sam3.cpp

C++ / ggml inference engine for SAM 3 (Segment Anything Model 3). Runs on Metal, CUDA, Vulkan, and CPU.

Quickstart

# Build
cmake -S . -B build && cmake --build build -j

# Run (auto-downloads weights from HuggingFace on first use, no extra deps)
python3 -c "
import sam3cpp
model = sam3cpp.Sam3Model('hf://rob-laz/sam3-gguf/sam3-image-f16.gguf')
pred = model.predict('photo.jpg', 'person')
print(f'{pred.count} detections')
"

Or use the CLI tools directly with a local GGUF:

./build/sam3-vision-trunk models/sam3-image-f16.gguf image.jpg output_prefix

Pre-trained Weights

Available on HuggingFace: rob-laz/sam3-gguf

The Python binding accepts hf://rob-laz/sam3-gguf/<filename> paths and will download + cache automatically (no extra dependencies — uses stdlib urllib).

Based on SAM 3 by Meta FAIR.

Architecture

• Vision backbone: 32-layer ViT with windowed + global attention (1008x1008 input)
• Text encoder: CLIP-based language backbone
• Encoder fusion: 6-layer transformer with self + cross-attention
• Decoder: 6-layer transformer decoder with iterative box refinement
• Heads: Grounding head (detection) + segmentation head (masks)

Building

cmake -S . -B build
cmake --build build -j

Metal is enabled by default on macOS. All other GPU backends are disabled but can be toggled via CMake options.

Python Binding

uv pip install pybind11 scikit-build-core
uv pip install -e bindings/python

python3 -c "
import sam3cpp

# Auto-download F16 weights from HuggingFace
model = sam3cpp.Sam3Model('hf://rob-laz/sam3-gguf/sam3-image-f16.gguf')
pred = model.predict('photo.jpg', 'person')

# pred.scores, pred.boxes_xyxy, pred.masks are numpy arrays
for i in range(pred.count):
    print(f'score={pred.scores[i]:.3f} box={pred.boxes_xyxy[i]}')

# Overlay visualization
overlay = sam3cpp.draw_overlay('photo.jpg', pred, mask_index=0)
overlay.save('output.png')
"

Rust Binding

cd bindings/rust
cargo build --example predict
cargo run --example predict

CLI Tools

All tools accept --cpu to force CPU execution. Default is GPU.

Converting Weights

python3 tools/convert_mlx_sam3_to_gguf.py \
  --input models/mlx-sam3/model.safetensors \
  --index models/mlx-sam3/model.safetensors.index.json \
  --output models/sam3-image-f32.gguf

# Quantize
python3 tools/convert_mlx_sam3_to_gguf.py \
  --input models/mlx-sam3/model.safetensors \
  --index models/mlx-sam3/model.safetensors.index.json \
  --output models/sam3-image-q8_0.gguf \
  --quantize q8_0

Performance

End-to-end pipeline benchmark on Apple M4 Max (1008x1008 input, Metal GPU):

Key GPU optimizations:

• Vision neck: transposed convolution decomposed into matmul + pixel shuffle (was 166s with CPU-only conv_transpose_2d)
• Encoder fusion: flash attention via ggml_flash_attn_ext (was 1.8s with materialized O(n²) score matrices)
• Segmentation head: group norm with interleaved channels runs on GPU (was 0.6s on CPU)
• Vision trunk: window partition/unpartition via reshape+permute (replaces CPU-only ggml_win_part)

License

Apache-2.0