[Maintain] Merge branch 'main' of https://github.com/tile-ai/TileOPs into dev

[Five-HZ]

No description provided.

[gemini-code-assist]

Code Review

This pull request introduces extensive updates across the repository, including refactoring the CI runner Dockerfile to a multi-stage build, locking dependency versions in constraints.txt, and significantly updating the benchmark and test suites to be manifest-driven. It also improves the PR review loop logic to trigger on body and label changes. The review feedback highlights a few key issues: a bug in .claude/skills/review-tileops/loop.sh where jq fails to seed hashes if they are initially null, an invalid lambda argument passed to torch.testing.assert_close in benchmarks/ops/attention/bench_grouped_gemm_baselines.py, and a potential stale error leakage in scripts/validate_manifest.py that can be resolved by resetting _last_error at the start of the function.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

In jq, null == "" evaluates to false. If .last_body_hash or .last_labels_hash is missing (or null) in an existing meta.json file, the if condition will fail and the fields will remain null instead of being seeded with the current hashes. This prevents subsequent body/label changes from being detected.

Using the alternative operator // ensures that null or missing values default to "" and are correctly seeded.

Suggested change

	jq --argjson n "$NEW_IDLE" --arg bh "$BODY_HASH" --arg lh "$LABELS_HASH" \
	'.consecutive_idle=$n
	| .last_body_hash = (if .last_body_hash == "" then $bh else .last_body_hash end)
	| .last_labels_hash = (if .last_labels_hash == "" then $lh else .last_labels_hash end)' \
	"$META" > "$META.tmp" && mv "$META.tmp" "$META"
	jq --argjson n "$NEW_IDLE" --arg bh "$BODY_HASH" --arg lh "$LABELS_HASH" \
	'.consecutive_idle=$n
	| .last_body_hash = (if (.last_body_hash // "") == "" then $bh else .last_body_hash end)
	| .last_labels_hash = (if (.last_labels_hash // "") == "" then $lh else .last_labels_hash end)' \
	"$META" > "$META.tmp" && mv "$META.tmp" "$META"

[gemini-code-assist]

torch.testing.assert_close expects msg to be a str or None. It does not accept a callable/lambda. Passing a lambda will result in a TypeError or print the lambda's representation instead of the formatted message.

Simplify this by passing a formatted string directly.

Suggested change

	torch.testing.assert_close(
	a, e, rtol=_RTOL, atol=_ATOL,
	msg=lambda m, i=i: f"{what} disagrees with 3WG at rows [{i}:{i + chunk}]:\n{m}")
	torch.testing.assert_close(
	a, e, rtol=_RTOL, atol=_ATOL,
	msg=f"{what} disagrees with 3WG at rows [{i}:{i + chunk}]")

[gemini-code-assist]

Setting _last_error on the function object is a convenient way to pass error state, but since it is not reset to None on successful executions, a stale error from a previous failed call could leak if a caller checks _last_error after a successful call.

Reset _last_error to None at the start of the function to prevent stale error leakage.

def _forward_positional_params(cls) -> list[str] | None:
    """Get positional parameter names of cls.forward(), excluding 'self'.

    Only POSITIONAL_ONLY / POSITIONAL_OR_KEYWORD count. KEYWORD_ONLY
    params (those after ``*``) are not part of the positional tuple
    that manifest ``signature.inputs`` aligns against. Shared by check_l1
    and the C4 forward-signature parity check so they stay in lockstep.
    """
    _forward_positional_params._last_error = None  # type: ignore[attr-defined]
    try:
        sig = inspect.signature(cls.forward)