Repository navigation
refactor(rocke/attention): parity 2D lane defaults to production spec (ceiling behind --variant hlpv) - #9466
Closed
AviralGoelAMD wants to merge 1 commit into
Conversation
AviralGoelAMD
force-pushed
the
users/avirgoel/ck/d256-parity-prod-default
branch
from
July 15, 2026 19:23
dcf2665 to
49c138f
Compare
AviralGoelAMD
changed the base branch from
users/avirgoel/ck/d256-gfx950-opt
to
users/avirgoel/ck/d256-parity-cleanup
July 15, 2026 19:23
AviralGoelAMD
force-pushed
the
users/avirgoel/ck/d256-parity-prod-default
branch
6 times, most recently
from
July 15, 2026 21:12
ea166bd to
55fe9b5
Compare
…; ceiling behind --variant hlpv Stacked on the builders/common hoist + PARITY_*-env->CLI cleanup. Addresses the branch with a variant switch. The 2D direct-launch lane now builds the PRODUCTION spec (`_tiled_spec_from_problem`) for ANY cohort by default -- a new cohort is measured with its real geometry with no per-cohort branch. The hand-tuned "aspirational ceiling" (num_warps=4 + half-local-PV) moves behind an explicit `--variant hlpv`. - Drops the `_d256_gfx950_fast` per-cohort branch from the harness. - Collapses the two duplicated build/compile/launch/time tails into one. - The shared tail consumes the CLI-param probes (--kq-swizzle / --kq-pad / --probe-occupancy) and the hoisted print_occupancy from the cleanup commit. GPU-verified on gfx950 (MI355X), 2D direct-launch: - prod default reproduces D256 (byte-identical kernel name; 391 us @ prefill_d256_gqa8_b16_sq4096) and the combo d64/b32 64x8 (54 us); - --variant hlpv reproduces the old ceiling (combo 64x8 47.7 us).
AviralGoelAMD
force-pushed
the
users/avirgoel/ck/d256-parity-prod-default
branch
from
July 15, 2026 21:25
55fe9b5 to
4531287
Compare
Base automatically changed from
users/avirgoel/ck/d256-parity-cleanup
to
users/avirgoel/ck/d256-gfx950-opt
July 15, 2026 21:28
Contributor
There was a problem hiding this comment.
Pull request overview
Refactors the gfx950 unified-attention parity/benchmark harness so the 2D direct-launch lane measures production-selected 2D geometry by default, while preserving the historical “aspirational ceiling” 2D spec behind an explicit --variant hlpv flag. This keeps new cohorts self-service (no harness edits) and makes parity probes cohort-agnostic.
Changes:
- Add
--variant {prod,hlpv}and plumb it into the 2D direct-launch lane (proddefault uses_tiled_spec_from_problemfor any cohort). - Remove the D256-special-cased fast-path import/branch and consolidate duplicated build/compile/launch/timing logic into a single shared tail.
- Apply KQ parity probes (
--kq-swizzle,--kq-pad) uniformly regardless of cohort/spec source.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+1448
to
+1450
| else: | ||
| spec = _tiled_spec_from_problem(problem) | ||
| cache_tag = ("prod_direct",) |
Contributor
Author
|
cherry picked into parent PR. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Change
The 2D direct-launch lane in
parity_unified_attention.pynow builds the production spec (_tiled_spec_from_problem) for any cohort by default; the hand-tuned "aspirational ceiling" (num_warps=4 + half-local-PV) moves behind an explicit--variant hlpv._d256_gfx950_fastimport/branch from the harness — new cohorts need no harness change.PARITY_KQ_SWIZZLE/PARITY_KQ_PAD) are now cohort-agnostic.Verification (gfx950 / MI355X, 2D direct-launch,
--skip-triton)Compared refactored
prodand--variant hlpvagainst this branch's current behavior, by kernel-name identity (noise-free) + timing:prod(new default)--variant hlpv(= old default)w2+interleave+ksb)w4base)w4+hlpv…)w4+hlpv)w4base)w2)w4+mfma32)prodreproduces D256 — byte-identical kernel name to this branch (395 vs 393 us).--variant hlpvreproduces the old ceiling default on every non-D256 shape tested.proddoes change what non-D256 lanes report (production geometry ≠ the hand-tuned ceiling) — and in every tested case production is equal-or-faster than the ceiling (d128 24 vs 37 us; combo 16x2 40 vs 43 us; combo 64x8 same kernel). Production has caught up to / surpassed the stale ceiling;--variant hlpvkeeps it measurable for any shape where you suspect it hasn't.it moves the non-D256 default numbers, but production ≥ ceiling on every shape measured (4 shapes; not exhaustive —
--variant hlpvis the escape hatch).Relates to AICK-1495 (S7).