Uplift Ttsim CI

Uplift TTSim CI

Follow the stages below in order. Do not move the regression sweep ahead of unskip triage, and do not stop after unskip triage.

Definition of Done

An uplift is complete only when all of these are true:

1. The release tag and both WH/BH assets are verified.
2. Every applicable simulator-skip node is inventoried before testing.
3. Every executable candidate is run in its own pytest process on each applicable architecture, and the skip edits are resolved.
4. The temporary noop edit is restored.
5. Every workflow-listed file is run, one file per pytest process, on both WH and BH after the skip edits.
6. Changed skip files are rerun on both architectures, the workflow is syntax-checked, and the final diff contains no triage-only edits.

If any stage is incomplete, report the coverage gap; do not call the uplift done.

Ground Rules

• .github/workflows/call-test-ttsim.yml is the source of truth for the test list, matrix, descriptors, and assets. Re-read it at the start; never reuse a stale test list copied into this skill.
• Cover wormhole_b0 and blackhole. Ignore Quasar/QSR unless requested.
• A TTSim error may terminate pytest with no summary, sometimes with exit code 1 rather than a signal-derived code. During unskip triage, one exact node per pytest process is mandatory. A whole-file run with skips disabled is not valid unskip evidence.
• Serialise candidate nodes within one architecture. WH and BH may run in parallel because they have separate simulator homes, libraries, descriptors, artifact paths, and logs.
• Use /opt/ttmlir-toolchain/venv/bin/python -m pytest; the local pytest wrapper may hide useful output.
• Before local commands, activate without set -u; activation expects some variables to be unset:

unset BUILD_DIR
source env/activate
export BUILD_DIR="$PWD/build"

Stage 1: Freeze the Release and CI Scope

1. Find the latest public release of tenstorrent/ttsim.
2. Verify that it contains both libttsim_wh.so and libttsim_bh.so.
3. Record the release tag, publication time, asset names, current tt-mlir SHA, and current tt-metal SHA.
4. Record the initial git status --short; preserve unrelated user changes.
5. Extract the exact .py arguments from the Run golden pytest on TTSim step into a shell TEST_FILES array and print it for review:

mapfile -t TEST_FILES < <(
  awk '/^[[:space:]]+pytest /,/--sys-desc/ {
    if ($1 ~ /\.py$/) print $1
  }' .github/workflows/call-test-ttsim.yml
)
((${#TEST_FILES[@]} > 0)) || {
  echo "No TTSim workflow tests found"
  exit 1
}
printf '%s\n' "${TEST_FILES[@]}"

6. Update only the workflow's ttsim-version default unless the current release requires a justified matrix/setup change. Keep this matrix:

- arch: wormhole_b0
  soc_desc: wormhole_b0_80_arch.yaml
  ttsim_asset: libttsim_wh.so
- arch: blackhole
  soc_desc: blackhole_140_arch.yaml
  ttsim_asset: libttsim_bh.so

Stage 2: Create Independent WH/BH Environments

Use a versioned root and separate result directories:

export TTSIM_VERSION=vX.Y.Z
export TTSIM_ROOT="${TMPDIR:-/tmp}/ttmlir-ttsim/$TTSIM_VERSION"
mkdir -p \
  "$TTSIM_ROOT/wormhole_b0" \
  "$TTSIM_ROOT/blackhole" \
  "$TTSIM_ROOT/results/wormhole_b0" \
  "$TTSIM_ROOT/results/blackhole"

curl -L --fail --retry 3 \
  -o "$TTSIM_ROOT/wormhole_b0/libttsim_wh.so" \
  "https://github.com/tenstorrent/ttsim/releases/download/$TTSIM_VERSION/libttsim_wh.so"
curl -L --fail --retry 3 \
  -o "$TTSIM_ROOT/blackhole/libttsim_bh.so" \
  "https://github.com/tenstorrent/ttsim/releases/download/$TTSIM_VERSION/libttsim_bh.so"

cp third_party/tt-metal/src/tt-metal/tt_metal/soc_descriptors/wormhole_b0_80_arch.yaml \
  "$TTSIM_ROOT/wormhole_b0/soc_descriptor.yaml"
cp third_party/tt-metal/src/tt-metal/tt_metal/soc_descriptors/blackhole_140_arch.yaml \
  "$TTSIM_ROOT/blackhole/soc_descriptor.yaml"

Override TTSIM_ROOT with a durable path when logs must survive a reboot, but keep generated binaries, artifacts, and reports outside the git working tree.

Set common variables once:

export TT_MLIR_HOME="$PWD"
export TT_METAL_HOME="$PWD/third_party/tt-metal/src/tt-metal"
export TT_METAL_SLOW_DISPATCH_MODE=1
export TT_METAL_DISABLE_SFPLOADMACRO=1
export TT_METAL_INSPECTOR=0
export TT_METAL_INSPECTOR_RPC=0
export LD_LIBRARY_PATH="$PWD/build/lib:${TTMLIR_TOOLCHAIN_DIR}/lib:${LD_LIBRARY_PATH:-}"

Build the current checkout before testing; results from stale tt-mlir or tt-metal binaries are invalid:

cmake --build "$BUILD_DIR"

Generate and retain one system descriptor per architecture. These commands may run in parallel:

export TT_METAL_SIMULATOR_HOME="$TTSIM_ROOT/wormhole_b0"
export TT_METAL_SIMULATOR="$TT_METAL_SIMULATOR_HOME/libttsim_wh.so"
ttrt query --save-artifacts \
  --artifact-dir "$TT_METAL_SIMULATOR_HOME/ttrt-artifacts" --quiet

export TT_METAL_SIMULATOR_HOME="$TTSIM_ROOT/blackhole"
export TT_METAL_SIMULATOR="$TT_METAL_SIMULATOR_HOME/libttsim_bh.so"
ttrt query --save-artifacts \
  --artifact-dir "$TT_METAL_SIMULATOR_HOME/ttrt-artifacts" --quiet

Do not share TT_METAL_SIMULATOR_HOME, --sys-desc, --path, or result logs between architectures.

Stage 3: Plan the Unskip Matrix

Complete the candidate plan before executing candidates.

1. Search only the workflow-listed files for sim skip sites. Record file, line, mark, parameter scope, and intended architecture in sim-mark-sites.txt.

rg -n '\bsim\b' "${TEST_FILES[@]}" > sim-mark-sites.txt

2. With the normal environment detection ("sim"), collect exact skipped node IDs on WH and BH without executing tests:

COLUMNS=10000 /opt/ttmlir-toolchain/venv/bin/python -m pytest \
  --setup-plan -vv --color=no "${TEST_FILES[@]}" \
  --sys-desc "$TT_METAL_SIMULATOR_HOME/ttrt-artifacts/system_desc.ttsys" \
  2>&1 |
  awk '/ SKIPPED/ {sub(/[[:space:]]+SKIPPED.*/, ""); print}' |
  LC_ALL=C sort -u

Save this once per architecture as all-skipped-before-noop.txt.

3. Temporarily change _get_current_environment() in test/python/golden/conftest.py to return "noop" when TT_METAL_SIMULATOR is set. Do not add noop to ALL_ENVIRONMENTS. Simulator marks no longer match, while unconditional skips, non-simulator skips, and only_config behavior remain active.
4. Repeat the setup-plan collection into still-skipped-with-noop.txt.
5. Both files must be sorted. Produce the executable candidate list with:

comm -23 \
  all-skipped-before-noop.txt \
  still-skipped-with-noop.txt \
  > candidates.txt

Review every source mark against the resulting WH/BH lists. Record sim-marked nodes that remain skipped under noop in not-executable.txt; do not count them as tested or remove their simulator mark without other evidence. Freeze the two candidate lists before execution and record their counts.

Stage 4: Run Every Candidate in Isolation

Keep the noop edit active. For each architecture, run the corresponding candidates.txt serially. Quote each complete node ID:

timeout --signal=TERM --kill-after=10s 180s \
  env PYTHONUNBUFFERED=1 \
  /opt/ttmlir-toolchain/venv/bin/python -m pytest -svvv "$node_id" \
  --sys-desc "$TT_METAL_SIMULATOR_HOME/ttrt-artifacts/system_desc.ttsys" \
  --path "$TT_METAL_SIMULATOR_HOME/pytest_artifacts_candidates" \
  --tb=short -rs

Write one log per node plus a TSV/JSON manifest containing architecture, node ID, elapsed time, exit code, and result. Continue after nonzero exits and print progress after every node. WH and BH loops may run concurrently, but never group multiple node IDs into one pytest invocation. Keep -s enabled so pytest capture cannot swallow the last TTSim error when the process exits.

Classify results from both exit status and the pytest terminal summary:

• passed: pytest reports the node passed. This is the only automatic unskip evidence.
• failed: pytest prints a normal failure/error summary.
• aborted: the process ends without a pytest terminal summary, even if its exit code is only 1.
• timed_out: the 180-second timeout expires.
• skipped, xfailed, or xpassed: record separately; none proves a clean simulator pass.

For an empty or unclear abort log, rerun that exact node with PYTHONUNBUFFERED=1 -svvv. Do not rerun a group.

After the matrix is complete:

• Remove a simulator mark only when every architecture to which it applies passed.
• Narrow mixed results to the smallest dtype, shape, op, or architecture scope. Single-chip WH is n150; single-chip BH is p150.
• If a mark on one parametrization axis mixes passing and failing combinations from another axis, restructure the parametrization so the mark names the exact combinations; do not retain the broad axis-level skip.
• A config list is AND'd; separate groups are OR'd. skip_config(["n150", "sim"]) means WH simulator only, while SkipIf("n150", "sim") means n150 or any simulator.
• Preserve unrelated skips. For example, if only sim is obsolete in SkipIf("ttnn", "emitc", "emitpy", "sim"), keep SkipIf("ttnn", "emitc", "emitpy").

If behavior differs from the previous TTSim version, run the smallest failing node against both versions with the same tt-mlir checkout and corresponding system descriptors.

Stage 5: Restore Triage State

Restore _get_current_environment() to return "sim" before any final validation. Inspect the diff rather than using a destructive checkout, and verify that conftest.py has no triage-only change. Keep the candidate manifests as validation artifacts, not source changes.

Stage 6: Run the Full Regression Sweep

This stage is mandatory and happens after skip updates and noop restoration.

Run every file in TEST_FILES, one file per pytest process, on WH and BH. Do not use one aggregate pytest command: an abort would hide the remaining files. Continue after failures. The two architecture loops may run in parallel.

timeout --signal=TERM --kill-after=15s 3600s \
  env PYTHONUNBUFFERED=1 \
  /opt/ttmlir-toolchain/venv/bin/python -m pytest -vv "$test_file" \
  --sys-desc "$TT_METAL_SIMULATOR_HOME/ttrt-artifacts/system_desc.ttsys" \
  --path "$TT_METAL_SIMULATOR_HOME/pytest_artifacts_regression" \
  --tb=short -rs

Some files may take more than 30 minutes, but an individual node should not take more than about three minutes. Monitor each live -vv log at least every three minutes. If progress stops, terminate the file run, isolate the last reported node with the 180-second candidate command, update the smallest necessary skip, and rerun that file. A generous file timeout is not a substitute for progress monitoring.

Record exactly one final status per file per architecture:

• passed: pytest exits 0 with a terminal summary; expected skips/xfails are allowed.
• failed: pytest reports normal failures.
• aborted: no pytest terminal summary, regardless of exit code.
• timed_out: the file timeout expires.

Any skip change made during regression requires a fresh run of that file on both architectures. Do not finish with an untriaged failed, aborted, or timed out file.

Stage 7: Final Verification and Handoff

1. Run focused WH and BH pytest for every file whose skips changed.
2. Verify all workflow-listed files have a final WH and BH status.
3. Syntax-check .github/workflows/call-test-ttsim.yml with available local tooling.
4. Run git diff --check and inspect git status and the complete diff.
5. Confirm conftest.py is restored and generated simulator assets/results are not in the source diff.
6. Summarize:
   • release tag and verified asset names;
   • descriptor YAML and generated .ttsys used per architecture;
   • candidate counts and result totals per architecture;
   • simulator skip removals, additions, and narrowings;
   • per-file regression status for WH and BH;
   • focused reruns and workflow validation;
   • any explicit remaining coverage gap.