Deterministic test backend: no NIF, no GGUF, no hardware.

Load it with backend => erllama_model_stub to exercise your own code against the full erllama API (completion, streaming, sessions, cache tiers) in unit tests:

{ok, M} = erllama:load_model(#{backend => erllama_model_stub}),
{ok, #{reply := Reply}} = erllama:complete(M, <<"hello world">>, #{response_tokens => 4}).

Tokenisation hashes whitespace-delimited words (erlang:phash2/1), generation derives the next token from the context hash, so the same prompt always yields the same tokens; kv_pack/kv_unpack serialise the token list, so cache saves and restores work end to end. Chat templates, embeddings-as-vectors and adapters behave as on a model without those features ({error, not_supported} / {error, chat_not_supported}).

Load-config keys specific to this backend: step_delay_ms (hold each decode step that long, to make queueing observable) and thinking_capable (emit thinking events).