embedding-free
PyTorch transformer input from raw bytes with zero learned input parameters: no tokenizer, embedding table, or learned input projection.
