Release the whole adapter dependency tuple

Record immutable base weights/configuration, adapter tensors and configuration, target module names, rank/scaling, tokenizer files, chat template, added tokens, and relevant library/runtime versions. PEFT 0.17.0 documents an adapter checkpoint as adapter state plus configuration rather than a complete frozen base. A base with identical dimensions but different weights changes the adapter's reference point. A tokenizer with the same vocabulary count can still change IDs and stopping behavior, so shape compatibility is only one validation layer.

If training also changes embeddings, biases, or other saved modules, include them in the release and rollback contract. Quantization settings and compute dtype belong with the final inference path. A merged full checkpoint is a different artifact from a base-plus-adapter pair and may have irreversible rounding changes. Keep source hashes and the exact merge recipe so provenance and reconstruction do not depend on mutable repository names or an operator remembering which adapter was active.

Test fixed prompts, token IDs, stop conditions, task metrics, retention, severe failures, and resource/latency behavior after loading the actual release. Compare separate-adapter and merged execution under the intended dtype and quantization, accepting only documented numerical differences. Canary traffic needs compatible cohorts and a defined stop condition. A training checkpoint's validation score does not cover a deployment that silently uses another template or truncation policy;those changes need their own evidence.

Shared serving adds state isolation. Routing must select the intended adapter and base for each request, and any prefix or key/value cache reuse must satisfy the implementation's model-state compatibility rules. Cache entries produced under one adapter are not automatically valid under another simply because prompt text matches. Include model/adapter revisions and other required state in cache identity or invalidate appropriately. Test concurrent tenants and in-flight requests during switches rather than checking only a quiet single-request load.

A rollback can route new requests to an immutable prior release while allowing or terminating old in-flight work under an explicit policy. Clear or segregate incompatible caches and preserve the old tuple long enough for controlled recovery. Disabling a pure adapter returns the intended base path only if no other trained modules or artifacts changed. After a quantized merge, loading the original base is more reliable than assuming an inverse subtraction reproduces its original bytes.

The cache lab is only a version-identity contrast. Reset and set both C=4, s=4, predict twelve hits among sixteen requests, then Step to add one hit. Its empty-start fully associative LRU stream caps at 4096 requests and has no model weights, KV state, tenants, or release routing. Use the contrast to explain that a cache hit is correct only under a valid identity contract;this toy's matching address cannot certify that a real prompt cache belongs to the newly selected adapter.