b70-optimization-lab

Laguna S 2.1 on 4x B70: published 102.971 / conventional 101.942 tok/s

This is the fail-closed reproduction packet for the approved Laguna S 2.1 INT4 four-B70 result. It embeds the sealed raw evidence, restores every known source and model revision, verifies actual native loader origins, runs one cold 13-prompt suite, and rejects token, text, cache, treatment, graph, teardown, or material performance drift.

Result and metric qualification

field value
LocalMaxxing/submitted historical convention 102.97143559613157 tok/s
Conventional first-to-100th-token interval rate 101.94172124017027 tok/s
Canonical-q1 equality 13/13 token IDs and output-text SHA-256
Cache state cached_tokens=0 on 13/13 rows
LocalMaxxing cms2ccv2d00lps201rej94pjy (APPROVED)

The historical helper divided 100 timestamped events by a span containing 99 inter-token intervals. The public value is an accurate receipt of that submitted convention, but it is not a conventional 102 tok/s result. See the accounting correction.

No duplicate LocalMaxxing submission is needed or allowed for this packet.

Exact identity

The FP8 label refers to draft projection weights, not KV. The record deliberately uses BF16 KV to preserve its canonical-teacher contract.

Verify the historical packet

This performs no GPU work and no network submission:

cd /home/steve/llm-optimizations
repro/laguna-s-2.1-int4-b70-102tps-20260726/verify-record.sh

The sealed raw benchmark, exactness report, server log, identity, service environment, metrics, cleanup status, and all pre/post idle snapshots are tracked below evidence/record-run/. Verification therefore fails if evidence is absent; it never degrades to a packet-only PASS. The command recomputes both throughput conventions and compares all 13 token streams and output-text hashes against the compact tracked oracles.

Restore source

The portable restore command fetches public prerequisites, imports all three tracked bundles, and creates five clean worktrees covering the mixed native provenance:

repro/laguna-s-2.1-int4-b70-102tps-20260726/restore-sources.sh \
  /mnt/fast-ai/laguna-repro-sources

The source bundles and reviewable combined patches are indexed in the snapshot README. Build and provenance details are in BUILD.md.

Restore or verify the two model payloads at their immutable Hugging Face revisions:

repro/laguna-s-2.1-int4-b70-102tps-20260726/restore-models.sh \
  --download /mnt/fast-ai/llm-models/laguna-s-2.1

The tracked manifest contains exactly 32 release files and excludes .cache locks, metadata, and incomplete downloads.

Runtime and model prerequisites

The exact lab replay expects:

The launcher pins and hashes the Python/vLLM entry points, oneCCL, SYCL, libtorch_xpu, four extension modules, six transitively loaded helper DSOs, xpumem_allocator, model configs, the release payload and original manifests, source trees, historical benchmark, interval-accounting qualifier, comparator, suite, and both compact teacher oracles. It also checks exact package versions, four B70 PCI identities, OS/kernel identity, cluster interface, model contents, and actual module and /proc/self/maps loader origins. Override REPRO_VLLM_TREE, REPRO_KERNEL_TREE, REPRO_VENV_ROOT, REPRO_XPUMEM_MODULE, or REPRO_CLUSTER_IP only when the same byte-identical artifacts live elsewhere.

The original single monolithic build invocation was not sealed. The packet therefore distinguishes a portable sealed-evidence audit, an artifact-exact originating-host replay, and a source-equivalent rebuild. A fresh rebuild that does not reproduce the pinned native binary hashes is a new environment and must pass the complete gate; do not weaken or replace the sealed hashes to make it look artifact-exact.

Preflight and run

Complete preflight performs no model launch. It hashes the full model payload, so it is intentionally slower than a shallow syntax check:

repro/laguna-s-2.1-int4-b70-102tps-20260726/run.sh --preflight

Run exactly one cold suite:

repro/laguna-s-2.1-int4-b70-102tps-20260726/run.sh --run

The policy is one start, one suite, and the first valid score. Do not retry to select a faster start, warm the model with a generation, omit prompts, enable prefix/history reuse, or move graph/setup work outside the measured contract.

Validate the resulting directory:

repro/laguna-s-2.1-int4-b70-102tps-20260726/verify-run.sh \
  /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/laguna-width12-dflash-fp8-repro-YYYYMMDDTHHMMSSZ

The validator reports both rate conventions. Identity, 13/13 token-and-text equality, cache-zero state, four treatment markers, 146/145 topology, service environment, and cleanup/idle evidence may not drift. To distinguish “correct but much slower” from reproducing the performance result, it also requires a conventional suite median of at least 96.844635178 tok/s (95% of the sealed conventional median).

The post-packaging hardening has been validated with component checks of the sealed evidence, runtime loader mapping, source bundles, and model manifest. The integrated run.sh --preflight and a fresh source restore were not rerun after packaging: the former is intentionally left for a clean committed checkout, and the latter encountered a pathological local partial-clone fetch loop. No additional score-bearing cold suite was consumed; the tracked scored evidence remains the sealed first-valid historical run.

Evidence

This site is open source. Improve this page.