b70-optimization-lab

Laguna S 2.1 persistent exact-attention metadata record candidate

Date: 2026-07-25 America/Toronto

Status: VERIFIED, INDEPENDENTLY AUDITED, SUBMITTED, APPROVED RECORD. LocalMaxxing ID cmrzrd4tf001ipa013xpx4kid. The conservative lower candidate start is 94.92003934159611 tok/s, compared with prior approved record cmrzjb7i906x4o401egrnm05m at 92.16352215694299 tok/s.

Result

The preregistered fresh-service graph-vs-graph A1/B1/B2/A2 campaign passed every quality, honesty, causal, reproducibility, practical-floor, and record gate:

Leg Treatment Median tok/s, tokens 1-100 after TTFT Canonical exact Cached tokens
A1 graph, metadata off 92.54961760665958 13/13 0 on 13/13
B1 graph, metadata on 94.92003934159611 13/13 0 on 13/13
B2 graph, metadata on 95.06654837534518 13/13 0 on 13/13
A2 graph, metadata off 92.87797142677606 13/13 0 on 13/13

The eligible value is the lower B start. It improves the prior approved record by 2.75651718465312 tok/s (+2.990898264455555%) and strictly exceeds the preregistered 92.393930962335 tok/s practical floor.

Both adjacent comparisons passed:

Pair Candidate row wins Headline gain Median paired gain Cycle saving Acceptance drift
B1 vs A1 13/13 +2.5612442236238477% +2.3511088081138385% 0.9112340643584034 ms 0.0003077489561834623
B2 vs A2 13/13 +2.3564004628315667% +2.5609196498322975% 1.6476436413154687 ms 0.0003070338847266929

All 52 leg outputs match the canonical q1 greedy teacher bitwise. The cross-leg bundle is 39/39 exact, every cached-token count is zero, and long-then-next and the 863-token rollover row pass on every leg. Metrics began at zero and each fresh service handled exactly 13 unique prompts once.

What changed

All four starts use the approved exact DFlash7, fused W1-route-W2, route-interleaved expert GEMM, shared-elementwise, QKNorm/RoPE, W1 N64, and Breakable PIECEWISE graph stack. The sole treatment difference is:

A control: VLLM_XPU_LAGUNA_M8_PREBUILT_EXACT_ATTN_METADATA=0
B candidate: VLLM_XPU_LAGUNA_M8_PREBUILT_EXACT_ATTN_METADATA=1

The candidate moves exact q2..q8 attention metadata into builder-owned, fixed-address XPU buffers. Query offsets, growing KV lengths, and expanded block tables are refreshed in place on the current stream. Metadata object, base pointer, offset, active-view, and owner signatures are guarded and fail closed on drift. VLLM_XPU_LAGUNA_M8_CAPTURE_ATTENTION_GRAPHS=0 throughout; the attention kernels remain eager boundaries inside the same 146-graph, 145-boundary topology.

Every start performed its first lazy graph capture inside the first measured cold request and logged one audited capture and replay on ranks 0 through 3. There was no warm-up generation, retry, prefix cache, history/ngram acceleration, response reuse, context checkpoint, qualification timing, or prior-run output.

Frozen identity

Models, private cache/RPC roots, and evidence all used internal EXT4 NVMe.

Sealed evidence

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/laguna-m8-metadata-formal-ea70d34fe-ef334233d-20260725T015736Z

Key SHA256:

1f4334ccdf45e3c17e68366c4e24ff95adc022809d60986146c10d00cdd460f9  full-analysis.json
6b4c1aebfdc7bcd5779d640e4276f9dec33abcc54d3a8b958c8f19f443d7e332  all-vs-teacher.json
1a3a85587d2409a5ee04740ce78504796766baecbc5e2e0036eda751e0be2fa9  cross-leg.json
5d7e9e49faac7046fdff4a6d83dc8ea7b4c6990ebb67d76d50cb42e98fdde6e2  B1-graph-metadata-on/bench.json
04e18dde23f53ebf9eeee0a3760bf7daeb3d0a20bc1f119fa2f43bc013bdde56  B1-graph-metadata-on/exactness-vs-q1.json
3f26de97073a3283b50c7b762e8124c70434308be96faaa035bf52bdc64b85ac  B1-graph-metadata-on/server.log

The committed analyzer emitted record_candidate. A separate read-only agent recomputed source/tool identities, all four leg gates, 52/52 canonical exactness, 39/39 cross-leg exactness, raw performance statistics, paired metrics, topology, idle intervals, cleanup, sealing, treatment isolation, and absence of retry/rescue roots, then approved submission of the lower B start.

For a read-only analyzer rerun, use the repository root as the current working directory. The controller recorded its self-path relative to that intended launch directory; this is a rerun hardening issue, not an evidence discrepancy in the sealed campaign.

Decision

The independently audited payload submitted only 94.92003934159611 tok/s, never the faster B2 value. LocalMaxxing returned HTTP 201 and approved it as cmrzrd4tf001ipa013xpx4kid. Preserve all four starts as the reproducibility record.

This site is open source. Improve this page.