agentleFS
intelintel/torch-xpu-ops/.claude/skills/fix-reproduce/SKILL.md

fix-reproduce

Skill121 starsChanged 5 months agoApache-2.0View on GitHub

Use when asked to reproduce a bug, verify a nightly CI failure, or confirm a failure still exists on latest source. Verifies whether a bug still reproduces before an orchestrator commits time to a fix. Runs a three-stage fallback (nightly wheel -> source build -> CI environment alignment) and returns REPRODUCED / NOT_REPRODUCED / NO_REPRODUCER / CANNOT_VERIFY. Called by the issue-handler orchestrator.

  • Deletes or force-pushes
  • Installs packages

The file

Show its rules and text

We split this skill into the rules it gives an agent

27 rules. About 0 are common, 4 are distinctive, and 23 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈4

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 23

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    reproducer_command present?: If missing or empty: NO_REPRODUCER(reason=no_command).

    line 69
  2. Very specific:

    If the input is a stack trace without a command, the orchestrator should not call this skill in the first place (that is issue-triage's …

    line 72
  3. Very specific:

    If needs_tree=no, skip this section and go to Stage 1.

    line 101
  4. Very specific:

    If pytorch_dir was provided as input: git -C $pytorch_dir fetch origin.

    line 105
  5. Very specific:

    If not provided, clone into the torch-xpu-ops repo's gitignored scratch dir.

    line 107
  6. Very specific:

    If the reproducer targets test/xpu/..., make the working torch-xpu-ops tree available at $pytorch_dir/third_party/torch-xpu-ops.

    line 127
  7. Distinctive:

    Stop. Do not fall through to Stage 2 — the source tree is the same across stages, so a collect-miss at Stage 1 will miss at 2 and 3 too.

    line 150
  8. Very specific:

    Always reproduce against the latest available XPU nightly.

    line 182
  9. Very specific:

    Do NOT run the nightly-wheel reproducer with cwd inside any pytorch source checkout.

    line 215
  10. Very specific:

    When writing a standalone reproducer for a TestCase.assertEqual failure, use the test's own assertion.

    line 240
  11. Very specific:

    If the failure log says AssertionError: Tensor-likes are not close, the assertion is torch.testing._comparison.assert_close, which has …

    line 246
  12. Distinctive:

    Run according to the reproducer form matched in "Reproducer forms":

    line 265
  13. Distinctive:

    Pytest form: run the pytest invocation.

    line 267
  14. Very specific:

    If the re-run FAILs → REPRODUCED (the skip was hiding it).

    line 281
  15. Distinctive:

    Python one-liner / shell block: run the command (or the reproduce step extracted from the shell block, per "Reproducer forms").

    line 288
  16. Very specific:

    If Prepare (above) already ran (needs_tree=yes), $pytorch_dir is detached at origin/main with submodules initialized — skip to "Build and …

    line 321
  17. Very specific:

    Leave HEAD detached at origin/main at exit (the ci_commit fallback below re-detaches to a different sha.

    line 330
  18. Very specific:

    If the origin/main build fails for a reason unrelated to the bug (broken trunk, upstream infra issue, etc.) and ci_commit is available, …

    line 339
  19. Very specific:

    Only reached when nightly wheel and source build at origin/main both pass.

    line 370
  20. Very specific:

    When this skill is invoked from within CI (e.g.

    line 379
  21. Very specific:

    If both queries return empty: CANNOT_VERIFY(stage=ci_env, blocker=no_recent_successful_torch-xpu-ops_run).

    line 454
  22. Very specific:

    If no qualifying run in the last 100: CANNOT_VERIFY(stage=ci_env, blocker=no_recent_successful_xpu_workflow_run).

    line 521
  23. Very specific:

    If xpu.yml grows a new build_env not covered here: CANNOT_VERIFY(stage=ci_env, blocker=unknown_build_env=<name>).

    line 543
  24. Very specific:

    Do not shallow-fetch (--depth 1) here — pytorch's nested submodules resolve against pins that require the full history to be reachable.

    line 611
  25. Very specific:

    Run in the current shell — no docker run wrapper.

    line 630
  26. Very specific:

    When aligning yields REPRODUCED, fold discovered pieces (env vars, flags, cwd) into refined_command per its contract in the Output section.

    line 646
  27. Very specific:

    Quoting: use double quotes for inline python: When the reproducer embeds python code, write it as python -c "..." (double quotes), not …

    line 719

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reproduce — Verify the Bug Exists

Runs a test and determines whether the bug reproduces. Uses a three-stage approach: nightly wheel first (fast), source build at CI commit second (precise), CI environment alignment third (last resort).

The orchestrator decides what to do with the output — this skill only reports the result.

Inputs

  • reproducer_command — the sequence of shell commands that triggers the failure. Any of the forms below is valid; Stage 1's "Reproducer forms" section routes execution:

    1. A pytest node id or pytest ... invocation: pytest -v test/xpu/test_ops.py::TestFooXPU::test_bar_xpu_float32
    2. A python -c "..." snippet or a single python script.py line.
    3. A multi-line shell block (as is common in issue bodies): env setup, git clone, pip install, followed by the actual failing command.

    Missing or unrunnable inputs return NO_REPRODUCER up-front — see "## Preflight" below. Non-NO_REPRODUCER verdicts come from the Stage 1/2/3 execution flow.

    Providers (set by the orchestrator, not this skill):

    • Issue body — extracted by issue-triage from the reproducer section (via issue-handler).
    • CI failure log — the failing pytest node id from the nightly CI report, passed through as a batch sub-item.
  • stage — which reproduction path to run. Default auto.
    • auto — run the full three-stage fallback chain (nightly → source_build → ci_env). Used by orchestrators that need a definitive verdict.
    • nightly — only run Stage 1 (nightly wheel). PASS returns NOT_REPRODUCED(checked_stages=[nightly]) immediately; do NOT fall through to source build or ci_env. Cheapest option — suitable for a fast "does this still reproduce on latest nightly?" answer.
  • ci_commit — upstream commit hash from the CI report. Only used as a fallback base when origin/main fails to build (optional; ignored when stage=nightly).
  • pytorch_dir — path to a local PyTorch checkout (optional). Used whenever Prepare determines needs_tree=yes (any pytest form with a repo-relative path) as well as by Stage 2's source build. If absent, clone to <torch-xpu-ops-repo-root>/agent_space_xpu/pytorch/ (agent_space_xpu/ is the gitignored scratch dir at the torch-xpu-ops repo root — see the containing repo's AGENTS.md).
  • ci_repo — which CI to align against in Stage 3: pytorch or torch-xpu-ops. Optional; when absent, Stage 3 infers from the reproducer path (see "Determine ci_repo" in Stage 3).

Preflight

NO_REPRODUCER is a pre-execution verdict — the skill decides that there is nothing to run before touching any stage. Every other verdict (REPRODUCED / NOT_REPRODUCED / CANNOT_VERIFY) comes from Stage 1/2/3 execution.

reproducer_command present? If missing or empty: NO_REPRODUCER(reason=no_command). Stop.

If the input is a stack trace without a command, the orchestrator should not call this skill in the first place (that is issue-triage's reproduction_missing=yes case); if it slips through, this check catches it.

The pytest collected 0 items check happens in Prepare below, after the source tree it needs to run against is in place.

Prepare

Some reproducer forms need a pytorch source tree even at Stage 1 (nightly wheel path) — either because the test file lives inside pytorch/test/ or because it lives under torch-xpu-ops/test/xpu/ and imports common test utilities via sys.path.append("../../../../test/functorch") relative paths that only resolve from <pytorch_dir>/third_party/torch-xpu-ops/test/xpu/.

When to prepare

Set needs_tree from the reproducer form:

Reproducer formneeds_tree
pytest, path is repo-relative (test/xpu/..., test/...)yes
pytest, bare node id without a file path (TestFoo::test_bar)yes — pytest rootdir discovery needs the tree
pytest, path is absolute and exists on diskno
python -c "..." / python /abs/path/script.pyno
shell block (issue body: clone + install + run)no — the block does its own setup

If needs_tree=no, skip this section and go to Stage 1.

Get the pytorch tree

If pytorch_dir was provided as input: git -C $pytorch_dir fetch origin.

If not provided, clone into the torch-xpu-ops repo's gitignored scratch dir. Resolve the path explicitly rather than relying on cwd:

XPU_OPS_ROOT=$(git -C <path-to-torch-xpu-ops-checkout> rev-parse --show-toplevel)
pytorch_dir="$XPU_OPS_ROOT/agent_space_xpu/pytorch"
if [[ ! -d "$pytorch_dir/.git" ]]; then
  git clone --filter=blob:none https://github.com/pytorch/pytorch.git "$pytorch_dir"
fi
git -C "$pytorch_dir" fetch origin
git -C "$pytorch_dir" checkout --detach origin/main
git -C "$pytorch_dir" submodule update --init --recursive

The tree is not built here — Stage 1 uses the nightly wheel for the runtime; the source tree only supplies test files and support modules. Stage 2 reuses the same tree and builds it there.

torch-xpu-ops test path

If the reproducer targets test/xpu/..., make the working torch-xpu-ops tree available at $pytorch_dir/third_party/torch-xpu-ops. The build's dev-override recipe (symlink or replace-clone) applies; see xpu-build-pytorch. From here on, Stage 1's cwd for pytest is $pytorch_dir/third_party/torch-xpu-ops/test/xpu/.

Collect-only check (pytest form)

Regardless of needs_tree, if reproducer_command matches the pytest form (starts with pytest, python -m pytest, or is a bare pytest node id), run:

pytest --collect-only <node_id>

Cwd:

  • needs_tree=yes → from the tree just prepared (for test/xpu/... targets, that's $pytorch_dir/third_party/torch-xpu-ops/test/xpu/)
  • needs_tree=no → from any non-pytorch directory (the reproducer's absolute path resolves on its own)

Output shows collected 0 items? → NO_REPRODUCER(reason=collected_zero). Stop. Do not fall through to Stage 2 — the source tree is the same across stages, so a collect-miss at Stage 1 will miss at 2 and 3 too.

Non-pytest forms have no equivalent pre-execution check.

Stage 1: Nightly Wheel (fast path)

Most failures reproduce here. Start here before doing anything heavier.

Reproducer forms

Three forms; each dispatches differently in "Run test" below:

  • Pytest form — reproducer_command starts with pytest, python -m pytest, or is a bare pytest node id (.../test_foo.py::TestBar::test_baz). The collect-only check (see "## Prepare" above) has already run.
  • Python one-liner / single-script form — python -c "..." or python path/to/script.py.
  • Shell-block form — a multi-line block copied out of an issue body: env setup + git clone + pip install + the failing command. Split it into setup steps (everything before the failing command) and the reproduce step (the last command that exercises the failing path). Only the reproduce step's outcome determines the verdict; setup-step failures return CANNOT_VERIFY(stage=<current>, blocker=<the failing setup step>).

The Working directory and Use the test's own assertion rules below apply to all three forms.

Install

Always reproduce against the latest available XPU nightly. Do not reuse a stale wheel from a previous session — a bug may already be fixed in a newer nightly, and re-verifying an old wheel produces misleading REPRODUCED results.

# Query available versions (informational — pip install --upgrade below
# will pick a resolvable one, which may lag the newest entry here by a
# day when the index metadata refreshes before all wheels land).
pip3 index versions torch --pre \
  --index-url https://download.pytorch.org/whl/nightly/xpu
pip3 install --pre --upgrade torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/nightly/xpu

Post-install, check that torch, torchvision, torchaudio are all from the same day — pip's resolver can leave a mixed set (either torch older than the auxiliary wheels, or the reverse). If they diverge, uninstall all three and reinstall together:

pip3 uninstall -y torch torchvision torchaudio
pip3 install --pre torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/nightly/xpu

Record the exact wheel version used (python -c "import torch; print(torch.__version__)") in the reproduce output and in any issue comment, so downstream stages and re-verifications know which nightly was tested.

Working directory

Do NOT run the nightly-wheel reproducer with cwd inside any pytorch source checkout. Python resolves import torch against the local torch/ package before site-packages, so it will load the in-tree torch/_C.so built at whatever revision that tree happens to be — typically stale relative to the installed wheel — and fail with ImportError: undefined symbol: .... Either cd $(mktemp -d) (or any non-pytorch dir) before running, or invoke the reproducer with an absolute path from outside the tree.

torch-xpu-ops test invocation

Applies when the reproducer targets a test under torch-xpu-ops/test/xpu/. Prepare has already ensured the pytorch tree exists at $pytorch_dir with the working torch-xpu-ops tree at $pytorch_dir/third_party/torch-xpu-ops. Invoke the reproducer from $pytorch_dir/third_party/torch-xpu-ops/test/xpu/ — the relative sys.path.append("../../../../test/functorch") in those tests only resolves from that cwd.

The pytorch tree does NOT need to be built for the nightly-wheel path — the wheel provides the runtime, the source tree only supplies test files and support modules.

Use the test's own assertion

When writing a standalone reproducer for a TestCase.assertEqual failure, use the test's own assertion. Do NOT substitute torch.allclose, torch.equal, or bare == — they have different (usually stricter) default tolerances and will manufacture false positives.

If the failure log says AssertionError: Tensor-likes are not close, the assertion is torch.testing._comparison.assert_close, which has dtype-specific defaults (bf16: rtol=0.016, atol=1e-5). Reproduce through assert_close or via TestCase.assertEqual:

import sys; sys.path.insert(0, "<pytorch>/test")
from torch._dynamo.test_case import TestCase   # or the base class the failing test uses

class T(TestCase):
    def test_x(self, device):
        ...
        self.assertEqual(out_ref, out)

T().test_x(device='xpu')

Run test

Run according to the reproducer form matched in "Reproducer forms":

  • Pytest form: run the pytest invocation. Result interpretation:
    • FAILED → REPRODUCED.
    • all skipped by @skipIfXpu → the marker is likely hiding the actual failure. Temporarily remove it in place, re-run once to check what happens without the skip, then revert the file so no change escapes this skill:

      # Remove @skipIfXpu from the target test file(s), then:
      pytest <node_id>
      # Regardless of outcome, revert:
      git checkout <test_file>

      If the re-run FAILs → REPRODUCED (the skip was hiding it). If it PASSes → treat as PASSED per the Decision table. Only return CANNOT_VERIFY when the skip is environmental (not an XPU marker, e.g. @skipIf(not has_cuda) shielding an unavailable dep).

    • xfailed → treat as FAILED (REPRODUCED).
    • PASSED → per the Decision table below.
  • Python one-liner / shell block: run the command (or the reproduce step extracted from the shell block, per "Reproducer forms"). Result interpretation:

    • Exit code non-zero and output matches the failure pattern named in the issue (traceback, error message, or specific assertion) → REPRODUCED.
    • Exit code zero → PASSED (per the Decision table below).
    • Exit code non-zero but cause is unrelated (missing dependency surfacing inside the reproduce step, permission error, missing device) → CANNOT_VERIFY(stage=nightly, blocker=<...>). Do NOT report REPRODUCED on a setup or infra failure.

    all skipped / xfailed do not apply to these forms — they are pytest-specific concepts.

Decision

ResultConditionAction
CANNOT_VERIFYenv problem (wheel install failed, runtime missing)Report to orchestrator, stop
REPRODUCEDFAILEDReturn REPRODUCED(stage=nightly, refined_command=...)
→ stage 2PASSED and stage=autoProceed to source build at origin/main to confirm
NOT_REPRODUCEDPASSED and stage=nightlyReturn NOT_REPRODUCED(checked_stages=[nightly]) — do NOT fall through

Stage 2: Source Build at origin/main

Nightly passing is not conclusive — it may lag behind CI. Build from origin/main to verify. Even when the failure came from a specific CI commit, we only consider fixes on top of origin/main — downstream stages branch off it.

Prepare pytorch checkout

If Prepare (above) already ran (needs_tree=yes), $pytorch_dir is detached at origin/main with submodules initialized — skip to "Build and run".

Otherwise (Prepare was skipped because needs_tree=no), run Prepare's "Get the pytorch tree" recipe now: resolve $pytorch_dir (from input or $XPU_OPS_ROOT/agent_space_xpu/pytorch), clone if missing, fetch, checkout --detach origin/main, submodule update --init --recursive.

Leave HEAD detached at origin/main at exit (the ci_commit fallback below re-detaches to a different sha; downstream stages branch off whatever this stage settled on).

Build and run

Load the xpu-build-pytorch skill and follow it for the build. Do not hand-roll the build here.

If the origin/main build fails for a reason unrelated to the bug (broken trunk, upstream infra issue, etc.) and ci_commit is available, fall back once:

git -C $pytorch_dir checkout --detach $ci_commit
git -C $pytorch_dir submodule update --init --recursive

Rebuild via xpu-build-pytorch. If this succeeds, proceed with the test on ci_commit and record base=<ci_commit_sha> in the output so the orchestrator branches its fix off the same base. If the fallback build also fails, escalate as CANNOT_VERIFY(blocker=trunk and ci_commit both fail to build) rather than silently reproducing on some other base.

Then run the reproducer following the form-specific rules in Stage 1 "Run test" (pytest interpretation vs python/shell interpretation).

Decision

Applies only when stage=auto — stage=nightly returns at Stage 1.

ResultAction
CANNOT_VERIFYReport to orchestrator, stop
REPRODUCEDReturn REPRODUCED(stage=source_build, base=origin/main|<ci_commit_sha>, refined_command=...)
PASSEDProceed to stage 3

Stage 3: CI Environment Alignment

Only reached when nightly wheel and source build at origin/main both pass. The failure may be specific to the CI environment: wheels built under CI toolchain, XPU/oneAPI stack pinned to a specific version, environment variables set by CI.

Assumption: the agent already runs inside the CI test container

Both pytorch/pytorch and torch-xpu-ops run their XPU tests inside a container (declared as container: image: in the workflow yaml). When this skill is invoked from within CI (e.g. via @torchxpubot fix), the agent is already inside that container — kernel modules, /dev/dri, oneAPI stack, and python are already the CI ones. This stage does not docker pull or docker run. It aligns the installed wheels + pytorch source checkout to the CI wheel, then runs the reproducer directly.

The skill logs the CI image reference it identified (for context in the report), but does not exec into it.

Clean up Stage 2 artifacts

Stage 2 may have left behind build/, torch/lib/*.so, or a modified third_party/xpu.txt from the dev-override. Left in place, Python will pick the host-built (stale-relative-to-CI-wheel) torch/_C.so off sys.path and error with undefined symbol before the reproducer runs.

# Restore xpu.txt to origin's pinned commit (in case Stage 2 rewrote it).
git -C "$pytorch_dir" checkout -- third_party/xpu.txt
# Discard stage-2 build outputs. `git clean` does not recurse into
# nested repositories by default, so third_party/torch-xpu-ops (a
# separate git repo) is preserved without needing `-e`.
git -C "$pytorch_dir" clean -fdx

Alternatively, run the reproducer with cwd outside $pytorch_dir (e.g. cd /tmp) so import torch resolves against site-packages, matching Stage 1's "Working directory" rule. Do at least one.

Determine ci_repo

Pick which CI to align against based on the reproducer:

Reproducer clueci_repo
Path contains test/xpu/ or torch-xpu-opstorch-xpu-ops
Path is pytorch/test/... or absolute path inside a pytorch treepytorch
Ambiguous / python -c snippet with no pathtry torch-xpu-ops first, fall back to pytorch

The orchestrator may also pass ci_repo explicitly; when set, use it and skip the heuristic.

Path A — ci_repo=torch-xpu-ops

Wheels come from intel/torch-xpu-ops's own build workflow and stay on GitHub Actions artifact storage. Fetch via gh run download, not S3.

A1. Find the latest successful wheel-producing run

The build job lives in _linux_build.yml (a reusable workflow called by pull.yml and nightly_ondemand.yml). It uploads the artifact Torch-XPU-Wheel-<pr|sha>-<runid>-<attempt>[-category].

# Nightly is the primary source (fresh main-branch build every night).
# Fall back to pull.yml when nightly has been failing for a stretch —
# pull.yml runs on PRs against main and its wheels are close enough for
# CI-env alignment.
RUN=$(gh run list --repo intel/torch-xpu-ops \
  --workflow nightly_ondemand.yml \
  --status success --limit 1 \
  --json databaseId,headSha,createdAt)
if [[ "$RUN" == "[]" ]]; then
  RUN=$(gh run list --repo intel/torch-xpu-ops \
    --workflow pull.yml \
    --status success --limit 1 \
    --json databaseId,headSha,createdAt)
fi
# Empty here → report CANNOT_VERIFY per the paragraph below (do not
# `exit`; the skill returns a verdict, it does not terminate the shell).
RUN_ID=$(jq -r '.[0].databaseId' <<<"$RUN")

If both queries return empty: CANNOT_VERIFY(stage=ci_env, blocker=no_recent_successful_torch-xpu-ops_run).

A2. Download the wheel artifact

CI_ENV_DIR="$XPU_OPS_ROOT/agent_space_xpu/ci_env"
WHEELS_DIR="$CI_ENV_DIR/wheels"
rm -rf "$WHEELS_DIR" && mkdir -p "$WHEELS_DIR"

# Artifact name is Torch-XPU-Wheel-<pr|sha>-<runid>-<attempt>[-category].
# `gh run download -n <name>` requires exact name; use pattern instead.
# If the run uploaded multiple category variants (target/baseline via
# `_linux_build.yml`'s `category` input), --pattern pulls all of them
# and the flatten below will clobber same-named wheels. In that case
# pass --name <specific-artifact> to pick one variant.
gh run download "$RUN_ID" --repo intel/torch-xpu-ops \
  --pattern 'Torch-XPU-Wheel-*' --dir "$WHEELS_DIR"

# gh unpacks each artifact into its own subdir; flatten:
find "$WHEELS_DIR" -mindepth 2 -name '*.whl' -exec mv {} "$WHEELS_DIR" \;
# Empty here → CANNOT_VERIFY(stage=ci_env, blocker=no_wheel_in_artifact).
# Do not `exit`; return the verdict via the skill's Output section.
find "$WHEELS_DIR" -maxdepth 1 -name '*.whl' | grep -q .

A3. CI image (reference only)

For torch-xpu-ops the test container is intelgpu/ubuntu-24.04-lts2:2523.40 (see .github/workflows/_linux_ut.yml). Record the tag for the report; do not pull. Kept for local-investigation convenience (someone reproducing outside CI can docker run this image manually) — the skill itself relies on the Assumption above.

Path B — ci_repo=pytorch

Wheels come from pytorch/pytorch's xpu workflow and land on gha-artifacts S3.

B1. Find the latest successful xpu workflow run

Accept only runs where every linux-*/ build job succeeded — partial runs still upload partial artifacts.

# Match by display name first; fall back to path in case pytorch/pytorch
# renames the workflow's `name:` field (the file path is more stable).
WF_ID=$(gh api "repos/pytorch/pytorch/actions/workflows?per_page=100" --paginate \
  --jq '.workflows[] | select(.name=="xpu" or .path==".github/workflows/xpu.yml") | .id' \
  | head -1)

RUN_ID=""
for page in 1 2 3 4 5; do
  while IFS=$'\t' read -r rid _ _; do
    conclusions=$(gh api \
      "repos/pytorch/pytorch/actions/runs/$rid/jobs?per_page=100" --paginate \
      --jq '.jobs[] | select(.name | test("^linux.*/ build$")) | .conclusion')
    [[ -z "$conclusions" ]] && continue
    grep -qv '^success$' <<<"$conclusions" && continue
    RUN_ID=$rid; break 2
  done < <(gh api \
    "repos/pytorch/pytorch/actions/workflows/$WF_ID/runs?status=completed&per_page=20&page=$page" \
    --jq '.workflow_runs[] | [.id, .head_sha, .created_at] | @tsv')
done

If no qualifying run in the last 100: CANNOT_VERIFY(stage=ci_env, blocker=no_recent_successful_xpu_workflow_run).

B2. Pick the right build_env

Per the Assumption above, the agent is already inside a compatible CI container — the image column below is a lookup for the report only, not a pull target. It is kept in case a local investigation (outside CI) wants to spin up the same image manually to reproduce; the skill itself does not use it.

pytorch/pytorch's xpu.yml currently defines only py3.10 linux builds:

build_envHardwareImage (reference)
linux-noble-xpu-n-py3.10PVCghcr.io/pytorch/ci-image:pytorch-linux-noble-xpu-n-py3-<docker-tree-hash>
linux-noble-xpu-n-py3.10-clientBMGghcr.io/pytorch/ci-image:pytorch-linux-noble-xpu-n-py3-client-<docker-tree-hash>
linux-jammy-xpu-n-1-py3.10PVCghcr.io/pytorch/ci-image:pytorch-linux-jammy-xpu-n-1-py3-<docker-tree-hash>

-client suffix = BMG (client GPU); no suffix = PVC (datacenter). Match the runner's hardware; default to PVC when unknown.

If xpu.yml grows a new build_env not covered here: CANNOT_VERIFY(stage=ci_env, blocker=unknown_build_env=<name>). Do not guess. When multiple envs match, iterate them in sorted order for determinism.

<docker-tree-hash> is the git tree hash of .ci/docker/ at the run's commit (see upstream _runner-determinator.yml "Compute .ci/docker tree hash"). Only needed if the report wants a fully-qualified image reference; skill does not pull the image.

B3. Download wheel artifacts

Artifacts live at:

https://gha-artifacts.s3.amazonaws.com/pytorch/pytorch/<run_id>/<build_env>/artifacts.zip

Probe availability with --range 0-0 -L (zero-byte GET); HEAD may be rejected by some intermediaries in front of this bucket, byte-range GET returns 200 or 206:

url="https://gha-artifacts.s3.amazonaws.com/pytorch/pytorch/$RUN_ID/$BUILD_ENV/artifacts.zip"
http_status=$(curl -s -o /dev/null -w "%{http_code}" --range 0-0 -L "$url")
# 200 or 206 -> ok; anything else -> skip this build_env

Download and extract:

CI_ENV_DIR="$XPU_OPS_ROOT/agent_space_xpu/ci_env"
WHEELS_DIR="$CI_ENV_DIR/wheels"
ARTIFACTS_ZIP="$CI_ENV_DIR/artifacts.zip"
rm -rf "$WHEELS_DIR" && mkdir -p "$WHEELS_DIR"
# --retry survives transient drops. A silent truncation of the 1.2 GB
# zip surfaces later as "cannot find zipfile directory" from unzip.
curl -sL -f --retry 3 --retry-delay 5 "$url" -o "$ARTIFACTS_ZIP"

# Layout varies: some build envs pack wheels under dist/, others at root.
unzip -o -j "$ARTIFACTS_ZIP" 'dist/*.whl' -d "$WHEELS_DIR" \
  || unzip -o -j "$ARTIFACTS_ZIP" '*.whl' -d "$WHEELS_DIR"
# Empty here → CANNOT_VERIFY(stage=ci_env, blocker=no_wheel_extracted).
find "$WHEELS_DIR" -maxdepth 1 -name '*.whl' | grep -q .

Install the CI wheel and align source

Same for both paths. Uninstall any existing torch stack first — the Stage 1 nightly is still resident:

pip uninstall -y torch torchvision torchaudio pytorch-triton-xpu triton_xpu 2>/dev/null || true
pip install --force-reinstall "$WHEELS_DIR"/*.whl

Align the pytorch source tree to the wheel's commit so tests that import support modules from pytorch/test/ see matching code (Prepare left $pytorch_dir detached at origin/main, which is not the wheel's commit):

TORCH_COMMIT_ID=$(python -c 'import torch; print(torch.version.git_version)')
git -C "$pytorch_dir" fetch origin
git -C "$pytorch_dir" checkout --detach "$TORCH_COMMIT_ID"
git -C "$pytorch_dir" submodule update --init --recursive

Do not shallow-fetch (--depth 1) here — pytorch's nested submodules resolve against pins that require the full history to be reachable; a shallow fetch surfaces later as "cannot find <sha>" in submodule update.

TORCH_COMMIT_ID (the wheel's build commit) is a temporary alignment only — it makes pytorch/test/ support modules match the installed wheel's binary so the test can run. It is not the fix base and is never returned as base. Downstream fixes always branch off origin/main (see Stage 2), so this stage still reports base=origin/main; TORCH_COMMIT_ID stays internal to Stage 3.

For torch-xpu-ops test paths, ensure the working torch-xpu-ops tree is at $pytorch_dir/third_party/torch-xpu-ops (Prepare already handled this if needs_tree=yes; if it didn't, do it now via the xpu-build-pytorch dev-override recipe).

Run the reproducer

Run in the current shell — no docker run wrapper. Working-directory rule from Stage 1 applies: for torch-xpu-ops tests use $pytorch_dir/third_party/torch-xpu-ops/test/xpu/; for a non-test-file reproducer, cd /tmp (or any non-pytorch dir).

Result interpretation is the form-specific rule from Stage 1 "Run test" — pytest FAILED / all skipped / xfailed, or non-pytest exit code

  • output-vs-failure-pattern.

What to check if the failure still doesn't reproduce

From the CI job log, extract and align remaining differences:

  • Full test command with all flags (--timeout, -x, specific env vars)
  • Any environment variables set in the CI job (ZE_AFFINITY_MASK, PYTORCH_TEST_WITH_XPU, IS_XPU_CI, etc.)

When aligning yields REPRODUCED, fold discovered pieces (env vars, flags, cwd) into refined_command per its contract in the Output section.

Decision

ResultAction
CANNOT_VERIFYReport to orchestrator, stop
REPRODUCEDReturn REPRODUCED(stage=ci_env, base=origin/main, refined_command=...) — base is origin/main, not the wheel's TORCH_COMMIT_ID the tree is currently detached at
PASSEDReturn NOT_REPRODUCED(checked_stages=[nightly, source_build, ci_env]) — issue no longer exists; orchestrator reports to user or triage collects reason

Output

Return one of these to the orchestrator:

REPRODUCED
  stage: nightly | source_build | ci_env
  base: origin/main | <ci_commit_sha>    # base for downstream build. Default origin/main (also for stage=ci_env). ci_commit_sha only when stage=source_build fell back to ci_commit. Stage 3's TORCH_COMMIT_ID wheel-alignment checkout is never returned as base.
  refined_command: <single shell-executable string>

Markdown block

Emit this block with the output above; the orchestrator appends it to the session comment.

<!-- agent:reproduce -->

## Reproduce

<one or two sentences: which build was used — nightly wheel version or
source sha — and which device.>

| Test case | Verdict | Observed |
|---|---|---|
| `test_foo_xpu_float8_e4m3fn` | REPRODUCED | `NotImplementedError: "bar_kernel" not implemented for 'Float8_e4m3fn'` |

<optional: one paragraph tying the signature to the CI job log, with a
link to the failing job; or naming what blocked a stage.>

*Automated by fix-reproduce.*

One row per test case (for a batch, label rows by sub-item number: 4. \test_foo...`). Observed` is the one-line failure signature, backticked — never a pasted traceback.

refined_command contract. A single shell-executable string that, run by itself, reliably triggers the failure. A downstream skill (a fix-verifier, a skip-list per-entry runner, etc.) invokes it directly (e.g. via bash -c "$refined_command") after applying a candidate fix to check whether the failure is gone. Consequences:

  • Include everything needed to reproduce. Env vars go as inline prefix (ZE_AFFINITY_MASK=0 pytest ...), a required cwd goes as a cd <dir> && prefix.
  • Do NOT include setup steps. git clone, pip install, wheel-download, etc. that appeared in the input shell block are excluded. The caller has already paid that cost; refined_command should re-trigger the failure, not re-provision the environment.
  • No docker run wrapper. Stage 3 runs the reproducer directly in the current shell (the CI job is already inside its container, and the caller of refined_command runs from an equivalent env). If the caller needs container-level isolation, that is its concern, not refined_command's.
  • Not just the input command. For the pytest form, refined_command may add -sv, --timeout <N>, or -x that Stage 1 used to get a usable failure signal. For the shell-block form, refined_command is the extracted "reproduce step" (usually the last line), not the whole block.
  • Quoting: use double quotes for inline python. When the reproducer embeds python code, write it as python -c "..." (double quotes), not python -c '...'. Downstream callers wrap the string in bash -c "$refined_command"; single-quoted python payloads compose poorly through that wrapping.
NOT_REPRODUCED
  checked_stages: [nightly] | [nightly, source_build, ci_env]
  reason: <what was checked and confirmed to pass>

NO_REPRODUCER
  reason: no_command | collected_zero
  (returned when either no reproducer_command was provided, or pytest
  reports `collected 0 items` for the provided command)

CANNOT_VERIFY
  stage: nightly | source_build | ci_env
  blocker: <what went wrong>

The orchestrator decides the next step based on this output.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.