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

synthetic_acceptance_sampler

synthetic_acceptance_sampler()

max.nn.sampling.synthetic_acceptance_sampler(draft_tokens, target_logits, base_acceptance_rate, num_draft_steps, seed)

source

Synthetic sampler for speculative decoding benchmarking.

Accepts each draft position independently with probability base_acceptance_rate. Once a position is rejected all subsequent positions are also rejected. Recovered tokens and bonus tokens are taken from the target argmax — generated text is not a faithful speculative decode; intended for throughput benchmarking only.

Parameters:

• draft_tokens (TensorValue) – Draft token ids [batch, num_steps].
• target_logits (TensorValue) – Verified target logits.
• base_acceptance_rate (float) – Per-position acceptance probability.
• num_draft_steps (int) – Number of speculative draft steps.
• seed (TensorValue) – Per-execute seed tensor. A rank-1 per-row seed uses its row-0 value: synthetic acceptance is batch-level benchmarking noise.

Return type:

tuple[TensorValue, TensorValue, TensorValue]

Returns (first_rejected_idx, recovered_tokens, bonus_tokens)