b70-optimization-lab

Laguna S 2.1 INT4 on four Intel Arc Pro B70s

Qualified result

The sealed width-12 / DFlash-depth-11 run is exact and approved by LocalMaxxing:

The historical helper counted 100 timestamped token events over a span containing 99 intervals. The approved row is preserved as the receipt of what was submitted; it must be described as the published legacy convention. Under conventional interval accounting, the 102 tok/s threshold was missed by 0.05827875982973 tok/s.

Identity

field value
target poolside/Laguna-S-2.1-INT4 at 4bbfc285f2f8b3b6b526274c133b7b17aae6c8cb
draft poolside/Laguna-S-2.1-DFlash-INT4 at 5e07c246915c86dc6920fead03d019989224f2ba
vLLM e596ef1543466ae1a05e5bb8091f58872e2b18ba
XPU kernels 6f9dd3c3a7b1b677a992ca4f431a968408f9c816
hardware 4x Intel Arc Pro B70 32 GB, TP4+EP4, one active generation
verifier / draft exact width 12 / DFlash depth 11
KV BF16
treatment 31 E4M3FN W8A16 draft projections per rank plus the exact auxiliary workspace

The FP8 treatment applies to disposable draft projections. It does not mean FP8 KV, and no gain is attributed to the intended separate draft LM-head path because its runtime preparation marker was absent.

Evidence and reproduction

The immutable queue and HTTP 201/public-verification receipt remain under data/. The submission is already approved; do not POST a duplicate.

The standalone packet now includes the sealed raw run, a release-only 32-file model manifest and download-at-revision helper, complete observed package and host identity, three kernel-source provenance points in addition to the final kernel tree, and checks for every direct and transitively loaded native library. Artifact-exact replay and source-equivalent rebuild are deliberately separate claims; rebuilt binaries require the full gate and are not silently substituted into the sealed environment.