Sampling and search

Inferlets that transform or search model distributions during decoding.

Inferlet	What it shows
sampling-primitives	Token, logits, entropy, probabilities, and log-probabilities from one forward pass.
consensus-decoding	Parallel candidate generation with string-similarity consensus ranking.
greenlist-watermarking	Greenlist token watermarking during generation.
mirostat-v2-sampling	Adaptive-surprise sampling with Mirostat v2.
contrastive-decoding	Same-model contrastive decoding with full-context expert and bounded-context amateur passes.
beam-search	On-device beam expansion over a shared prefix-tree KV cache.

Related guide

• Samplers and probabilities: every Sampler and Probe kind.
• The forward pass: the primitive these inferlets drive directly.
• Forking and saving: shared-prefix branching used by parallel candidates.