A delta-compressed tensor stores the residual between the tensor bytes and a reference (e.g. the matching tensor from a base model checkpoint). For fine-tuned models the residual concentrates on very few bits and compresses far better than the raw tensor. This module exposes the delta math only. To obtain a compressed blob, compose with :mod:`weights.codecs`:: residual = ptwm.delta.encode(finetune_bytes, reference=base_bytes) blob = weights.codecs.get_codec("huffman").encode(residual) # inverse: residual = weights.codecs.get_codec("huffman").decode(blob, original_len) recovered = ptwm.delta.decode(residual, reference=base_bytes) Use :func:`reference_hash` to record which reference a residual was produced against; :func:`decode` accepts an optional ``expected_reference_hash`` for cross-checking at call time.

If ``expected_reference_hash`` is given, the BLAKE3 digest of ``reference`` is checked against it before reconstruction, guarding against silently decoding with the wrong reference.

``data`` and ``reference`` must have identical byte length. Returns a buffer of the same length; the inverse is :func:`decode`.