Module constrained_decoding 

Grammar-constrained decoding for token-by-token generation.

This module provides the TokenConstraint trait and concrete implementations that restrict which tokens the model can emit at each decoding step:

• NoConstraint — passthrough, all tokens allowed
• RegexConstraint — restricts output to strings matching a regex pattern
• JsonConstraint — restricts output to syntactically valid JSON
• AllowListConstraint — restricts output to one of a finite set of token sequences
• SequenceConstraint — forces output to reproduce a specific token sequence
• LengthConstraint — enforces hard minimum and maximum generation lengths

The ConstrainedSampler wraps a crate::sampling_advanced::SamplerChain and applies a mask to logits before sampling so that only valid continuations are drawn.

Example

use oxibonsai_runtime::constrained_decoding::{ConstrainedSamplerBuilder, TokenConstraint};

let mut sampler = ConstrainedSamplerBuilder::new(128, 42)
    .with_json_constraint();
assert!(!sampler.is_complete());

Module structure

Phase 30B split the monolithic constrained_decoding.rs (1966 lines) into focused sub-modules; all external crate::constrained_decoding::* access paths are preserved through the re-exports below.

• [error_trait][] — ConstraintError, the TokenConstraint trait, and the passthrough NoConstraint.
• [regex][] — NFA-based RegexConstraint.
• [json][] — JSON-grammar JsonConstraint and its JsonParseState.
• [sampler][] — ConstrainedSampler and ConstrainedSamplerBuilder.
• [allow_list][] — AllowListConstraint.
• [sequence][] — SequenceConstraint.
• [length][] — LengthConstraint.

Structs

AllowListConstraint

A constraint that forces the generated token sequence to exactly match one of a finite set of allowed token-id sequences (e.g., multiple-choice answers).

ConstrainedSampler

Wraps a crate::sampling_advanced::SamplerChain with a TokenConstraint.

ConstrainedSamplerBuilder

Ergonomic builder for ConstrainedSampler.

JsonConstraint

Constrains generation to syntactically valid JSON.

LengthConstraint

A constraint that enforces hard minimum and maximum token-count limits.

NoConstraint

A passthrough constraint that places no restriction on the vocabulary.

RegexConstraint

Constrains generation to strings that match a regular expression.

SequenceConstraint

A constraint that forces the generated output to reproduce a specific, pre-determined token sequence.

Enums

ConstraintError

Errors that can arise when building or running a token constraint.

JsonParseState

Internal parser state for JsonConstraint.

Traits

TokenConstraint

A constraint that restricts which tokens are valid at each decoding step.