Parse + store mrope_section / mrope_interleaved in RopeParameters
(previously accepted-but-ignored). RotaryEmbedding gains:
- inv_freq + per-axis column masks (mask_t/h/w) built from mrope_section;
- plain_cos_sin(pos, seq_len): narrow the precomputed tables (text/decode);
- mrope_cos_sin(position_ids (3,seq)): per-axis freqs blended at the
interleave columns (vision);
- apply_cos_sin(q,k,cos,sin): the rope_slow application, factored out.
The existing apply(q,k,offset) is retained (delegates to
plain_cos_sin + apply_cos_sin) so current callers are unchanged; Stages
3–4 move cos/sin construction into the model forward and thread the 3D
position ids for image tokens.
Tests: masks partition the half-dim; interleave drives the right axis
per column; and the load-bearing invariant — mrope_cos_sin reduces
bit-for-bit to plain_cos_sin when the three axes are equal (so text
inference is unchanged).
Refs the MRoPE-gap diagnosis (vision spatial misread). Pure non-cuda;
no behaviour change until wired.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
rob thijssen
5731f4c318
