Preprocess an image: resize to img_size × img_size, convert to float, normalize. Returns a float tensor in [C, H, W] layout (channel-first), range nor

Splits text into word tokens following the CLIP pattern: <|startoftext|> | <|endoftext|> | 's|'t|'re|'ve|'m|'ll|'d | [\p{L}]+ | [\p{N}] | [^\s\p{L}\p{

Print a per-op-type summary of a built graph.

Register all tensor names in the model struct so we can look them up by name when loading from the binary file. This creates ggml tensors with no_allo

CPU-side ROI Align matching torchvision.ops.roi_align behavior. Features in ggml [C, W, H] layout. Box in XYXY format scaled to feature grid coords. U

Select memory frames for propagation (most recent + evenly spaced).

Sinusoidal positional encoding for box coordinates. Matches Python PositionEmbeddingSine.encode_boxes(cx, cy, w, h). Output: [258] = [pos_y(128), pos_

Single text encoder block forward pass. Input x: [E, L] where E=text_width=1024, L=seq_len (typically 32). causal_mask: [L, L] F16 additive mask for g

TwoWayAttentionBlock forward

Split a UTF-8 string into individual unicode characters.

Single ViT block forward: pre-norm → attn (window or global, with RoPE) → residual → pre-norm → MLP → residual x: [E, W, H, B] in ggml layout (followi

Encode box (cx, cy, w, h) with sinusoidal PE. Returns [N, 258].

Encode (x, y) with sinusoidal PE. x, y: [N] normalized to [0,1].

Match PositionEmbeddingSine.forward() exactly.

build a table that decodes both magnitude and value of small ACs in one go.

decode image to YCbCr format

@OPTIMIZE: store non-zigzagged during the decode passes, and only de-zigzag when dequantizing

after we see SOS