b70-optimization-lab

Laguna separate shared gate+up native M=8 MM Stage-0 pass

Date: 2026-07-23 America/Toronto / 2026-07-24 UTC

Outcome

The first and only authorized tensor-bearing Stage-0 campaign for the separate shared gate+up native-M8 treatment passed its frozen one-card exactness protocol. The production analyzer returned exactly stage0-valid.

This is narrow component-preparation evidence. It proves two separate, role-bound native BF16 M=8 matrix multiplications, in gate-then-up order, through the actual checkpoint-selected LagunaMLP.forward path on physical card 0. It is not a merged gate/up operation, timing result, four-card component result, model generation, endpoint result, record claim, or LocalMaxxing submission authorization.

The pass authorizes construction, CPU validation, source freezing, and independent audit of pair-specific four-card component tooling only. Component execution, timing, other-card work, counters, model generation, network use, payload construction, reboot, and submission remain unauthorized until a separate packet-only commit freezes the new toolchain.

Frozen lineage

• runtime-guard reaffirmation: experiments/laguna-s-2.1-xpu-b70/notes/2026-07-23-shared-gate-up-runtime-guard-fix-and-reaffirmation.md
• reaffirmation SHA-256: 20ec56aeec8f8f0a1843cfa430ca118e7be9c6f97977e6b8b80d2beba45e194f
• Stage-0 tooling commit: 79577851f76f078d3150a8300bad670670b4d48c
• authorization-only child commit: 8bb2af9ef2657aa17687bf323f310a2efaf6c902
• authorization packet: data/laguna-s-2.1-shared-gate-up-m8-stage0-authorization.json
• authorization raw-file SHA-256: 550ca03817f81d74233e3a89a874e9ef36e0f81a64fdd345f78265b9a90ff00e
• authorization canonical-JSON digest without the tracked newline: 670cd0c7212708c8a7974a9f37e60dc09b67ca475af05eee039d029551f2bacd
• vLLM commit: 503f7784cf9d1704109b1e4650427fb4f417d604
• XPU-kernel commit: c59aaadbbfd350c2b5f4ad663e247c2811ae3181

The authorization commit changed exactly the packet and has the tooling commit as its parent. Main, vLLM, and kernel worktrees were clean for execution and production analysis. The boot remained 0b7f98a5-e50a-46a5-81ea-15938b55317a.

Fixture and authorized command

The canonical fixture is on internal NVMe:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/authorizations/shared-gate-up-m8-stage0-fixture-v1-79577851f.json

• file SHA-256: c47d3f7e0fef29dd483fb1397ca4d36a93a923c7d6a9af18859023d98de7618c
• manifest SHA-256: 87e40cc2367ebc44df5620f27861a0b2d87d2c5a3a8bbd28e28a51784e7b5300
• ordered-epoch-hashes SHA-256: e907c35d4726a8178abfa40c98bfe369ad4704c9a4540ff43a9e6458c377c851

The exact authorized command was:

/home/steve/llm-optimizations/experiments/laguna-s-2.1-xpu-b70/tools/run_laguna_shared_gate_up_mm_stage0.sh \
  --authorization /home/steve/llm-optimizations/data/laguna-s-2.1-shared-gate-up-m8-stage0-authorization.json \
  --fixture /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/authorizations/shared-gate-up-m8-stage0-fixture-v1-79577851f.json \
  --output-root /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-up-m8-stage0-card0-79577851f-v1

It exited zero. The result began at 2026-07-24T04:41:46.514405Z and completed at 2026-07-24T04:42:16.864499Z.

Sealed evidence

Campaign root:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-up-m8-stage0-card0-79577851f-v1

• stage0-result.json: 726bfa5b19dd6e8722aedda0679603d9cc43214acb1952b07c15dc9562b558b0
• pre-tensor-identity-checkpoint.json: 335ccb1f1663a96ac95ab499986fef53319f77869d9eb5a65b53e3d44aad8b4c
• tensor-work-started-checkpoint.json: d1db2b1724bcdc7b8db6f3aad49372ea44da3412ee6ac66103e8895ee482ca62
• runtime-card0-binding-checkpoint.json: 8589a4bf0747075b7c51c7127bdf8225ec6ffed3df865f23fc4919cbee819d22
• constructor-scope-proof.json: 27a1398ef501a9e2f5dc643f2732aae62f51c175ef3f5e7b47ebe3eaa78706e4
• dispatch-proof.json: 0f76865d0c38a06ae521671c47acf93dd52c6062b79a238907beaa6ec7537f0a

The terminal result is stage0_exactness_pass, with all 128 epochs durable. All 1,152 declared comparisons passed both raw little-endian BF16/uint16 equality and torch.equal. Every one of the 16 output labels had 128 unique raw and canonical hashes. Inputs and all three projection weights remained immutable across host copy, transfer, and execution.

The exact boundaries were:

• gate BMM control versus native-MM candidate and candidate repeat;
• up BMM control versus native-MM candidate and candidate repeat;
• gate SiLU;
• BF16 SiLU-gate times up;
• shared down projection;
• shared-plus-routed add; and
• three-peer sequential fixed-rank reduction.

The actual-forward dispatch proof recorded exactly two native MMs, one for gate and one for up, in that order. Across the full proof it recorded 2 MM, 22 incumbent BMM, and 0 fallback calls. The 22 incumbent cases cover unmarked M8, marked M1-M7, prefill, dense, draft, routed, and shared-down paths. All 30 named layout, marker, pair-scope, peer, selector, and record-stack corruptions raised the frozen RuntimeError before any MM, BMM, or fallback primitive.

The constructor proof binds only:

• model.layers.1.mlp.shared_expert.gate_proj; and
• model.layers.1.mlp.shared_expert.up_proj.

Shared down, dense, draft, and routed representatives remain unmarked. The runtime bound one visible XPU to physical card 0, DRM /dev/dri/card3, BDF 0000:23:00.0, and the frozen B70 UUID.

Production validation

The independent production command was:

PYTHONDONTWRITEBYTECODE=1 \
  /home/steve/.venvs/deepseek-v4-xpu/bin/python \
  experiments/laguna-s-2.1-xpu-b70/tools/analyze_laguna_shared_gate_up_mm_stage0.py \
  --fixture /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/authorizations/shared-gate-up-m8-stage0-fixture-v1-79577851f.json \
  --authorization /home/steve/llm-optimizations/data/laguna-s-2.1-shared-gate-up-m8-stage0-authorization.json \
  --result /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/shared-gate-up-m8-stage0-card0-79577851f-v1/stage0-result.json

It returned exactly stage0-valid. Before authorization, independent audits also caught and closed two evidence weaknesses: dispatch records now bind both input and output width, including shared down 256 -> 3072, and an identical raw tensor with a forged torch_equal=false can no longer be classified as a proven raw mismatch. The final CPU-only tool suite passed 85/85.

Next boundary

Construct new pair-specific four-card component tools from the corrected shared-gate component structure. The frozen treatment remains two separate gate-then-up MMs—no N=512 merge, logical-B16 packing, fusion, reordering, overlap, or shared-down treatment.

Each card must independently pass 128 pre-timing and 32 post-timing changing exactness epochs, at least 28 of 31 A-B-B-A blocks, and at least 0.20 ms median saving per complete 47-layer gate+up cycle. The component toolchain must be CPU-tested, independently audited, committed, and followed by its own authorization-only child commit before any component execution.

This site is open source. Improve this page.