cesnimda
4.5 KiB
Model/cache qualification architecture and fixture identity
Source baselinef17dba2,25October2026. Actual Python consumers are AutoTokenizer and AutoModelForSeq2SeqLM for fixed BART summarisation in tools/summarizer/app.py:96; they initialise lazily, or eagerly only under the existing environment flag. Document parsing uses pypdf/PyMuPDF/python-docx/Pillow and resource-bounded child processes. OCR uses the native Tesseract executable and its English traineddata, not a Transformers OCR model. No Python embedding/reranking/Docling model is implemented in this source. Ollama is an HTTP client boundary with its own separately managed model storage; no Ollama model/cache/server was accessed.
The selected full repository sshleifer/distilbart-cnn-12-6 advertises Apache-2.0 and immutable revision a4f8f3ea906ed274767e9906dbaede7531d660ff. Its1,222,317,369-byte PyTorch checkpoint has authoritative LFS SHA256 3bac65d18c99463302d12ca75c2220ea714f9c81ce235f205fa818efe71df6ea. It was not available in the checked local user cache or existing local sidecar image. Local Docker context is desktop-linux; the only volume is the distinct MariaDB development volume. No production filesystem/cache was inspected. Old image contents are not approved current-runtime artifacts.
Small task-authorised fixtures
ARTIFACTS.json records two independently hashed, isolated fixtures. Their tokenizer/vocabulary/merges/model-card and reference configuration come from the immutable upstream revision. Every downloaded ordinary Git file is checked against the published Git blob identity, then pinned with SHA256. Only1,358,673 bytes of upstream metadata/tokenizer content were downloaded. No pretrained checkpoint or remote Python code was downloaded.
Synthetic BART weights are generated locally with fixed seed73421 using the Apache-2.0 Transformers BART implementation: vocabulary50265, dimension16, one encoder/decoder layer, two heads,32-wide feed-forward layers and1024 positions. The842704 parameters are saved once as Safetensors and once as a known state dictionary. These files are approximately3.58MB each. The second format exercises the same PyTorch serialization family as the selected full checkpoint, with explicit weights_only=True. No third-party pickle or trained-weight licence is assumed. The actual tokenizer's Apache-2.0 declaration is preserved from its pinned card; no model weights are redistributed in Git.
These are real native BART/tokenizer/generation artifacts, with deliberately untrained/random outputs. They establish the tested library/API/format/CPU compatibility scope. They do not establish full DistilBART output quality, production memory/latency, GPU compatibility, intended model semantics or C11/P5-13 acceptance. No owner approval of a production model or formal risk exception is implied by task-authorised test preparation.
The committed manifest is the integrity authority, not an arbitrary cache-local JSON file. The verifier rejects missing/corrupt/linked/duplicate/traversal/unlisted artifacts before loading. The fixture mount is read-only; runtime networking is none, with explicit Hub/Transformers offline flags, private tmpfs cache and twoCPU/2GiB/128PID limits. Each container has an owned label, bounded parent deadline and removal verification. Replacing source/checkpoint identities requires a new manifest and qualification.
The production model choice is unchanged. A minimal loader repair now pins tokenizer and weights to the same reviewed revision, denies remote code and explicitly requests restricted weights loading. Its new unit test failed for the missing revision before the repair; the complete native Windows suite passed59 cases afterward. That suite alone is not real-model evidence; Linux fixture inference is separate.
Cache ownership and supported transition
filelock4 changes shared-cache lock protocols. All qualification caches are new isolated disposable directories with a single library generation. No production or developer cache was reset, rewritten or silently mixed with old lock participants. A later production rollout must stop old participants and explicitly select a separately prepared/verified cache or establish documented protocol migration compatibility. That operator action is separate from development fixture compatibility and remains withheld.
Actual compatibility, negative/offline/cache contention/OCR results and any limitations are recorded separately as they execute. The full checkpoint remains absent; no large download was used to manufacture a qualification claim.