Self-attention with optional RoPE, head-wise RMSNorm on Q/K, and an optional FlashAttention path.

Muon: Momentum-Orthogonalized by Newton-Schulz. Applies only to 2D matmul-shaped parameters. 1D/embedding/bias parameters should be routed to

Run 32-step generation + gen_ppl + sample_entropy at every epoch end.

Frozen text encoder contract for ELF. The ELF transformer needs only three things from an encoder: 1. `d_model` — the hidden size of the la

Generative-PPL + unigram-entropy evaluator (single-process).

Frozen T5 encoder. NO internal op is changed — exact reproduction of HF T5. Notes: * Runs in fp32 (matches ELF JAX). Do not override unless