Recursively replaces every nn.Linear inside *module* with BitLinear, skipping the token embedding and lm_head (which stay full-precision).

Convert a HuggingFace LlamaForCausalLM to a BitNet b1.58 model in-place. Steps (exactly as described in the training manuscript): 1. R

Generate text continuation for *prompt*. Args: model: The BitNet LlamaForCausalLM. tokenizer: Matching

Return True if at least one module in the model is already a BitLinear.

Per-token absmax activation quantisation to int8 range [-127, 127]. scale = max(|x|, dim=-1) / 127 (per token, shape [batch, seq, 1]) x

Absmean weight quantisation → {-1, 0, +1}. scale = mean(|W|) + ε (per weight-matrix, scalar) W_q = round_clamp(W / scale) vi

Remove the pre-attention and pre-FFN RMSNorms from every decoder layer. Per the paper: "Remove RMSNorm before attention and SwiGLU because

Load and tokenise the dataset. The dataset is expected to have a "text" column. Each example is tokenised and truncated / padded to `co

Build a small LLaMA 2 model from scratch, then convert to BitNet.

Construct a BitLinear from an existing nn.Linear, copying its weight.

Return a bf16 autocast context when supported, otherwise a no-op.

Load model and tokenizer from *model_path*. If the saved model does not yet contain BitLinear layers (e.g. a plain LLaMA checkpoint), th