Map a PyTorch state_dict key to the flat ggml name. Returns None if the tensor should be skipped.

Fuse RepVGG 3-way parallel branches + optional final BN into single DW 3×3. Branches: 1. DW Conv 3×3 + BN (prefix.conv.c.weight + prefix.c

Run ViT forward pass. If use_f16_weights, quantize all 2D weights to f16.

Build FPN neck graph for SAM2. Input: 4 stage outputs from Hiera backbone Output: 3 FPN levels (after scalp=1) in fpn_outs[] as [D, W, H, 1]

Build full Hiera backbone graph. Input: [img_size, img_size, 3, 1] preprocessed image Output: 4 stage outputs in stage_outs[] as [stage_dim, W, H, 1]

Compute Hiera positional embedding on CPU. Returns [E, H, W] in ggml column-major layout: element (e, x, y) at flat index e + x*E + y*E*W. PE = bicubi

Build memory attention graph. curr_tokens: [D, N, 1] — current frame image tokens (flattened from 72x72 = 5184) src_pos: [D, N, 1] — sinusoidal PE

Build prompt and prompt_pos tensors CPU-side for memory attention. mem_slot_feats[i]: [MD * H * H] raw spatial features (ggml layout) mem_slot_pes[i]:

Build the full ViT graph. Input: [img_size, img_size, 3, 1] (ggml convention: [W, H, C, B]) Output: [E, W, H, 1] where E=1024, W=H=72

Match the official image predictor prompt ordering: box corners (if any) → user points → trailing padding point. Even when a box is present, SAM keeps

CXBlock: depthwise conv + LayerNorm + pointwise MLP with residual scaling.

Fill small holes in a binary mask using BFS connected components.

Unit separator (0x1F) cannot appear in byte-encoded BPE tokens.

Pre-compute geometry encoder input on CPU: box embeddings + CLS token. Reads model weights from GPU, computes embeddings, returns float vector. Layout

Remove small foreground sprinkles.

Bilinear resize of a [H, W, 3] uint8 image to [dst_h, dst_w, 3].

Resolve overlapping masks: higher-scoring instances take priority.

Tokenize text into a fixed-length token ID vector [ctx_len]. Format: [SOT, bpe_tokens..., EOT, 0, 0, ..., 0]

returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.

adds an extra all-255 alpha channel dest == src is legal img_n must be 1 or 3

create the png data from post-deflated data

this function is designed to support animated gifs, although stb_image doesn't support it two back is the image from two frames ago, used for a very s

decode one 64-entry block--

get some unsigned bits

returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow

returns 1 if the product of two ints fits in a signed short, 0 on overflow.

set up the kernels

read 16bit value and convert to 24bit RGB