Repository navigation

t27c PTX backend: ternary x int8 GEMM kernel emitted by a .t27 spec (int8 mma.sync), ptx-run glue #5863

Workflow file for this run

.github/workflows/cli-tri.yml at f3628bd

	# This repository has 35 workflows and, until this one, not a single one built
	# the `tri` crate. The consequence was exactly what you would expect: `cargo
	# build -p tri` on main failed because dlc10 embeds a bitstream with
	# `include_bytes!` that was committed to a feature branch and never to main, so
	# nobody could build the CLI from a clean checkout. An autonomous loop went on
	# committing into cli/tri/src/ the whole time.
	#
	# A crate nothing builds is a crate nobody can use, and the breakage is silent
	# because there is no signal to be red.
	name: cli-tri

	on:
	push:
	branches: [main, master]
	paths:
	- 'cli/**'
	- '.claude/skills/**'
	- 'Cargo.toml'
	- 'Cargo.lock'
	# `tri census pin --gate` pins `gates quiet` and `gates shell`, whose
	# subject is this directory. A trigger narrower than the subject cannot
	# fire on the change it exists for -- and worse, the ledger would go
	# stale on a workflow-only commit and the NEXT cli/** commit would fail,
	# blaming an author who changed nothing. Priced over the last 200 commits
	# on master: 24 touch .github/workflows/, 2 of those already touch cli/,
	# so this adds this job on 22 of 200 commits (11%).
	- '.github/workflows/**'
	- 'scripts/ci/test_a_moved_census_names_its_file.py'
	# NO `paths:` here, deliberately. A required context that does not POST is not
	# "passed" -- it is pending forever, so a paths-filtered workflow cannot be
	# required without hanging every PR that skips it. Measured over the last 40
	# merged PRs: 15 of them (38%) touched none of the paths this filter listed.
	# The `push:` filter above is kept: a master baseline only needs taking when
	# the subject moves.
	#
	# Cost of running on every PR, measured rather than estimated: +38% of runs on
	# a job whose median is 2.9 minutes, green 58 of its last 60 (43 of 45 on
	# pull_request). A gate that reddens rarely is one that can be required.
	pull_request:
	workflow_dispatch:

	# A newer push or PR edit makes this run's verdict stale: cancel it (queue jam 2026-10-09).
	concurrency:
	group: cli-tri-${{ github.ref }}
	cancel-in-progress: true

	jobs:
	build:
	# A UNIQUE context name. GitHub matches a required check by CONTEXT, and
	# `build-vivado-image.yml` and `coq-kernel.yml` both emit `build` too -- so
	# "require build" names three workflows and can be satisfied by the wrong
	# one, or hang on a job that does not run. First of two things that had to
	# be true before this gate could be required at all.
	name: cli-tri
	runs-on: ubuntu-latest
	timeout-minutes: 25
	steps:
	# chips/phi is a gitlink (160000) to gHashTag/tt-trinity-phi, and
	# checkout's default is submodules: false -- so the directory the last
	# step points `tri rtl check` at was EMPTY on the runner, and the CLI died
	# with a bare "No such file or directory (os error 2)". The binary was
	# fine; it had nothing to read. That step had never executed once in this
	# workflow's history, because the two causes before it (#2303 yosys
	# ordering, #2305 dry_run_sweep_ok) aborted the job first.
	#
	# All three chips/* submodules are public, so the default GITHUB_TOKEN --
	# which is scoped to this repo alone -- can clone them; ~10.2 MB total.
	# None of the three has a nested .gitmodules, so `true` is the minimal
	# setting that works and `recursive` would buy nothing.
	- uses: actions/checkout@v4
	with:
	submodules: true
	- uses: dtolnay/rust-toolchain@stable

	# From a clean checkout, with nothing copied in by hand. That is the
	# condition that was broken and the only one worth asserting.
	- name: cargo build -p tri --all-targets
	# Tests are targets too. The plain build stayed green twice while the
	# test build was broken on master (#2227, #2236) -- every PR inherited
	# a break no gate reported. --all-targets closes that gap.
	run: cargo build -p tri --all-targets

	# #5482. `tri hooks pre-push` runs the NOW gate's script in a
	# contributor's own clone (on request only since #5935; `tri gates
	# preview` did too until then), and until #5482 every pass there
	# rewrote .git/config and appended to .gitattributes. This job has the
	# binary, so the `tri` caller is asked here, in a scratch repository;
	# untrusted-input-gate.yml runs the script's own arms without one.
	# Placed before the tests so a red test does not skip it.
	- name: The NOW gate's tri caller writes nothing to the clone
	run: python3 scripts/ci/test_now_gate_writes_nothing.py --tri ./target/debug/tri

	# yosys has to be on PATH BEFORE the tests, not after them. This install
	# used to sit inside the last step's run block, which is ordered after
	# `cargo test -p tri`, and ubuntu-latest ships no yosys. So
	# fpga::tests::test_smoke_gate_json_synthetic_verify_lean -- which calls
	# the real smoke_gate(), whose `passed` flag ANDs in `yosys_ok` -- hit
	# "[smoke-gate] SKIP: yosys not on PATH" and failed on every single run.
	# 11 runs on master, 11 failures, no green in this workflow's history:
	# the job could not pass by construction, and the job never even reached
	# the install because a failed step aborts the rest.
	#
	# The test is right to demand the binary. A smoke gate that "passes"
	# while skipping its only synthesis phase asserts nothing -- the same
	# absence-is-not-a-value defect closed in #2285 and #2287. Fix the
	# ordering, not the assertion.
	- name: install yosys
	run: sudo apt-get update -qq && sudo apt-get install -y -qq yosys

	# zig before the tests, for the same reason as yosys above: `tri mutate
	# spec` compiles a mutant with zig and runs its tests, and its own tests
	# (#7148) do both on small lowered specs. Without zig they would fail, and
	# rightly: a hang or a compile error told apart from a kill without
	# running zig would assert nothing. Same pin as bootstrap-tests.yml.
	- name: Install zig
	env:
	ZIG_VERSION: "0.16.0"
	run: \|
	set -euo pipefail
	curl -fsSL "https://ziglang.org/download/${ZIG_VERSION}/zig-x86_64-linux-${ZIG_VERSION}.tar.xz" -o /tmp/zig.tar.xz
	mkdir -p "$RUNNER_TEMP/zig"
	tar -xJf /tmp/zig.tar.xz -C "$RUNNER_TEMP/zig" --strip-components=1
	echo "$RUNNER_TEMP/zig" >> "$GITHUB_PATH"
	"$RUNNER_TEMP/zig/zig" version

	# --nocapture --test-threads=1, and the log tee'd out, so the ratchet
	# below can measure something `cargo test` otherwise throws away.
	#
	# Three smoke-gate tests need a Lean toolchain that no workflow installs
	# (#2370). They do not use #[ignore] -- they println! a reason and return.
	# So libtest counts them as PASSED and the summary line reads
	# "0 ignored", and `cargo test` CAPTURES the stdout of a passing test, so
	# the reason never reaches the log at all. Without --nocapture there is
	# nothing to measure. Without --test-threads=1 the reason cannot be
	# attributed to a test, because libtest lands it on the "test NAME ... "
	# line it has already opened and parallel workers interleave those.
	#
	# set -o pipefail keeps tee from swallowing a real test failure.
	- name: cargo test -p tri
	run: \|
	set -o pipefail
	cargo test -p tri -- --nocapture --test-threads=1 2>&1 \| tee "$RUNNER_TEMP/tri-test.log"

	# A green run of this job is consistent with those three never having
	# executed. This step puts the count and the reasons in the job summary so
	# a reader of a green run cannot mistake it for coverage, and ratchets the
	# set: it FAILS if the set grows. A fourth test starting to skip silently
	# is the regression nobody would otherwise see.
	#
	# It deliberately does not fail on today's three. Turning existing skips
	# red is the owner's policy decision and a skip is not a failure; #2370
	# stays open for the real fix, which is installing the toolchain.
	#
	# The baseline is emitted by this same script (--emit-baseline) from a real
	# test log. It is never hand-written: a hand-typed "3" would restate what
	# we already know and would not survive a rename.
	- name: runtime skip ratchet
	if: always()
	run: python3 cli/tri/skipwatch.py "$RUNNER_TEMP/tri-test.log" cli/tri/skip-baseline.txt

	# `tri rtl check` reports numbers, and a binary that runs but reports
	# nothing would pass the two steps above. yosys is already on PATH from
	# the install step above, so the command is exercised against a real
	# design rather than assumed.
	# `tri skill check` finds two sections sharing a number, prints PROBLEMS
	# and exits 1 -- and until now NOTHING called it. `grep -rn "skill check"`
	# across .github/workflows/, scripts/ and tools/ returned zero.
	#
	# It was not hypothetical: two sections were both numbered 504 on master,
	# landed 26 minutes apart by two sessions, each appending above the highest
	# number it had read. The checker would have caught it on the second PR.
	#
	# This job already builds the binary, so the marginal cost is one command.
	# Its `paths:` grows to .claude/skills/** in the same commit, because a
	# gate whose trigger is narrower than its subject cannot fire on the change
	# it exists for.
	- name: No two skill sections share a number
	run: ./target/debug/tri skill check

	# A census that moves silently is a number nobody re-read. Measured over
	# the 39 most recent transitions on master with one fixed instrument:
	# 8 commits moved a census and only 4 said so -- and one of the silent
	# four had made 45 of 50 red workflows in a sibling repository invisible
	# (#3176). The A/B that finds it is two commands, and nothing ran them.
	#
	# This is cheap for a reason that is structural rather than hopeful: each
	# pinned census's POPULATION IS A DIRECTORY, so it cannot move unless a
	# file in that directory changes. All 8 of the 8 observed moves had edited
	# that census's own subject; not one was a side effect. So the re-bless
	# falls only on commits already working in that area.
	#
	# Deliberately excluded: `dead` and `unmeasured` read the GitHub API, so
	# their answer moves when the world moves and pinning them would redden
	# this gate on somebody else's push. A test refuses their addition.
	- name: No census moved without saying so
	run: ./target/debug/tri census pin --gate

	# The gate above prints the FIRST line that differs and stops, which says a
	# census moved and not what moved it. On #5787 that first line was `files
	# read 22 -> 23` and the cause sat further down (`0 do not -> 1 do not`: a
	# helper spelling "lower bound" in lowercase), and the PR read "open" for six
	# ticks while this job was red on it. `tri census explain` lists every moved
	# line and names the changed file behind each by putting that file back
	# alone in a scratch copy. This step proves it names the mover and not the
	# bystander, writes nothing, and leaves no copy behind (#5825).
	- name: A moved census names the file that moved it
	run: python3 scripts/ci/test_a_moved_census_names_its_file.py --tri ./target/debug/tri

	# `cargo test` cannot report this: a test that lost its `#[test]` and one
	# that gained a second CANCEL in the total, so the suite count is
	# unchanged and green. Commit f7c1ff5 shipped exactly that pair -- one
	# test stopped running, another was registered twice -- and the pass that
	# wrote it verified "11 attributes against 11 functions" and read a match.
	- name: Every test runs, and runs once
	run: ./target/debug/tri gates tests --gate

	- name: the CLI actually produces a report
	run: \|
	set -uo pipefail
	./target/debug/tri rtl check chips/phi --json > /tmp/r.json 2>/tmp/r.err \|\| true
	cat /tmp/r.err \| head -5
	N=$(python3 -c "import json;print(len(json.load(open('/tmp/r.json'))['checks']))" 2>/dev/null \|\| echo 0)
	echo "verdict lines: $N"
	if [ "${N:-0}" -lt 5 ]; then
	echo "::error::tri rtl check emitted $N verdicts; five checks should each emit one"
	exit 1
	fi

Navigation Menu

Repository navigation

t27c PTX backend: ternary x int8 GEMM kernel emitted by a .t27 spec (int8 mma.sync), ptx-run glue #5863

Workflow file

t27c PTX backend: ternary x int8 GEMM kernel emitted by a .t27 spec (int8 mma.sync), ptx-run glue #5863

Uh oh!

Workflow file for this run