Files
jobtrackingapp/docs/redesign/phase-9/p9-sec-02/full-model/PLAN.md
T
cesnimda b48f25c4ff
MariaDB persistence qualification / mariadb (push) Has been cancelled
test: qualify immutable models and Phase 9 regression
2026-10-25 13:43:50 +01:00

2.2 KiB

Actual configured checkpoint follow-up

The small immutable BART fixtures passed library, native CPU, serialization and API compatibility. They do not establish that the actual older configured DistilBART checkpoint/configuration loads under the major Transformers transition. This is the remaining reason for one exact pretrained download; no unrelated or larger generative model is introduced.

The user authorises necessary trusted, licensed, size-checked downloads into an approved nonproduction environment. Local disk has over240GiB free and host memory is about32GiB; qualification is bounded to fourCPU/eightGiB/128PIDs/300seconds with network none, read-only checkpoint and no production/Ollama access. The source is the Apache-2.0 repository at immutable revisiona4f8f3ea906ed274767e9906dbaede7531d660ff. Only the1,222,317,369-byte PyTorch checkpoint with SHA2563bac65d18c99463302d12ca75c2220ea714f9c81ce235f205fa818efe71df6ea is fetched; existing verified tokenizer/config bytes are reused.

The entire size and digest must match before any model load. A partial or invalid download remains failed evidence and is never loaded. The loader is the patched actual sidecar path, with explicit weights_only and no trusted remote code. The test starts only after ordinary, provider/PDF and exclusive15minute qualification has finished. It does not compete with measured workloads.

This is a new follow-up, not a rewrite of the earlier no-full-download checkpoints. Results remain PENDING until actual execution. Full-model compatibility does not establish semantic/C11 acceptance, GPU behaviour, live cache ownership/integrity or deployment readiness.

Execution25October2026: PASS. The exact checkpoint size/hash matched; the actual source loader used the default pinned repository revision through a private offline Hub snapshot. CPU model/config/tokenizer loading,1024-token processing, finite logits, generation, authentication and repeat-cache contract passed in27.72seconds with peakRSS2858928KiB. The nonroot, read-only, network-none container exited0 and positive daemon evidence verified removal. The preceding PENDING statement is the preserved pre-execution plan. validation.json records the actual successor, manifest hash and remaining scope limits.