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 class

AcceptanceSampler

AcceptanceSampler

class max.nn.sampling.AcceptanceSampler(synthetic_acceptance_rate=None, num_draft_steps=1, use_stochastic=False, draft_proposal='argmax', vocab_size=None, relaxed_topk=None, relaxed_delta=None)

source

Bases: object

Dispatches between greedy, synthetic, and stochastic acceptance.

• synthetic_acceptance_rate set → synthetic (benchmarking) mode. The per-position acceptance probability is calibrated so that the mean joint acceptance across num_draft_steps matches the configured rate, via compute_synthetic_acceptance_base_rate().
• use_stochastic=True → stochastic rejection sampling. The caller must then pass per-row sampling params (temperature, top_k, max_k, top_p, min_top_p) at call time.
• Otherwise → greedy (accept iff draft token == target argmax).

Synthetic mode takes priority over stochastic when both are configured, but still samples with the stochastic params.

relaxed_topk / relaxed_delta require draft_proposal="argmax"; the relaxed rule assumes the drafted token is the draft’s own argmax, so it does not carry over to a sampled proposal.

Parameters:

• synthetic_acceptance_rate (float | None)
• num_draft_steps (int)
• use_stochastic (bool)
• draft_proposal (Literal['argmax', 'sampled'])
• vocab_size (int | None)
• relaxed_topk (int | None)
• relaxed_delta (float | None)

acceptance_rule

property acceptance_rule: Literal['synthetic', 'stochastic', 'greedy']

source

The rule __call__() dispatches to, in its precedence order.

This is the only place the acceptance rule in effect is decided: SpeculativeConfig.rejection_sampling_strategy is read by nothing.