completers

Guide for using completers — TokenCompleter and MessageCompleter for text generation during RL rollouts and evaluation. Use when the user asks about generating text, completing messages, or using comp

SKILL.md

GitHub

Completers

Completers wrap SamplingClient for convenient text generation. Two levels of abstraction:

  • TokenCompleter — low-level, returns tokens + logprobs
  • MessageCompleter — high-level, returns parsed Message objects

Reference

Read these for details:

  • tinker_cookbook/completers.py — Implementation
  • docs/completers.mdx — Usage guide

TokenCompleter

Generates tokens from a ModelInput prompt. Used internally by RL rollouts.

from tinker_cookbook.completers import TinkerTokenCompleter, TokensWithLogprobs

completer = TinkerTokenCompleter(
    sampling_client=sc,
    max_tokens=256,
    temperature=1.0,
)

result: TokensWithLogprobs = await completer(
    model_input=prompt,
    stop=stop_sequences,  # list[str] or list[int]
)
# result.tokens: list[int]
# result.maybe_logprobs: list[float] | None

MessageCompleter

Higher-level: takes a conversation (list of Messages), returns a Message. Handles rendering and parsing internally.

from tinker_cookbook.completers import TinkerMessageCompleter

completer = TinkerMessageCompleter(
    sampling_client=sc,
    renderer=renderer,
    max_tokens=256,
    temperature=1.0,
    stop_condition=None,  # Override stop sequences
)

response_message: Message = await completer(messages=[
    {"role": "user", "content": "What is 2+2?"},
])
# response_message = {"role": "assistant", "content": "4"}

When to use which

  • TokenCompleter: RL rollouts, custom generation loops where you need logprobs and token-level control
  • MessageCompleter: Evaluation, tool-use environments, multi-turn RL where you work with Messages

Custom completers

Both are abstract base classes you can subclass for non-Tinker backends:

from tinker_cookbook.completers import TokenCompleter, MessageCompleter

class MyTokenCompleter(TokenCompleter):
    async def __call__(self, model_input, stop) -> TokensWithLogprobs:
        ...

class MyMessageCompleter(MessageCompleter):
    async def __call__(self, messages) -> Message:
        ...

Common pitfalls

  • Create a new completer (with a new SamplingClient) after saving weights
  • TokensWithLogprobs.maybe_logprobs can be None if logprobs weren't requested
  • MessageCompleter uses the renderer for both prompt construction and response parsing