b70-optimization-lab

Laguna separate shared gate+up native M=8 MM component pass

Date: 2026-07-24 America/Toronto

The separate gate-then-up native-M8 treatment passed its first and only authorized four-card component campaign. This is exact component evidence, not an endpoint benchmark, model generation, or a LocalMaxxing record.

Sealed campaign

The tooling and authorization identities are:

• component tools commit 4cef996c94502ad06233caa55d5be019d13a5114;
• packet-only authorization commit f04d7431224017859ef892b1251f2a87fc1dee4a;
• packet data/laguna-s-2.1-shared-gate-up-m8-component-authorization-20260724T051216Z.json;
• packet raw SHA-256 415411d1cce8ce2a9210032e1d54543bd66eff3f26fa08ffc6fa8436002d302b;
• packet canonical JSON SHA-256, excluding its final newline, 75c13ddee2c1e0c8a1256341fa97b928de3c960c2e18a9420ef4f9be58326780;
• vLLM 503f7784cf9d1704109b1e4650427fb4f417d604; and
• XPU kernels c59aaadbbfd350c2b5f4ad663e247c2811ae3181.

The fresh live-evidence root is entirely on local NVMe:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-up-m8-component-4cef996c9-20260724T051216Z

The coordinator acquired the campaign root at 2026-07-24T05:16:31.543275Z. All four cards, the aggregate analyzer, the finalizer, and the final verifier exited successfully. The independently rerun final verifier also passed from the sealed files.

The principal final artifacts are:

• component aggregate SHA-256 d9a8e9ce8a55d79bfb9a33f4a3e926c6ce3c1ceb2039445de137e2102d3a414c;
• campaign terminal SHA-256 f3e477d11bc515fea2f387de4c4d30fcf3958eaec979fd77f8b69bc66099afc3;
• final manifest SHA-256 8aa2a45a8bcc31c5d2e84e5f55568ad43144ebd6c380d6d914cf91f77d10a10d; and
• final manifest coverage of all 675 pre-manifest files, with the complete campaign containing 676 regular files.

Timing result

Timing retained two separate, ordered projections. The control issued a gate BMM and then an up BMM; the candidate issued a gate native MM and then an up native MM. Merged N512, logical B16, packing, fusion, reordering, overlap, and shared-down treatment remained forbidden.

Each arm used 47 distinct inputs, 47 distinct gate weights, 47 distinct up weights, four independent 47-slot output rings, one 128 MiB touch, 20 untimed warm cycles, and 64 complete cycles with exactly 6,016 projection calls. The analyzer recomputed all 31 A-B-B-A blocks from raw nanoseconds.

Rank	BDF	Wins	Control median ms/cycle	Candidate median ms/cycle	Median saving ms/cycle	Relative
0	0000:23:00.0	31/31	3.021198	2.737756	0.285200	9.4400%
1	0000:27:00.0	31/31	2.978052	2.662922	0.308360	10.3544%
2	0000:43:00.0	31/31	3.341755	3.020350	0.321073	9.6079%
3	0000:47:00.0	31/31	3.117406	2.769954	0.348841	11.1901%

Every physical card independently exceeded the frozen requirements of at least 28/31 wins and at least 0.20 ms median saving per complete 47-layer gate+up cycle. The narrowest result was rank 0 at 0.285200 ms; no cross-card average rescued a card.

Exactness and identity

Each card passed 128 pre-timing changing epochs and 32 post-timing exact replays. Every epoch checked nine boundaries: gate, gate repeat, up, up repeat, gate SiLU, the actual BF16 multiply, shared down, shared+routed add, and fixed-rank reduction. That is 1,440 raw-BF16 plus torch.equal comparisons per card and 5,760 across the campaign, all equal.

All four cards independently reproduced the same exactness digests:

• pre fixture e480bfbee66add59da38b786b403872ea7ed7e2dade35a53820f1a4b34eca440;
• pre output 7c6ab1b9d4e62355321a4567e248cd42dfd7b87382f9d7980667f8ba36ce42bc;
• post fixture b963e1c58965e0612879d3dc300101a460c5db7778eb82dc527d0c11a6dd8a75; and
• post output 94e5fd23e806664325c72acedd5571895ca5f47f793a4c5063a830416d92fe0a.

The sealed dispatch proof retained exactly two ordered native MMs for the marked pair, incumbent behavior for unmarked paths, and all rejection classes. The analyzer also revalidated four distinct physical UUID/BDF mappings, Torch Level Zero UUID byte reversal, one boot 0b7f98a5-e50a-46a5-81ea-15938b55317a, the source, binaries, runtime, model configuration, Stage-0 certificate, packet-only Git lineage, and strict artifact inventory.

The per-card result SHA-256 values are:

• rank 0: 15638454c3d004349cc0946707d74274325df5a8f5bd6700522c12a7e33fb5a2;
• rank 1: 077a816328db9abff39e9bb0688a26c992cf25deeba795bf08757b71f193ecab;
• rank 2: 08949ddd560c5fd0ec9e900ec89ab1d1eeacdd6b8f1d271be602bf2a03299e68; and
• rank 3: 8360164baf317a0a34bb25aea5f4895e5a35ecbac860931f5a29abc5be03bd3b.

Authorization boundary

Only the verified final manifest enables counter_tooling_construction_authorized=true. Counter execution, endpoint or service startup, model generation, payload creation, network access, reboot, record claims, and LocalMaxxing submission remain false.

The next permitted action is to construct, CPU-test, freeze, and independently audit a fresh cold-counter campaign. Counter execution still requires its own separate packet-only authorization child. The external USB remains backup-only.

This site is open source. Improve this page.