tinker_cookbook.completers.TinkerTokenCompleter

class tinker_cookbook.completers.TinkerTokenCompleter(TokenCompleter)

Token completer that uses a tinker.SamplingClient to sample actions.

completer = TinkerTokenCompleter(sampling_client, max_tokens=512)
result = await completer(model_input, stop=["<|endoftext|>"])
print(result.tokens, result.logprobs)

Fields:

• sampling_client (tinker.SamplingClient)
• max_tokens (int)
• temperature (float, default: 1.0)
• context_window (int | None, default: None)

__call__(model_input, stop, max_tokens)

Sample an action from the policy given an observation.

A per-call max_tokens override is combined with the completer's own max_tokens by taking the minimum, and the context_window cap (when configured) still applies on top.

Parameters:

• model_input (tinker.ModelInput)
• stop (StopCondition)
• max_tokens (int | None)

Returns: TokensWithLogprobs