returns 1 if the product is valid, 0 on overflow. negative factors are considered invalid.

Compare two tensors and return detailed metrics.

Forward through MLP layers from state dict.

Manual multi-head attention forward.

Resolve effective img_size / feat_size from state (which may override hp defaults).

clean up the temporary component buffers

decode a jpeg huffman value from the bitstream

after a restart interval, stbi__jpeg_reset the entropy decoder and the dc prediction

mallocs with size overflow checking

initialize a callback-based context

Fuse Conv2d + BatchNorm2d into a single Conv2d with weight and bias.

Sort key: family → size → precision

Build full fusion encoder graph (6 layers). image_feats: [D, N, B] where N=5184 (72*72), D=256 prompt_tokens: [D, T, B] text/exemplar features pos_enc

Build the SimpleFPN neck graph for one path (detector or tracker). Input: ViT output [E, W, H, B] with E=1024, W=H=72 But the conv ops expect [W, H, C

Lazily compute and cache PE/RoPE data that is identical across all propagation calls.

Read SAM prompt encoder weights from GPU and cache them in state. Also pre-computes the dense PE grid and no-mask tiled embedding. These never change

Convert screen position to image coordinates

combined JPEG 'receive' and JPEG 'extend', since baseline always extends everything it receives.

returns 0..31 for the highest set bit

extract an arbitrarily-aligned N-bit value (N=bits) from v, and then make it 8-bits long and fractionally extend it to full full range.

returns STBI_rgb or whatever, 0 on error

Stage 0: Dump preprocessed image tensor.

Build the full RepViT backbone graph. Input: [W, H, 3, 1] image Returns 4 feature maps as stage_outs[0..3], each [C_stage, W_stage, H_stage, 1]