Crate large 

LARGE — Lightweight Architecture for Running Generative Engines.

An educational, from-scratch LLM inference engine written in Rust, targeting CPU inference on Qwen3-0.6B using the GGUF model format.

Modules

• gguf — GGUF file format parser with memory-mapped tensor access
• tensor — Dequantization and math operations (mat-vec, RMSNorm, RoPE, etc.)
• tokenizer — GPT-2 style byte-level BPE tokenizer
• model — Qwen3 transformer model (GQA, SwiGLU, KV cache)
• sampler — Token sampling strategies (greedy, temperature, top-p)

Modules

gguf

GGUF file format parser.

model

Qwen3 transformer model for inference.

sampler

Token sampling strategies for autoregressive generation.

tensor

Tensor operations for LLM inference.

tokenizer

BPE tokenizer for Qwen3 models.