Load SAM2 Hydra YAML config and extract hyperparameters.

Load a tensor dumped by sam3_dump_state_tensor. Returns data in ggml flat order and the ggml shape (ne[0], ne[1], ne[2], ...).

Load a tensor saved by the Python dumper (standard row-major).

Compute IoU between two binary masks.

72

SAM3 Python preprocessing pipeline.

Map a SAM2 PyTorch state_dict key to a flat ggml tensor name. Returns None if the tensor should be skipped.

Map shared component keys (FPN, SAM, memory) to ggml names. Returns None if the tensor should be skipped.

Run C++ ggml inference for one variant.

Run Python SAM2 inference for one variant.

Bicubic interpolation of a [C, H_in, W_in] tensor on CPU to [C, H_out, W_out]. Used for background positional embedding interpolation.

Full SAM2 image encoding: preprocess → Hiera → FPN → state

Single Hiera MultiScaleBlock forward pass.

Set per-block metadata: stage_idx, dim_in/out, window_size, has_q_stride

Register all SAM2 tensors (Hiera backbone + FPN neck + shared)

Window partition: [B, H, W, C] -> [B*nW, ws, ws, C] with padding if needed. In ggml layout: input ne = {C, W, H, B}, output ne = {C, ws, ws, B*nW}. Pa

Window unpartition: reverse of sam2_window_partition. Input ne = {C, ws, ws, nW*B}, output ne = {C, orig_W, orig_H, B}. Reverses the partition steps:

Compute IoU between two boxes [x0, y0, x1, y1].

Apply BPE merges to a byte-encoded word string. Returns space-separated BPE tokens (e.g. "he llo</w>").

Build a causal (lower-triangular) attention mask for the text encoder. Returns: [L, L] F16 tensor. mask[kv][q] = 0 if kv <= q, -inf otherwise. Marked

Build query positional encoding from reference boxes via sine embed + ref_point_head MLP. ref_boxes: [4, NQ, 1] — after sigmoid sine_dim_t: [1, 64] Re

Build sinusoidal positional embedding for 4D reference points in the ggml graph. ref_boxes: [4, NQ, B] — (cx, cy, w, h) after sigmoid, B=1 sine_dim_t:

Build box-relative positional bias for DETR cross-attention. ref_boxes: [4, N_q, B] — (cx, cy, w, h) in [0,1] rpb_coords: [feat_hw] — normalized coord

Convert (cx, cy, w, h) in [0,1] to (x0, y0, x1, y1) in pixel coordinates.

Single DETR decoder layer. queries: [D, N_q, B] where N_q = 201 (200 object queries + 1 presence token) query_pos: [D, N_q, B] positional encoding for

DotProductScoring: classify queries against text features via dot product. Python reference (DotProductScoring.forward): 1. prompt_mlp(prompt) → resi

Compute feat_size for an arbitrary img_size.

Single fusion encoder layer. x: [D, N, B] image features (N=5184), prompt: [D, T, B] text/exemplar tokens, pos: [D, N, B] Returns: updated x [D, N, B]

1-D sinusoidal PE matching Python get_1d_sine_pe(pos_inds, dim, temperature). Output: [dim] floats — first dim/2 are sin, second dim/2 are cos.

Maps each byte 0-255 to a unique unicode character (as UTF-8 string). Printable bytes map to themselves; non-printable bytes map to U+0100..U+0143.

inverse_sigmoid: log(x / (1 - x)), clamped to avoid inf Python reference uses eps=1e-3: x1 = x.clamp(min=eps), x2 = (1-x).clamp(min=eps)

Load tensors from the binary file into the already-registered ggml tensors

Compute mask IoU between two binary masks.

Standard multi-head attention with fused in_proj. q_in, k_in, v_in: [D, N, B] (if fused_qkv, only q_in is used and contains QKV stacked) in_proj_w: [

Non-maximum suppression on detections, sorted by score descending. Returns indices of kept detections.

Build pixel decoder: progressively upsample FPN features. fpn_feats[0]: [D, 288, 288, B] (highest res) fpn_feats[1]: [D, 144, 144, B] fpn_feats[2]: [D