Sampler adapter that wraps a llama_sampler* chain. More...

#include </home/runner/work/entropic/entropic/src/inference/llama_cpp_sampler.h>

Inheritance diagram for entropic::LlamaCppSampler:

Collaboration diagram for entropic::LlamaCppSampler:

Public Member Functions
	LlamaCppSampler (llama_sampler chain, llama_context ctx)
	Construct with an already-built llama_sampler chain.

	~LlamaCppSampler () override
	Free the underlying llama.cpp sampler chain.

	LlamaCppSampler (const LlamaCppSampler &)=delete

LlamaCppSampler &	operator= (const LlamaCppSampler &)=delete

int32_t	sample () override
	Sample one token from the current logits via the wrapped chain.

void	reset () override
	Reset llama_sampler internal state.

llama_sampler *	native_chain () const
	Expose the underlying chain for legacy call sites that have not yet been ported to the Sampler API.

Detailed Description

Sampler adapter that wraps a llama_sampler* chain.

Owns the chain (freed in destructor). Stores a non-owning llama_context* snapshot captured at factory creation time; the backend guarantees the context outlives every Sampler it vends.

Definition at line 47 of file llama_cpp_sampler.h.

Constructor & Destructor Documentation

◆ LlamaCppSampler()

entropic::LlamaCppSampler::LlamaCppSampler	(	llama_sampler *	chain,
		llama_context *	ctx
	)

Construct with an already-built llama_sampler chain.

Construct an LlamaCppSampler wrapping a pre-built sampler chain.

Parameters

chain	Owning pointer to llama_sampler chain.
ctx	Non-owning llama_context (snapshot at sampler birth).

Version: 2.3.10

Definition at line 120 of file llama_cpp_sampler.cpp.

◆ ~LlamaCppSampler()

entropic::LlamaCppSampler::~LlamaCppSampler ( )

override

Free the underlying llama.cpp sampler chain.

Definition at line 128 of file llama_cpp_sampler.cpp.

Member Function Documentation

◆ native_chain()

llama_sampler * entropic::LlamaCppSampler::native_chain ( ) const

inline

Expose the underlying chain for legacy call sites that have not yet been ported to the Sampler API.

Production callers in the decode loop should use sample() directly; this accessor exists for transition-period code and for tests asserting that production wiring really produces a non-null chain.

Returns: Borrowed llama_sampler* (never null for the production Sampler; do not free).

Definition at line 88 of file llama_cpp_sampler.h.

◆ reset()

void entropic::LlamaCppSampler::reset ( )

overridevirtual

Reset llama_sampler internal state.

Reset stateful samplers (e.g.

Forwards to llama_sampler_reset so callers re-seeding a generation pick up a fresh RNG / penalty-window state.

Version: 2.3.10

repeat-penalty history) on the chain.

Reimplemented from entropic::Sampler.

Definition at line 151 of file llama_cpp_sampler.cpp.

◆ sample()

int32_t entropic::LlamaCppSampler::sample ( )

overridevirtual

Sample one token from the current logits via the wrapped chain.

Returns: Sampled token id, or -1 if the chain or context is missing.

Implements entropic::Sampler.

Definition at line 141 of file llama_cpp_sampler.cpp.

The documentation for this class was generated from the following files:

inference/llama_cpp_sampler.h
inference/llama_cpp_sampler.cpp

Public Member Functions