Skip to content

quantization-artifacts.md

Latest commit

 

History

History
58 lines (46 loc) · 3.06 KB

quantization-artifacts.md

File metadata and controls

58 lines (46 loc) · 3.06 KB
Raw

Quantization artifact retention

The repository's historical benchmark notes describe experiments, not a guarantee that every temporary binary remains in the checkout. Cleanup consolidates local evidence under ignored local-evidence/quantization-2026-09-09/; no model, dataset or generated result is committed by this operation.

Historical cleanup counts and validation are recorded in the original report. They do not establish current artifact availability or test status. Verify the local inventory before reuse; this guide owns the restore procedure.

Retained locally

  • reports-inputs.tar.gz: reports, metrics/predictions, prepared NPZ inputs, experiment scripts/configurations, logs, traces and small synthetic image files. Original tmp-* relative paths are preserved inside the archive.
  • retained/: final last.pt models for the flat/hierarchical three-seed studies and the longer hierarchical pair; floating/materialized and calibrated ONNX deployment bundles for both representative heads. Their original relative directory structure is preserved.
  • inventory.json: archived/retained paths, sizes and SHA-256 digests, plus the discarded-file inventory and runtime-directory totals. Archive content and retained models are verified before their temporary source directories are removed.
  • README.md: local cleanup totals and restore instructions.

Prepare optional GPU runtimes explicitly on the target using the inference guide; the archive does not include a portable working environment.

Removed as disposable

Intermediate and duplicate checkpoint/model exports, synthetic large-head binary models, laptop-specific TensorRT engines, temporary package installations, and completed pytest outputs. Final optimizer-resume snapshots are not retained in this compact handoff: exact continuation from discarded experiments requires regenerating the run. Retained final model weights support inference/quality re-evaluation, not a claim of complete optimizer/RNG restoration.

Restore selected evidence

Run from the repository root. List the archive before choosing paths:

tar -tzf local-evidence/quantization-2026-09-09/reports-inputs.tar.gz
tar -xzf local-evidence/quantization-2026-09-09/reports-inputs.tar.gz \
  tmp-cpu-deployment-flat tmp-cpu-deployment-hierarchical

This restores reports and their child evidence for archival/review without engines or model weights. To rerun inference, restore the corresponding inputs and copy the selected bundle from retained/ to the intended location. Rebuild TensorRT engines on the target. Original reports keep original provenance paths and hashes; do not rewrite them to make discarded binaries appear available.

The archive is a local handoff, not a remote backup or a live result-hosting service. Transfer it deliberately with the required retained models when target access is available. Public compact history contains only the selected reporting projection.