A multimodal conversation is an ordered media-and-token contract

Versioned Transformers 4.57.1 multimodal chat documentation describes message content as a list of typed blocks rather than the plain content string used by its text-only examples. Its processor can combine preprocessing, tokenization, and chat templating. This is a documented API family, not a universal format for every model service. A specific checkpoint can have different supported roles, separators, media markers, or image-count behavior. Preserve the actual template and processor artifacts with the model, and verify their outputs instead of hand-building a plausible-looking prompt from an unrelated example.

Assign stable identities to media and retain their order through loading, decoding, preprocessing, and feature expansion. If asynchronous fetching completes in another order, restore the declared conversation mapping before merging features. Repeated references and shared cached payloads need explicit ownership semantics. Count and validate media using the selected contract rather than relying on total tensor size. The LLaVA documentation for this pinned version also warns that technical multi-image input support does not establish that a checkpoint was trained to handle several images well; interface acceptance and task capability remain separate questions.

A rendered image sentinel may expand into many model positions, while other architectures keep visual memory separate. For a concatenated interface, compare the processor's expected inserted positions with the connector's retained feature counts. Truncation, an incompatible feature-selection rule, or a missing payload can break this relation. Do not repair it by silently repeating or dropping arbitrary features. Reject or apply a documented supported fallback before the model consumes a prompt with a different media meaning than the user supplied.

During autoregressive instruction training, the loss mask must match the intended target tokens and role boundaries. Original LLaVA training illustrates supervising assistant responses under its particular conversation format; another objective may supervise a different set. Padding masks and target-ignore conventions are separate from causal and media-ownership masks. Test the final token IDs, labels, attention mask, feature lengths, and association together. A batch with valid individual examples can still fail if left/right padding or token offsets shift the merge or target positions differently.

Use two deliberately distinguishable media objects and questions that require knowing which one is referenced. Test single examples, batched examples with different lengths, missing files, repeated references, history truncation, and round-trip inspection of the rendered prompt. Save expected identities and compatible shapes as a fixture for the chosen library and checkpoint revision. Test generation start and stop boundaries independently from training labels. A fluent answer is a weak serialization check because the language model can produce one even when it ignored or misattached the evidence.

The modular lab shows a known ordered sequence, not a conversation parser. Reset, set both m=10 and additive a=3, predict residues three and six after one and two Steps, then inspect. Reset and set both m=10 and a=5 for a period-two orbit returning to zero after two Steps. Each Step advances one addition per 0.5 seconds. Its finite recurring labels carry no media, roles, tokenizer, template, feature expansion, or target mask. It only illustrates why a declared order must be followed and why a small recurring number is not a complete identity contract.