LFM2.5-350M review validation — 2026-10-05

This bundle records correctness and integration checks of source commit 3e2cbcf9f6e6f2db05642739efd46f84b9db4d29 against upstream main f594d7dfc4c2bef812e23f7ed73573be9625b287. The tested source archive SHA-256 is f32bc518e8e60e5fc66c67087d2234d7fed34a46fccfe865e6de32a235784b68. Both successful runs recorded 215 source files in their hash manifests. Provenance gives job states, pins, and repository-relative source hashes. These checks do not add performance measurements to the historical results.

CPU retry job 184536 completed 0:0 on cpu-2 in 31 seconds. The LFM suite had 32 passed; the Rust workspace had 54 passed, 9 ignored, with three pinned fixture tests and six native CUDA kernel tests ignored; the explicit frontend command had 12 passed (included in the workspace total). Rust fmt, Clippy and release build passed. Strict MkDocs build, benchmark fixture validation and 8 benchmark unit tests passed. The raw batch output is review-cpu-184536.out; cpu-checks.tar.gz losslessly wraps the downloaded cpu-184536.tar, which contains 18 check logs, environment/pin records and hash records.

GPU job 184541 completed 0:0 on one NVIDIA L40S (driver 580.178.04) in 54 seconds, using FP16 and pinned LiquidAI/LFM2.5-350M weights. verify.json.gz preserves the full verifier JSON: 17 cases × 6 batch sizes = 102 score/token-ID/selection checks passed against the uncached oracle; maximum absolute score error 0.1276397705078125 was below the empirical 0.15 gate. All six attention and ten convolution cache layers passed isolation and distribution checks. Four near-tie probes passed at all six sizes (24 checks), with minimum oracle margin 0.0462646484375 and no selection flips. The multi-question API check passed independent answers and additive, unpadded input usage (193 + 162 = 355). The direct worker and Rust frontend matched in 10/10 cases, including health, valid responses and 400/422 errors; the full byte-parity report is frontend.json.gz.

Initial CPU job 184529 was FAILED 1:0 even though its individual checks printed PASS and the binary hash was written. Its batch output and Rust step log are retained. The corrected retry used a non-login shell wrapper and completed; the available logs do not uniquely prove which exit hook or container step caused the first failure. No source code or test gate was changed for this retry.

To reproduce, use the LFM recipe for the pinned environment, weights, reference checkout and loopback worker/frontend setup. From the repository root at the tested commit, the CPU checks are python -m pytest -q tests/lfm2, cargo fmt --all --check, cargo clippy --workspace --locked --all-targets -- -D warnings, cargo test -p omni-jev --test frontend --locked, cargo test --workspace --locked, and cargo build --workspace --release --locked. For the docs and benchmark fixture checks, use a separate tools environment so the pinned LFM environment stays intact:

python3.12 -m venv .venv-checks
.venv-checks/bin/python -m pip install -r docs/requirements.txt -r benchmarks/requirements.txt
.venv-checks/bin/python -m mkdocs build --strict
.venv-checks/bin/python benchmarks/bench.py validate benchmarks/smoke.jsonl
.venv-checks/bin/python -m unittest discover -s tests/benchmarks -p 'test_*.py' -v

On a Slurm GPU node, run python recipe/lfm2/verify.py --reference-dir <pinned-reference> --device cuda --output <verify.json>; launch the worker with --candidate-batch-size 8 and compare its loopback endpoint with the frontend using recipe/lfm2/compare_with_frontend.py. The exact run settings and source hashes are in the reports and provenance file.

All three .gz files have gzip mtime zero and decompress to the exact downloaded source bytes. Run sha256sum -c SHA256SUMS here to verify this bundle. The score tolerance is a gate for these cases, not a uniform FP16 guarantee. This run did not repeat latency or memory benchmarks.