CI · FBTriton

All workflows live in .github/workflows/. The GPU test workflows run on push, pull request, and a nightly schedule; nightly failures are filed as issues automatically.

GPU test workflows

Workflow	Runner	Jobs
`b200.yml`	`nvidia-dgx-b200`	`b200-tritonbench-test`, `b200-tlx-test`, `b200-gluon-test`
`h100.yml`	`linux-gcp-h100`	`h100-tritonbench-test`, `h100-tlx-test`
`mi350.yml`	`linux-fb-triton-mi350-1` (gfx950 / CDNA4)	`mi350-tritonbench-test`, `mi350-tlx-test`, `mi350-gluon-test`
`torchtlx.yml`	`nvidia-dgx-b200`, `linux-fb-triton-mi350-1`	`b200-torchtlx-test`, `mi350-torchtlx-test`

What each job class runs:

*-tritonbench-test — TritonBench performance coverage, against

meta-pytorch/tritonbench. Skips are listed in .ci/tritonbench/fbtriton_skip_tests.yaml.

*-tlx-test — TLX unit tests (python/test/unit/language/test_tlx_*.py),

the tutorial correctness suite (third_party/tlx/tutorials/testing/test_correctness.py), and runtime unit tests. The test_tlx_*.py glob covers manual TLX (tlx.async_tasks) and structurally excludes the autoWS suites, which are tested separately.

*-gluon-test — pytest python/test/gluon/ (see Gluon below).

*-torchtlx-test — test_torchtlx_templates.py and

test_torchtlx_fusions.py, against a nightly PyTorch.

Compiler tests

compiler.yml runs the lit test suite (lit-tests) on ubuntu-latest — no GPU required. This is where compiler-side AutoWS coverage lives.

ci.yml runs integration tests and pre-commit.

Pre-commit

pre-commit.yml enforces formatting; pre-commit-autofix.yml can apply fixes on demand. Run it locally before pushing:

pre-commit run --all

Nightlies and releases

Workflow	Purpose
`nightly-fbtriton.yml`	Build and publish nightly `.dev` wheels
`wheels_fb.yml`	Build fbtriton wheels (`TRITON_WHEEL_NAME=fbtriton`)
`publish_fbtriton.yml`	Publish to PyPI
`create_release.yml`	Cut a release
`release-testing.yml`	Release validation
`llvm-build.yml`	Build the pinned LLVM
`documentation.yml`	Build docs

Nightly failure handling

Scheduled runs parse their JUnit output with .github/scripts/parse_junit_failures.py and bucket failures per job. Three workflows then act on that:

Workflow	Purpose
`report-nightly-failure.yml`	File an issue for a nightly failure
`reconcile-nightly-issues.yml`	Close issues that no longer reproduce
`bisect-nightly-failure.yml`	Bisect a nightly failure (manual dispatch)

TLX-AMD CI

AMD tutorial kernels are exercised by mi350.yml on a gfx950 (MI350 / CDNA4) runner, mirroring the H100 job in h100.yml:

mi350-tlx-test — TLX unit tests (python/test/unit/language/test_tlx_*.py)

plus the tutorial correctness suite (third_party/tlx/tutorials/testing/test_correctness.py). AMD and IKBO cases run; Hopper/Blackwell and gfx1250 cases auto-skip via the arch gates.

mi350-tritonbench-test — TritonBench performance coverage (the AMD perf

scripts in TLX &rsaquo; Testing are for local runs; perf-regression tracking lives in TritonBench).

Both run on push, PR, and the nightly schedule; nightly failures are filed as issues via report-nightly-failure.yml.

Gluon

Gluon (python/triton/experimental/gluon/) is upstream-synced and not a first-class DSL here (TLX is the focus) — do Gluon feature/bug work upstream, not in this fork. But since fbtriton is a secondary Triton, we run fundamental Gluon CI so we don't silently break it.

CI: the b200-gluon-test / mi350-gluon-test jobs run pytest python/test/gluon/ (every test_*.py). It is green: ~200 passed, ~1600 skipped, 0 failed. The real signal is the compile-only, target-agnostic frontend suite (test_frontend.py, one run covers NVIDIA + AMD codegen via mock GPUTarget); the version-skewed cases below are skipped via python/test/gluon/conftest.py rather than failing the job.

Fork-side fixes (Gluon frontend itself unmodified): core.py reduce reduction_ordering compat, semantic.py dot() allow_tf32 default, a test_core.py collection fix (a bad cherry-pick left an IndentationError), and regenerated test_frontend.py goldens (upstream-synced — overwritten on next sync).

Skipped, needs upstream Gluon re-sync — TODO(gluon-ci): the GPU-execution suites (test_core, test_lowerings, test_consan, test_fpsan, one kernel in test_layout_format_view) were synced from much newer upstream (e.g. test_fpsan.py via bundle #1956) than the pinned Gluon frontend (_semantic.py ~2026-06-29). They require Gluon behavior the frontend doesn't have yet — raw pointer gl.load/gl.store inferring a *distributed* layout — so they fail wholesale with expected ... distributed_type but got block_type. The fix is to sync the Gluon frontend forward (upstream cherry-pick / re-sync), not a local patch. Also skipped: ~9 frontend per-target golden tests (single inline golden can't match every parametrized target) and the create_lds_barrier_wait pybind mismatch. conftest.py lists these; remove/trim it after the re-sync.