open-standards/.github/workflows/wai-quantum-dataflow.yml
Transaction Science 6517401f01 CI: the gate that decides whether a workflow runs was queued behind the work it gates
Every path-gated workflow begins with one small job -- `does this change touch
it`, or `which workspaces this change touches` for the standards build -- that
all of its real jobs declare in `needs`. All 55 of those gate jobs ran on
`linux-arm64`.

That is the one label the fleet cannot spare. Sampled while 267 runs were
queued, every waiting run had the same shape:

    [waiting]  linux-arm64    does this change touch it
    [blocked]  linux          member · dataflow device, simulated
    [blocked]  macos          member · dataflow device absent

237 runs, each holding a gate in the longest queue, and every real job blocked
behind it. Not only `linux` jobs: `macos`, `linux-hw`, `linux-arm64-hw` and
`macos-m3-ultra` jobs were blocked too. The measured effect was `linux-arm64`
263 deep against `linux` **zero**, with jetson-hub's 20 cores taking no job for
a full hour -- starved not for want of work but because the work could not be
unblocked. A gate waiting on the resource it is rationing is head-of-line
blocking, and it makes every other runner in the fleet idle on its behalf.

The gates move to `linux`. Nothing else changes: the diff is 55 `runs-on` lines
and no others, and every touched file still parses.

Why `linux` and not elsewhere. The gate is not free -- it compares
`BASE...HEAD`, which needs the merge base, so `fetch-depth: 0` is a full clone
rather than a ten-second script. `macos` is served by hosts on a Mac Studio
sitting at load 185; `linux` is a separate 20-core box that was measurably
empty. This moves a serialisation point to the least contended label; it does
not remove it, and the gate is still a single point of failure for its
workflow. The structural fix is an early-exit first step inside each job, which
costs one checkout per leg and is a larger change.

Follow-up, deliberately not bundled: jetson-hub runs at capacity 2 with 15 GB of
RAM, so 55 clones still serialise two at a time. Raising capacity helps gates
and hurts the heavy emulated-architecture jobs that share the label, so it wants
measuring on its own.

Co-Authored-By: Claude Opus 5 <[email protected]>
2026-11-03 15:19:13 -05:00

216 lines
11 KiB
YAML

name: wai-quantum dataflow member
# The fabric's dataflow member: the gate's fixed-point arithmetic on the first
# RISC-V core of a dataflow accelerator's compute tile, driven through the
# device's library at the level of configuration space, memory windows and
# clocks. The library here is the device's open-source simulator (Apache-2.0),
# built from source at a pinned commit, so every certificate is labelled
# simulator. Two device generations, each its own leg.
#
# Before the device runs anything, the RV32IM encoder that builds its programs
# is checked against an independent assembler: every listing, conforming and
# mutant, must assemble to the encoder's words exactly.
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true
on:
pull_request:
paths:
- 'wai/wai-quantum/**'
- 'rust-toolchain.toml'
- '.github/workflows/wai-quantum-dataflow.yml'
schedule:
- cron: '4 8 * * *'
jobs:
# Does this pull request's own diff touch the paths below? This forge matches a
# workflow's `paths:` against the push, so a merge of main into a branch fires the
# workflow whatever the pull request changes; every job here waits on this answer.
# A push to main, a schedule, or anything the script cannot determine answers
# `true` (scripts/pr-touches-paths.py).
relevant:
name: does this change touch it
runs-on: linux
outputs:
run: ${{ steps.gate.outputs.run }}
steps:
- uses: actions/checkout@v5
with:
fetch-depth: 0
- id: gate
run: |
set -euo pipefail
r=$(python3 scripts/pr-touches-paths.py .github/workflows/wai-quantum-dataflow.yml \
"${{ github.event.pull_request.base.sha }}" \
"${{ github.event.pull_request.head.sha }}")
echo "this change touches wai-quantum-dataflow: $r"
echo "run=$r" >> "$GITHUB_OUTPUT"
dataflow:
name: member · dataflow device, simulated (pci ${{ matrix.pci }})
needs: relevant
if: needs.relevant.outputs.run == 'true'
runs-on: linux
timeout-minutes: 60
strategy:
fail-fast: false
max-parallel: 1
matrix:
include:
# `vector`: the generation whose vector engine the vector programs
# drive, and whose instruction table their encoder is checked against.
- pci: '1e52:b140'
build: bh
key: dataflow-1e52-b140,dataflow-1e52-b140-vector
vector: 'true'
- pci: '1e52:401e'
build: wh
key: dataflow-1e52-401e
vector: ''
container:
image: debian@sha256:913f6706df59a68922d1dd08f78c2476560a8d367897200a6005b00e5f67c2d5
options: --memory=6g --memory-swap=6g
env:
SIM_REPO: https://github.com/tenstorrent/ttsim.git
SIM_COMMIT: 283f621e7384d86f01e4ac5a9c83b60704f77d92
WAI_QUANTUM_DATAFLOW_EXPECT: ${{ matrix.key }}
WAI_QUANTUM_DATAFLOW_SIMULATOR: device simulator 283f621e (built from source, ${{ matrix.build }})
steps:
- name: Tools
shell: bash
run: |
set -euo pipefail
apt-get update -qq
apt-get install -y -qq --no-install-recommends ca-certificates curl git build-essential python3 binutils-riscv64-linux-gnu >/dev/null
# No Node for a checkout action: the forge's archive of the exact commit,
# its embedded commit id checked.
- name: Fetch the source
shell: bash
run: |
set -euo pipefail
want="${{ github.event.pull_request.head.sha || github.sha }}"
rm -rf src src.tar.gz src.tar && mkdir src
curl -sSf -o src.tar.gz "http://100.64.224.0:3400/Transaction-Science/open-standards/archive/$want.tar.gz"
gzip -dc src.tar.gz > src.tar
got="$(git get-tar-commit-id < src.tar)"
[ "$got" = "$want" ] || { echo "::error::the archive is of $got, not $want"; exit 1; }
tar -xf src.tar -C src --strip-components=1
rm src.tar.gz src.tar
- name: The simulator, built from source at the pinned commit (cached), and Rust
shell: bash
working-directory: src/wai/wai-quantum
run: |
set -euo pipefail
cache=/ci-cache/wai-quantum-dataflow
mkdir -p "$cache"
exec 9>"$cache/.lock"
flock 9
lib="$cache/sim-$SIM_COMMIT/${{ matrix.build }}/libttsim.so"
# The table is part of the cache entry: an entry built before it was
# kept is rebuilt, not used without it.
if [ ! -f "$lib" ] || [ ! -f "$(dirname "$lib")/tensix_isa.json" ]; then
b="$cache/sim-build"
rm -rf "$b" && git init -q "$b"
git -C "$b" fetch -q --depth 1 "$SIM_REPO" "$SIM_COMMIT"
git -C "$b" checkout -q FETCH_HEAD
[ "$(git -C "$b" rev-parse HEAD)" = "$SIM_COMMIT" ] || { echo "::error::the simulator is not at the pinned commit"; exit 1; }
(cd "$b" && ./make.py "src/_out/release_${{ matrix.build }}/libttsim.so" > build.log 2>&1) || { tail -30 "$b/build.log"; exit 1; }
mkdir -p "$(dirname "$lib")"
# The instruction table too: the vector programs' encoder is checked against it.
cp "$b/data/${{ matrix.build }}/tensix_isa.json" "$(dirname "$lib")/tensix_isa.json"
cp "$b/src/_out/release_${{ matrix.build }}/libttsim.so" "$lib.tmp" && mv "$lib.tmp" "$lib"
rm -rf "$b"
fi
export RUSTUP_HOME="$cache/rustup" CARGO_HOME="$cache/cargo"
export PATH="$CARGO_HOME/bin:$PATH"
if ! command -v rustup >/dev/null; then
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --no-modify-path --profile minimal --default-toolchain none
fi
rustup toolchain install
- name: The encoder against an independent assembler
shell: bash
working-directory: src/wai/wai-quantum
run: |
set -euo pipefail
cache=/ci-cache/wai-quantum-dataflow
exec 9>"$cache/.lock"
flock 9
export RUSTUP_HOME="$cache/rustup" CARGO_HOME="$cache/cargo" CARGO_TARGET_DIR="$cache/target"
export PATH="$CARGO_HOME/bin:$PATH"
out="$RUNNER_TEMP/rv32"
cargo run --locked --release -q --example rv32_listing --features quantum_fabric_dataflow -- "$out"
riscv64-linux-gnu-as --version | head -1
n=0
for s in "$out"/*.s; do
p="${s%.s}"
riscv64-linux-gnu-as -march=rv32im -mabi=ilp32 "$s" -o "$p.o"
riscv64-linux-gnu-objcopy -O binary -j .text "$p.o" "$p.as.bin"
cmp "$p.bin" "$p.as.bin" || { echo "::error::$(basename "$p"): the encoder's words differ from the assembler's"; exit 1; }
n=$((n + 1))
done
[ "$n" = 14 ] || { echo "::error::$n of 14 programs were compared"; exit 1; }
echo "all $n programs: the encoder's words are the assembler's"
- name: The member's tests
shell: bash
working-directory: src/wai/wai-quantum
run: |
set -euo pipefail
cache=/ci-cache/wai-quantum-dataflow
exec 9>"$cache/.lock"
flock 9
export RUSTUP_HOME="$cache/rustup" CARGO_HOME="$cache/cargo" CARGO_TARGET_DIR="$cache/target"
export PATH="$CARGO_HOME/bin:$PATH"
export WAI_QUANTUM_DATAFLOW_LIB="$cache/sim-$SIM_COMMIT/${{ matrix.build }}/libttsim.so"
# The vector programs' encoder is a transcription of this generation's
# table; another generation lays some instructions out differently.
if [ -n "${{ matrix.vector }}" ]; then export WAI_QUANTUM_TENSIX_ISA="$cache/sim-$SIM_COMMIT/${{ matrix.build }}/tensix_isa.json"; fi
cargo test --locked --release -j 6 --features quantum_fabric_dataflow --lib -- quantum_fabric_dataflow:: quantum_fabric::rv32:: quantum_fabric::tensix:: quantum_fabric::vector_engine:: --nocapture 2>&1 | tee "$RUNNER_TEMP/unit.log"
grep -qE '^test result: ok\. [1-9][0-9]* passed; 0 failed' "$RUNNER_TEMP/unit.log" || { echo "::error::no unit test ran"; exit 1; }
# The vector engine's encoder against the simulator's own table (the
# test returns early without the table, so require its line). Tests
# print from parallel threads, so the line may follow another test's
# "ok" on the same line: it is matched anywhere in a line.
if [ -n "${{ matrix.vector }}" ]; then
grep -qE 'tensix layouts: [0-9]+ instructions match the table' "$RUNNER_TEMP/unit.log" || { echo "::error::the instruction table was not checked"; exit 1; }
fi
report="$RUNNER_TEMP/dataflow-report.txt"; : > "$report"
WAI_QUANTUM_DATAFLOW_REPORT="$report" cargo test --locked --release -j 6 --features quantum_fabric_dataflow --test dataflow_member -- --test-threads=1 2>&1 | tee "$RUNNER_TEMP/device.log"
grep -qE '^test result: ok\. 9 passed; 0 failed' "$RUNNER_TEMP/device.log" || { echo "::error::not every device test ran and passed"; exit 1; }
cat "$report"
grep -qx "dataflow devices: as declared ($WAI_QUANTUM_DATAFLOW_EXPECT)" "$report" || { echo "::error::the devices found are not the declared ones"; exit 1; }
# Each declared member (a generation with a vector engine has two).
for key in ${WAI_QUANTUM_DATAFLOW_EXPECT//,/ }; do
grep -q "^dataflow member: key=$key .*simulator:device simulator 283f621e" "$report" || { echo "::error::$key: the member does not say it ran on the pinned simulator"; exit 1; }
n="$(grep -c "^dataflow certified: key=$key .* parity=bit-exact " "$report")"
[ "$n" = 4 ] || { echo "::error::$key: $n of 4 certificates were earned"; exit 1; }
grep -qx "dataflow mutants: key=$key killed 3 of 3" "$report" || { echo "::error::$key: a kernel mutant survived"; exit 1; }
done
dataflow-absent:
name: member · dataflow device absent (macos)
needs: relevant
if: needs.relevant.outputs.run == 'true'
runs-on: macos
timeout-minutes: 60
env:
WAI_QUANTUM_DATAFLOW_EXPECT: none
steps:
- uses: actions/checkout@v5
# launchd's PATH omits cargo on the macOS host runner.
- name: Put cargo on PATH
shell: bash
run: echo "$HOME/.cargo/bin" >> "$GITHUB_PATH"
- name: No device library, and the member says so
shell: bash
working-directory: wai/wai-quantum
run: |
set -euo pipefail
report="$RUNNER_TEMP/dataflow-report.txt"; : > "$report"
nice -n 10 cargo test --locked --features quantum_fabric_dataflow --lib -- quantum_fabric_dataflow:: quantum_fabric::rv32:: 2>&1 | tee "$RUNNER_TEMP/unit.log"
grep -qE '^test result: ok\. [1-9][0-9]* passed; 0 failed' "$RUNNER_TEMP/unit.log" || { echo "::error::no unit test ran"; exit 1; }
WAI_QUANTUM_DATAFLOW_REPORT="$report" nice -n 10 cargo test --locked --features quantum_fabric_dataflow --test dataflow_member -- --test-threads=1 2>&1 | tee "$RUNNER_TEMP/device.log"
grep -qE '^test result: ok\. 9 passed; 0 failed' "$RUNNER_TEMP/device.log" || { echo "::error::not every device test ran and passed"; exit 1; }
grep -qx 'dataflow devices: as declared (none)' "$report" || { echo "::error::the absent case was not established"; exit 1; }