gensbi.experimental.models.pixeldit.rope

Rotary positional embedding tables for the PixelDiT port.

Ports precompute_freqs_cis_2d (reference modules.py:132-145) and fetch_pos_text (reference pixeldit_t2i.py:232-241) from the PyTorch PixelDiT code, but emits the real 2x2-rotation tensor consumed by gensbi.models.flux1.math.apply_rope instead of complex cis values.

The two formats are numerically equivalent: the reference rotates consecutive value pairs (x[2j], x[2j+1]) by the complex phase freqs_cis[n, j], and apply_rope applies the matching 2x2 rotation to the same consecutive pairs.

Functions

_rotation_from_angles(angles)	Build the (1, 1, N, head_dim/2, 2, 2) rotation table from per-pair angles.
precompute_freqs_cis_1d(head_dim, length[, theta])	1D integer-position rope for cond tokens (reference fetch_pos_text).
precompute_freqs_cis_2d(head_dim, height, width[, ...])	2D axial rope on fractional positions in [0, scale].

Module Contents

gensbi.experimental.models.pixeldit.rope._rotation_from_angles(angles)

Build the (1, 1, N, head_dim/2, 2, 2) rotation table from per-pair angles.

angles has shape (N, head_dim/2); each angle becomes a 2x2 rotation matrix matching flux1.math.rope.

gensbi.experimental.models.pixeldit.rope.precompute_freqs_cis_1d(head_dim, length, theta=10000.0)

1D integer-position rope for cond tokens (reference fetch_pos_text).

Same output layout as precompute_freqs_cis_2d() with N = length. The full head dim is rotated by the single axis, so there are head_dim/2 frequencies (step 2).

gensbi.experimental.models.pixeldit.rope.precompute_freqs_cis_2d(head_dim, height, width, theta=10000.0, scale=16.0)

2D axial rope on fractional positions in [0, scale].

Returns a (1, 1, H*W, head_dim//2, 2, 2) float32 rotation table. The x and y angle pairs are interleaved exactly as the reference interleaves the complex x_cis/y_cis phases.