Fold a Llama-family model (tensor names model.embed_tokens / model.layers.N.
{input_layernorm, post_attention_layernorm, self_attn.{q,k,v,o}_proj,
mlp.{gate,up,down}_proj} / model.norm) per the turboquant-weights recipe:
untie → γ-fold → R1 → R2. Returns lazy bf16 tensors for the writer, with
norm gains replaced by ones and a separate folded lm_head (untied).
Fold a Llama-family model (tensor names model.embed_tokens / model.layers.N. {input_layernorm, post_attention_layernorm, self_attn.{q,k,v,o}_proj, mlp.{gate,up,down}_proj} / model.norm) per the turboquant-weights recipe: untie → γ-fold → R1 → R2. Returns lazy bf16 tensors for the writer, with norm gains replaced by ones and a separate folded lm_head (untied).