Skip to content

27B: test an Omega 32-row prompt kernel - #252

Closed
aien-spark wants to merge 2 commits into
perf/q38-head-boundsfrom
perf/q38-prefill-shape
Closed

aien-spark wants to merge 2 commits into
perf/q38-head-boundsfrom
perf/q38-prefill-shape

Conversation

@aien-spark

@aien-spark aien-spark commented Nov 1, 2026 •

Copy link
Copy Markdown
Collaborator

Short prompts currently use a 64-row matrix kernel even when fewer rows are populated. This adds an Omega admission policy for one 32-row, four-warp shape, reusing the existing parameterized generator. The default shape, arithmetic order, instruction encodings, wait distance and ownership checks remain unchanged. All other shapes stay refused.

This is stacked on #251. It introduces no serving selection policy. The existing test setter must be called before prefill buffer sizing or tile construction.

Validation:

  • The new cases fail on the parent because the generator refuses the shape. Host references and all eight simulator cases pass after the change, including tails, direct/indexed rows, both output layouts, two warp schedules and arithmetic mutants.
  • Admission/refusal cases and four compiled Omega mutants pass. Pinned binding regeneration and crumb verification pass.
  • The 30-token chip screen passes across four 27B matrix dimensions. Repeated full output buffers are identical. Three sampled rows per matrix match the incumbent reference contract: exact F32, at most one BF16 ulp for the SwiGLU transcendental path. All deliberate arithmetic faults are caught. The receipt is in om/prefill-shape/validation.json.

The short serving screen now passes exact128, seeded sampled32 and exact128 after sampling. First-visible latency remains around0.5 seconds; the chip timings are mixed, so no default change is proposed. See om/prefill-shape/serving-validation.json.

Full256 conversation/streaming checks, longer prompts, current-head CI and controlled performance comparisons remain pending. Keep this draft. No whole-model speed or first-token improvement is claimed. Production implementation is Omega through the pinned Bridge; Python is external verification only. Source notices are preserved and the staged diff passes whitespace checks.

aien-spark added a commit that referenced this pull request Nov 2, 2026
#259 #260)

native-x86-host on this head hit its 40 min timeout (cancelled, not failed); the same code at 9ade9f5 (diff since: evidence JSON + .crumb only) passed it in 34 min, and main e6598c7 passes it. All 10 other checks green on the merge with main. Independent review APPROVE-WITH-NOTES (see PR comment).
@aien-spark

Copy link
Copy Markdown
Collaborator Author

Merge desk (AI), 2026-11-02. Landed on main as part of the reviewed Ribbon stack through #260 (merge 183fa4f). This head (2d7a41c) is contained in main. Closing as landed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant