Expand description
Core quantization primitives for LittleBit (NeurIPS 2025) and LittleBit-2 (ICML 2026).
Re-exports§
pub use binary::binarize_sign_no_grad;pub use binary::smooth_sign;pub use binary::ste_binary;pub use config::compute_eff_bits;pub use config::estimate_split_dim;pub use config::finalize_split_dim;pub use config::parse_eff_bit_layers;pub use config::BiasMode;pub use config::BitBudget;pub use config::CastPolicy;pub use config::InferenceMode;pub use config::InitMode;pub use config::ItqMode;pub use config::LayerShape;pub use config::LittleBitConfig;pub use config::LittleBitConfigOverrides;pub use config::QuantFunc;pub use config::ResidualMode;pub use config::TrainMode;pub use config::WeightState;pub use config::SMOOTH_SIGN_DEFAULT_ALPHA;pub use error::LittleBitError;pub use error::Result;pub use itq::apply_joint_itq;pub use itq::compute_itq_rotation;pub use itq::random_orthogonal_matrix;pub use linear::pow2_at_least;pub use linear::reconstruct_branch_matrix;pub use linear::FactorBranch;pub use linear::LittleBitLinear;pub use linear::PackedLittleBitLinear;pub use linear::Param;pub use linear::FP16_MAX_SAFE;pub use packer::binary_packer;pub use packer::binary_unpacker;pub use packer::pack_i8_slice;pub use packer::unpack_i32_slice;pub use packer::PackedBinaryMatrix;pub use safetensors_io::inspect_checkpoint;pub use safetensors_io::load_and_unpack_state_dict;pub use safetensors_io::load_raw_safetensors_dir;pub use safetensors_io::save_packed_safetensors;pub use safetensors_io::CheckpointFormat;pub use safetensors_io::CheckpointStats;pub use safetensors_io::LayerInspection;pub use safetensors_io::LoadedStateDict;pub use safetensors_io::SafetensorsReader;pub use safetensors_io::ShardedSafetensorsWriter;pub use safetensors_io::TensorLocation;pub use svd::exact_thin_svd;pub use svd::rank_one_decompose;pub use svd::svd_lowrank;pub use svd::ThinSvd;pub use svd::DEFAULT_LOWRANK_NITER;
Modules§
- binary
- Binary activation/weight quantization functions (
STEBinaryandSmoothSign). - config
- Quantization configuration, algebraic modes, and bit-budget calculation.
- error
- Typed errors for
littlebit-core. - itq
- LittleBit-2 Joint Iterative Quantization (Joint-ITQ) latent rotation.
- linear
LittleBitLinearand compact 1-bitPackedLittleBitLinearmodules.- packer
- LSB-first 32-bit binary matrix packing, unpacking, and compact 1-bit matmul.
- safetensors_
io - Safetensors checkpoint loading, binary unpacking, packed saving, and header inspection.
- svd
- Pure-Rust thin SVD and randomized low-rank SVD backed by
faer.