Crate oxibonsai_core 

oxibonsai-core

GGUF Q1_0_g128 format parser, tensor types, and model configuration for OxiBonsai — the Pure Rust 1-bit LLM inference engine.

This crate provides the foundational data types and parsing logic used by the rest of the OxiBonsai stack:

• GGUF v3 binary format parsing — header, metadata key-value store, and tensor info directory (see gguf).
• Q1_0_g128 block type — the 18-byte packed representation used for 1-bit weights (see tensor::BlockQ1_0G128).
• Memory-mapped tensor loading — zero-copy access to weight data from disk via memmap2.
• Model configuration — config::Qwen3Config extracted from GGUF metadata or constructed for known Bonsai variants (8B, 4B, 1.7B).

GGUF Q1_0_g128 Format

Each block is 18 bytes: 2-byte FP16 scale + 16 bytes (128 sign bits). Weight = bit ? +scale : -scale. Effective 1.125 bits per weight.

Crate Organisation

Module	Purpose
config	Qwen3Config with named constructors for each variant
gguf	Low-level GGUF v3 reader (header, metadata, tensors)
quant_ternary	BlockTQ2_0_g128, BlockTQ2_0, TernaryCode — ternary block types
tensor	BlockQ1_0G128 and OneBitTensor types
error	BonsaiError / BonsaiResult

Re-exports

pub use config::Qwen3Config;

pub use error::BonsaiError;

pub use error::BonsaiResult;

pub use gguf::compat::build_compat_report;

pub use gguf::compat::check_gguf_header;

pub use gguf::compat::CompatError;

pub use gguf::compat::ExtendedQuantType;

pub use gguf::compat::GgufCompatReport;

pub use gguf::compat::GgufVersion;

pub use gguf::header::GgufHeader;

pub use gguf::metadata::MetadataStore;

pub use gguf::metadata::MetadataValue;

pub use gguf::model_card::keys as model_card_keys;

pub use gguf::model_card::extract_known_fields;

pub use gguf::model_card::extract_model_card;

pub use gguf::model_card::ModelCard;

pub use gguf::streaming::GgufStreamParser;

pub use gguf::streaming::GgufValue;

pub use gguf::streaming::StreamState;

pub use gguf::streaming::StreamedGguf;

pub use gguf::streaming::StreamedTensorInfo;

pub use gguf::tensor_info::TensorInfo;

pub use gguf::tensor_info::TensorStore;

pub use gguf::types::GgufTensorType;

pub use gguf::types::GgufValueType;

pub use gguf::writer::MetadataWriteValue;

pub use gguf::writer::GgufWriter;

pub use gguf::writer::TensorEntry;

pub use gguf::writer::TensorType;

pub use gguf::writer::WriteError;

pub use quant_fp8::fp8_e4m3_decode;

pub use quant_fp8::fp8_e4m3_encode;

pub use quant_fp8::fp8_e5m2_decode;

pub use quant_fp8::fp8_e5m2_encode;

pub use quant_fp8::BlockFP8E4M3;

pub use quant_fp8::BlockFP8E5M2;

pub use quant_fp8::BLOCK_FP8_BYTES;

pub use quant_fp8::FP8_E4M3_MAX;

pub use quant_fp8::FP8_E5M2_MAX;

pub use quant_fp8::QK_FP8;

pub use quant_k::BlockQ2K;

pub use quant_k::BlockQ3K;

pub use quant_k::BlockQ4K;

pub use quant_k::BlockQ8K;

pub use quant_k::BLOCK_Q2_K_BYTES;

pub use quant_k::BLOCK_Q3K_BYTES;

pub use quant_k::BLOCK_Q4_K_BYTES;

pub use quant_k::BLOCK_Q8K_BYTES;

pub use quant_k_ext::BlockQ5K;

pub use quant_k_ext::BlockQ6K;

pub use quant_k_ext::BLOCK_Q5K_BYTES;

pub use quant_k_ext::BLOCK_Q6K_BYTES;

pub use quant_std::BlockQ4_0;

pub use quant_std::BlockQ8_0;

pub use quant_std::BLOCK_Q4_0_BYTES;

pub use quant_std::BLOCK_Q8_0_BYTES;

pub use quant_std::QK_Q4_0;

pub use quant_std::QK_Q8_0;

pub use quant_ternary::BlockTQ2_0;

pub use quant_ternary::BlockTQ2_0_g128;

pub use quant_ternary::TernaryCode;

pub use quant_ternary::BLOCK_TQ2_0_BYTES;

pub use quant_ternary::BLOCK_TQ2_0_G128_BYTES;

pub use quant_ternary::QK_TQ2_0;

pub use quant_ternary::QK_TQ2_0_G128;

pub use tensor::BlockQ1_0G128;

pub use tensor::OneBitTensor;

Modules

config

Qwen3 model configuration extracted from GGUF metadata.

error

Error types for OxiBonsai core operations.

gguf

GGUF v3 binary format parser.

quant_fp8

FP8 quantization block types: E4M3FN and E5M2.

quant_k

K-quant block types for Q2_K, Q3_K, Q4_K, and Q8_K quantization formats.

quant_k_ext

K-quant block types for Q5_K and Q6_K quantization formats.

quant_std

Standard GGUF quantization block types: Q4_0 (4-bit) and Q8_0 (8-bit).

quant_ternary

Ternary quantization block types for TQ2_0_g128 and TQ2_0 formats.

tensor

Q1_0_g128 tensor types and 1-bit data access.