Crate torg_mask 

LLM logit masking for TØR-G constrained decoding.

This crate provides utilities for integrating TØR-G’s valid_next_tokens() with LLM inference engines.

Overview

TØR-G graphs are constructed by feeding tokens to a Builder. At each step, only certain tokens are valid (returned by Builder::valid_next_tokens()). This crate provides:

• Token mapping: Bidirectional mapping between TØR-G tokens and LLM vocab IDs
• Logit masking: Efficiently mask invalid tokens in logit vectors
• Constrained decoding: High-level orchestrator for decode loops

Example

use torg_mask::{ConstrainedDecoder, MaskGenerator, TokenMapping};

// Create a token mapping (in production, map to actual LLM vocab IDs)
let mapping = TokenMapping::sequential(256);
let generator = MaskGenerator::new(mapping.clone(), 1000);
let mut decoder = ConstrainedDecoder::new(generator);

// Simulate a decode loop
let token_ids = [
    5,   // InputDecl
    9,   // Id(0)
    3,   // NodeStart
    10,  // Id(1)
    0,   // Or
    9,   // Id(0)
    7,   // True
    4,   // NodeEnd
    6,   // OutputDecl
    10,  // Id(1)
];

for &id in &token_ids {
    let mask = decoder.next_mask();
    assert!(mask.is_allowed(id));
    decoder.feed_token(id).unwrap();
}

let graph = decoder.finish().unwrap();
assert_eq!(graph.nodes.len(), 1);

Re-exports

pub use decoder::ConstrainedDecoder;

pub use decoder::DecodeError;

pub use generator::MaskGenerator;

pub use mapping::TokenMapping;

pub use mapping::TokenMappingBuilder;

pub use mask::LogitMask;

pub use torg_core;

Modules

decoder

High-level constrained decoding orchestrator.

generator

Mask generator for creating logit masks from builder state.

mapping

Token mapping between TØR-G tokens and LLM vocabulary IDs.

mask

Logit mask for LLM inference.