You are viewing a plain text version of this content. The canonical link for it is here.
Posted to discuss-archive@tvm.apache.org by GitBox <gi...@apache.org> on 2026/08/28 05:57:58 UTC

[GH] (tvm/fix/issue-20048-dlight-reduction): Workflow run "Lint" failed!

The GitHub Actions job "Lint" on tvm.git/fix/issue-20048-dlight-reduction has failed.
Run started by GitHub user Junius-Wynn (triggered by Junius-Wynn).

Head commit for run:
1fb55209a75af2cd7a927d0738b49ba5ebd2482b / Junius-Wynn <jw...@nyu.edu>
[Fix][S-TIR][DLight] Guard non-affine reduction write-back

Issue #20048 reports that relax.build can abort with ScheduleError while compiling a valid CUDA conv2d. The GPU Reduction rule fuses all spatial loops, but its inner-spatial schedule derives the thread extent from only the innermost spatial extent. For some output shapes, rfactor and reverse_compute_at then produce non-quasi-affine write-back bindings. Normalized accesses may also split an iterator, making the effective fused extent differ from its original domain.

Detect the unsafe geometry before applying the inner-spatial schedule and return None so the next applicable schedule rule can handle the block. Derive both spatial order and effective extents from the normalized dominant read. For accesses with a non-unit lower factor, use split_expr.extent so the guard and thread extent match the loops produced by _normalize. Keep unexpected ScheduleError failures visible at the rule-selection layer instead of catching them globally.

Add regression coverage for the reported convolution-like access, a plain reduction without mixed spatial/reduction indices, reordered access, a normalized divided access using x // 2, affine write-back, and propagation of unexpected ScheduleError. Fixes #20048

Report URL: https://github.com/apache/tvm/actions/runs/33070088621

With regards,
GitHub Actions via GitBox


---------------------------------------------------------------------
To unsubscribe, e-mail: discuss-archive-unsubscribe@tvm.apache.org
For additional commands, e-mail: discuss-archive-help@tvm.apache.org