mlnomad/gelu-d12-chinchilla-261M-pytorch

Fix: lazy non-persistent RoPE buffers — HF from_pretrained's meta-init was leaving persistent rope_cos/rope_sin uninitialised, producing NaN logits. Buffers are now computed on first forward.

Fix: lazy non-persistent RoPE buffers — HF from_pretrained's meta-init was leaving persistent rope_cos/rope_sin uninitialised, producing NaN logits. Buffers are now computed on first forward.

Fix: lazy non-persistent RoPE buffers — HF from_pretrained's meta-init was leaving persistent rope_cos/rope_sin uninitialised, producing NaN logits. Buffers are now computed on first forward.

Fix: lazy non-persistent RoPE buffers — HF from_pretrained's meta-init was leaving persistent rope_cos/rope_sin uninitialised, producing NaN logits. Buffers are now computed on first forward.

Fix: GQA → MHA (n_head = n_kv_head, no head sharing)

Add wikitext-103 eval: loss=3.840, PPL=46.52

Initial PyTorch port of GELU d12 Chinchilla (261M) — AutoModelForCausalLM compatible

initial commit