b70-optimization-lab

Shared-gate component threshold miss

Date: 2026-07-23 EDT / 2026-07-24 UTC

Classification

• Outcome: valid rank-0 component measurement; preregistered timing threshold miss.
• Authorization packet: data/laguna-s-2.1-shared-gate-m8-component-authorization-20260724T025700Z.json
• Packet SHA-256: f748730bbb30be5268117015ffd8b8f88e81749eba4af9c3ca6ef6d99079c4ef
• Authorization commit: 3c2579115875463a81d5fe7edb4373af79f9b8d1
• Tools commit: c800d00580df3624b29e027ed22938936e1ed4ad
• Artifact root: /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-m8-component-c800d0058-20260724T025700Z
• Card 0 result: component_failed_stop_before_counters, with kind=timing_threshold.
• Cards 1 through 3: not started, as required after the first per-card miss.
• Analyzer and counters: not started.
• Every downstream authorization remained false, including model generation, endpoint, service, network, payload, reboot, and submission.

This packet and artifact root are terminal and must not be reused.

Exactness and dispatch

The marked M=8 shared-gate path used one torch.mm call with zero torch.bmm or fallback calls and exactly matched its literal raw-BF16 reference. The M=1-through-M=7, unmarked M=8, and marked-prefill controls stayed on their required torch.bmm paths.

All 128 distinct pre-timing epochs and all 32 distinct post-timing epochs reported both raw_uint16_equal=true and torch_equal=true for every comparison. The timing preflight also established raw equality among the literal BMM, literal MM, preallocated control output, and preallocated candidate output before any warm-up or timed arm.

Frozen timing result

The valid isolated primitive protocol used:

• 47 distinct 1,572,864-byte weights;
• 20 warm cycles per arm;
• 31 A-B-B-A paired blocks;
• 64 47-call cycles per arm and 3,008 calls per arm;
• one 128 MiB eviction pass before each arm;
• device synchronization only at arm boundaries;
• distinct preallocated input, weight, control-output, and candidate-output storage.

The native M=8 BF16 MM candidate beat the literal stride-zero B=8/M=1 BF16 BMM control in all 31 of 31 blocks. Savings ranged from 0.103621140625 ms to 0.133871273437 ms per 47-call cycle, with:

• mean saving: 0.12204077872983868 ms;
• median saving: 0.12085622656249995 ms;
• frozen minimum wins: 28 of 31;
• frozen minimum median saving: 0.150 ms.

The win-count condition passed, but the median-saving condition did not. No block individually reached the frozen 0.150 ms saving. The stable positive signal does not override the preregistered threshold and is not an accepted optimization.

Preserved campaign evidence

• campaign-start-checkpoint.json: a22fe7b6833a03194c4ba33581888c494ae47c6224327f5007e56115597736c1
• card0/pre-tensor-identity-checkpoint.json: 385be6570bd70bb857a9fe9f0eed48b281bbab61b0f6355b8784e495ab82220c
• card0/tensor-work-started-checkpoint.json: 857bc82c2af0b9618d285b28c5bb27285ce1a9316fafa74fd6e55e1e1cd5da9c
• card0/runtime-card-binding-checkpoint.json: d6216442d7f0551f088d34f92ce3e16968fd417936cbd831bebe89ee9d4ed90f
• card0/constructor-scope-proof.json: 9fbff9fe7f4ddada18e40749ffc03c3208abcc462dfdb5d4d0847cfa82d9b1dc
• card0/dispatch-proof.json: 58ca2d1777c484656f59812f8c47e07efb41868a9269ff32f633e0d5108ebc6e
• card0/timing.json: 0562d29903b22475db30a23d1b1a674b5c37fae0d1c9d5b5ec414183dcdce313
• card0/component-result.json: ebf755488cb71b506f259ccaf9c10f7b3f1677348544ebbac41f6620091354c1
• rank-0-terminal.json: 02d0e6c57d996a8362e8005fcc86655104212206a643860f0dd37650e7b9aa66
• campaign-terminal.json: 5b40fe51537a7e5999e263c2b5ad604d1e3b61efd38ea4f281ef1f173a8964ff

Decision

Do not rerun this packet and do not weaken the frozen threshold. The simple shared-gate BMM-to-MM substitution is a reproducible sub-threshold positive, not a bankable component win. Any follow-up must make a materially stronger candidate—such as a separately justified combined occupancy change or a deeper shared-expert kernel improvement—and preregister its own acceptance rule before device measurement. This result alone does not authorize counters or an endpoint benchmark.

This site is open source. Improve this page.