For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python function

token_sampler

token_sampler()

max.pipelines.sampling.token_sampler(sampling_config, device, return_logits=False, needs_bitmask_input=None, custom_extensions=(), unpadded_vocab_size=None)

source

Builds a sampling graph that samples tokens from logits.

Parameters:

• sampling_config (SamplingConfig) – Sampling configuration (top-k, temperature and so on).
• device (DeviceRef) – Device for the graph inputs and ops.
• return_logits (bool) – Whether the graph should expose logits as an output.
• needs_bitmask_input (bool | None) – Whether to wire a bitmask input into the graph. See TopKSampler.
• custom_extensions (Iterable[Path]) – Custom-op extension paths to compile the graph with. Empty by default.
• unpadded_vocab_size (int | None) – The tokenizer’s token count. See TopKSampler.

Returns:

A graph that takes logits (and optional penalty inputs) and outputs tokens.

Return type:

Graph