Module common_sampler 

llama.cpp’s assembled sampler chain, and the reasoning-budget sampler.

LlamaSampler wraps the individual llama_sampler_* primitives, leaving the caller to build a chain. This wraps common_sampler — the chain upstream assembles for its own tools — which is a different trade: less control, but it gets right several things that are easy to miss when hand-rolling.

• Ordering. Penalties and DRY run before truncation samplers, which run before temperature. A chain in the wrong order silently samples from the wrong distribution.
• Grammar prefill. Output-format and tool-call grammars are advanced past the generation prompt; user grammars are not. See ChatParams::grammar_sampler for what goes wrong otherwise.
• Model-declared suppress tokens are merged into the logit bias, so tokenizer.ggml.suppress_tokens is honoured without the caller knowing it exists.
• Reasoning budget, created whenever a lazy grammar is active so a thinking block can be force-closed.
• Speculative acceptance via CommonSampler::sample_and_accept_n.

Structs

CommonSampler

llama.cpp’s assembled sampler chain.

CommonSamplerParams

Parameters for CommonSampler::new, seeded from llama.cpp’s defaults.

ReasoningBudget

A sampler that caps how long a model may think.

Enums

CommonSamplerType

Which of llama.cpp’s samplers to run, and in what order.

GrammarSource

How a grammar was obtained, which decides whether the generation prompt is prefilled into it.

ReasoningBudgetState

Where a ReasoningBudget is in its state machine.

Type Aliases

CommonSamplerError

Errors from the common-sampler layer.

CommonSamplerScalars

The scalar half of llama.cpp’s sampling parameters.