b70-optimization-lab

Laguna M8 sharded gather: Stage-0 fixture and operational preflight

Date: 2026-07-24

Outcome

The host-only portion of Stage 0 is complete for the standalone M8 MoeGather sharded-occupancy lane. The committed operational preflight passed against the installed observer, and the fixed 288-epoch production fixture corpus was generated and independently re-read successfully from the internal NVMe.

This does not authorize a candidate primitive yet. Phase A and conditional Phase B still require their immutable packets to be built, corruption-tested, reviewed, and committed together. No candidate/native import, Torch XPU import, XPU tensor allocation, component primitive, model load, endpoint, generation, benchmark, payload, network request, submission, service change, or reboot occurred.

Corrections made before the official preflight

The first uncommitted mock-only parser incorrectly required process_name to be an absolute path. Independent review caught that the installed xpu-smi ps -j reports the basename xpu-smi. A separate read-only schema probe then confirmed the exact installed shape:

• top-level key device_util_by_proc_list;
• four self-observer rows, one for each device ID 0..3;
• exact row fields device_id, mem_size, process_id, process_name, and shared_mem_size;
• basename process_name: "xpu-smi".

The sanitized retained fixture is data/laguna-s-2.1-xpu-smi-ps-installed-schema-sanitized-20260724.json, SHA-256 f1ed4c4f41999a2f79ce77f8b7fb251d15d5f2d87ac6a62b83a1ddfddfaa40ef.

Further read-only audits found and corrected, before the official run:

• permissive self-row schema/cardinality checks;
• an observer executable CLI override;
• missing strict duplicate-key JSON rejection;
• missing raw failure capture;
• cleanup gaps for ordinary communication errors and BaseException at proc-directory, optional-pidfd, and communication boundaries;
• missing internal-NVMe/ext4/source/block-device attestation;
• incomplete fixture-metadata binding.

The final preflight is committed at 3f6eda32c406fca71aa406ad7f7f0adc613329de. Its 17 CPU-only tests, Ruff, AST parsing, whitespace check, and independent review pass. The installed observer is fixed to /usr/bin/xpu-smi, SHA-256 2b5b128edf28b38da8637413fe8bfe3a4a40e8113210ba9ddaed945bd56d826e. The JSON basename is treated only as schema data; executable identity comes from the retained Popen PID plus a live proc-directory descriptor and /proc/<pid>/exe, with a pidfd retained when the Python build exposes it.

Official operational result

The one committed operational invocation passed:

• artifact: /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/m8-gather-sharded-operational-preflight-20260724T104851Z/report.json;
• artifact SHA-256: 1c08e4e22fb24931258124f7dee3b31e5d117d715fae239c024e69ffb28b3649;
• exact result: status=passed;
• accepted mode: self_observer_rows;
• rows/device IDs: 4 / [0,1,2,3];
• every row PID matched the retained observer child;
• every row basename was xpu-smi;
• no foreign or mixed XPU process was present;
• backing storage: /dev/nvme0n1p2, ext4, 259:2, with an NVMe sysfs ancestry;
• report mode/ownership: 0600, steve:steve;
• report encoding: canonical JSON with raw stdout/stderr bytes retained as base64 plus SHA-256.

This was a read-only management query. It did not import or execute the candidate or any model/runtime path.

Production fixture

The fixture generator/analyzer is committed at 30b043b2b97223cf3305822512d2277da1f635a4. Its 16 CPU-only tests, Ruff, AST parsing, whitespace check, and two independent audits pass. Inputs are regular-file/no-follow and size bounded; the analyzer regenerates every expected tensor independently rather than trusting manifest hashes.

Production root:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/evidence/m8-gather-sharded-fixtures-30b043b2b-20260724T1050Z

The corpus is 205 MiB and contains exactly 288 epochs: 256 pre-timing plus 32 post-timing. Manifest and analysis identities:

• manifest.json SHA-256: b1bef1cfeb72502074c0408ab73fa8bdf6f30732862410360658284b74161bd0;
• analysis.json SHA-256: 04a897553820581f22f6d7ff62d72b08b6494f233caf6398308c667eeab6c8d4;
• analyzer status: passed;
• manifest hashes: matched;
• independently regenerated bytes: matched.

All frozen coverage proofs are true:

• all 65,536 raw BF16 patterns;
• exact FP32 edge inventory, including signed zero/subnormal/finite/infinity and signed payload-NaN cases;
• all 1,024 local/remote masks;
• all ten dedicated single-slot probes;
• all-local, all-remote-zero, and mixed rows;
• literal zero for every remote row;
• deterministic nonzero formula for local rows;
• signed-int32 canonical route map;
• ordered-cancellation and BF16-midpoint witnesses.

Next action

Build and independently corruption-test the exact Phase-A and conditional Phase-B runner/analyzer plus their mutually bound authorization packets. Commit both packets together before the first candidate/native import or device primitive. The candidate still receives one bounded campaign and no repair/retry after packet freeze.

This site is open source. Improve this page.