ext.toml | ||
| !.github/workflows/triton-utlx.yaml |
Collaborator
There was a problem hiding this comment.
Also, why not just rename the workflow publish-utlx.yaml (or publish-micro-tlx) to avoid this?
imperatormk
added a commit
to imperatormk/triton-ext
that referenced
this pull request
Jul 14, 2026
- MetalTGGlobalCoalesce: shrink over-provisioned global_smem to the real async-copy high-water (reclaims the upstream allocator's dead staging tail), ~20% faster 64x64 GEMM via 2 threadgroups/core. Bails on any dynamic-GEP global with no async-copy writer (int8 dot stages via store-float dynamic GEPs with no DMA tile) so it never clips live staging. - DotOpAppleMmaConversion: always emit the post-SG-load threadgroup_barrier on the multi-slot SMEM dot path. The num_stages>=3 barrier elision was unsound (write_slot == read_slot mod slots every iter -> DMA raced in-flight reads), fixing the 64x64x16:ns3 / ns3 GEMM miscompiles (triton-lang#98). - AsyncCopyConversion: partition cross-warp async_copy_2d into per-warp row bands, each warp DMAs+waits its own band. Fixes the slice_scatter cross-warp race (warp-0-only copy + per-simdgroup wait left siblings reading stale staging) while keeping the async fast path (triton-lang#100). 1x traffic, no sync fallback regression. Full inductor MPS suite: 1143 passed / 0 failed.
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
2 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.
Adds .github/workflows/triton-utlx.yaml: builds the uTLX extension (extensions/utlx) into a pip wheel and publishes it as the distribution "micro-tlx" to PyPI on a v* tag, or to TestPyPI/PyPI on a manual run (workflow_dispatch defaults to a TestPyPI dry run).
Also un-ignores the workflow from .gitignore, which would otherwise match the root-anchored 'triton-*' dependency-dir pattern (same collision the existing '!triton-ext.toml' exception handles).
Assisted-by: Claude Code/claude-opus-4-8