Repository navigation
test(spmd): retire the mixed-kernel xfail and cover documented flags - #2946
georgebisbas wants to merge 1 commit into
Conversation
|
Important Review skippedAuto incremental reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (5)
Included review availability: This review used your included allowance. Your plan provides up to 2 included reviews per hour; 1 remain after this review. 📝 WalkthroughWalkthroughThe changes add SPMD coverage for early resolve and selective dumps, remove an expected-failure marker, activate a paged-attention test case, and update the related example command and test-helper comment. ChangesSPMD Early Resolve
SPMD Selective Dump
Paged Attention SPMD
Priority: ⬇️ Low Estimated code review effort: 3 (Moderate) | ~20 minutes Change: Other Merge Risk: ⚪ Minimal · up to This change adds test coverage and small documentation fixes without changing product behavior. No merge-blocking risk was found. 🚥 Pre-merge checks | ✅ 4✅ Passed checks (4 passed)
✨ Finishing Touches 💡 1🛠️ Fix failing CI checks 💡
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. A rabbit hops through four neat rows, Comment |
…ag surface - retire the stale xfail on test_spmd_mixed_kernel: the case XPASSes on a2a3sim (x3) and a5sim, and pypto#2887's device log recorded the same XPASS - add an allow_early_resolve=True SPMD dispatch ST case (documented on all three pl.spmd forms; previously parser/codegen coverage only) - add grid-scope dumps= correctness + --dump-args manifest tests (the third selective-dump front-end, next to the dump_tag and submit ones) - enable the ragged context_len paged-attention ST case (varctx path) - fix a stale path in the golden-writer UT comment; add a Run: line to the SPMD example
16c0776 to
6ff3aa1
Compare
What this PR offers
A tests-and-example-only hardening pass over the SPMD ST suite — no compiler or runtime changes (5 files, +247/−7). It retires a stale
xfailthat was misreporting a fixed bug, adds the first executed coverage for two documentedpl.spmdkwargs, enables a long-commented ragged-context case, and fixes two hygiene items.test_spmd_mixed_kernelprecisionxfail(test_spmd.py)test_golden_writer.pycommenttests/st/codegen/test_paged_attention_spmd.py(does not exist)context_lenpaged-attention casevarctxpath now has end-to-end coverageallow_early_resolve=Trueonpl.spmdSPMDEarlyResolveProgram/test_spmd_early_resolvedumps=[...]onpl.spmdpl.submitonlytest_dump_tag.py: correctness + two--dump-argsmanifest assertions (single dispatch retained; roles cover input + inout)Run:line (examples/models/09_paged_attention_spmd.py)Verification
Sim Docker (same commit,
a2a3sim/a5sim):tests/st/runtime/cross_core/test_spmd.py-k mixed_kernel, before the marker removaltest_paged_attention_spmd_ptoas[4-16-128-128-context_len3-32768]tests/st/runtime/framework_and_models/test_dump_tag.pyfull,--dump-argstests/ut/codegen/test_golden_writer.pypre-commiton the changed filesDeveloper verification — NPU (Ascend 910B2 / a2a3, via the shared task queue, at branch commit
16c07767):tests/st/runtime/cross_core/test_spmd.py(a2a3)test_spmd_mixed_kernel[a2a3]PASSED as a normal test (not xfail/XPASS/skipped);test_spmd_early_resolve[a2a3]PASSEDtests/st/runtime/framework_and_models/test_dump_tag.py -k SpmdDumps --dump-args(a2a3)task_id; rolesinput+inoutThe only warning in the hardware run is the known
attrs={"split": …}deprecation (F9 below) — recorded, not a failure. The branch touches no distributed/HCCL surface.Notes for review
test_spmd_mixed_kernelretirement is evidence-backed (sim XPASS ×4 + pypto#2887's device log) and was re-confirmed on hardware before this PR — a future need to re-xfailwould itself be a finding.test_spmd_early_resolveproves the flag survives lowering and does not change the computed result; it does not observe whether dispatch was actually staged early (that needs a swimlane / L2 record) — stated in the test docstring.optimizations=[pl.cross_core_slot(...)]onpl.spmdreaching the pipe-buffer sizing), F5 (example@pl.jitmigration — own PR), F7 (behaviouralsync_starttest), F9 (corpus-wideattrs={"split": …}→pl.func_attrmigration, 22 sites / 5 files).