Crate ferrum_sampler 

Ferrum Sampler

MVP sampler implementation for Ferrum inference stack.

This crate provides a thin wrapper around the sampling interfaces defined in ferrum-interfaces, offering convenient factory functions and utilities for building sampling pipelines from SamplingParams.

Design

• Re-export Interface Types: All core types from ferrum-interfaces::sampler
• Factory Pattern: Simple factory for creating samplers and configs
• Zero Overhead: Direct delegation to interface implementations

Usage

use ferrum_sampler::{build_sampling_config, sampler_from_params};
use ferrum_types::SamplingParams;

let params = SamplingParams::default();
let config = build_sampling_config(&params);
let sampler = sampler_from_params(&params);

Modules

json_mode

JSON mode logits processor.

Structs

DefaultSamplerFactory

Default sampler factory for creating samplers and configurations.

GreedySampler

Common samplers Greedy sampler (always picks highest probability token)

LogitsProcessorChain

Logits processor chain for composing multiple processors

MultinomialSampler

Multinomial sampler for probabilistic sampling

RepetitionPenaltyProcessor

Repetition penalty processor

SamplingConfig

Complete sampling configuration

SamplingConfigBuilder

Sampling configuration builder

SamplingContext

Sampling context passed to logits processors and samplers

SamplingParams

Sampling parameters for generation

SamplingPipeline

Sampling pipeline that combines config and execution logic.

SamplingStats

Sampling statistics for monitoring

TemperatureProcessor

Common logits processors Temperature scaling processor

TokenId

Token identifier used across the inference pipeline.

TopKProcessor

Top-k filtering processor

TopPProcessor

Top-p (nucleus) filtering processor

Enums

ProcessorPriority

Priority levels for logits processors

Traits

LogitsProcessor

Logits processor trait for modifying raw model outputs

MultiSampler

Multi-sample capability for beam search and parallel sampling

Sampler

Token sampler trait for selecting next token from processed logits

Functions

build_sampling_config

Build sampling configuration from parameters.

greedy_sampler

Create a greedy sampler (always picks highest logit).

multinomial_sampler

Create a multinomial sampler (probabilistic sampling).

pipeline_from_params

Build complete sampling pipeline from parameters.

sampler_from_params

Create sampler instance from parameters.

Type Aliases

Result

Result type used throughout Ferrum