I built a GPU inference engine in Rust that runs Meta's SAM 3 end-to-end on Vulkan. No CUDA, no PyTorch at deployment.
7.9 MB runtime. 1 GB int8 weights.
314 ms vs 346 ms (PyTorch CUDA bf16, RTX 4070, image + 6 text prompts).
Code: github.com/automataIA/onnx-v…
Quantized ONNX models + runnable examples (text, point & box prompts):
huggingface.co/Autognosi/sam…
Full technical breakdown, benchmarks, quantization trade-offs and limitations:
autognosi.substack.com/p/seg…
Rewrote my coding agent: from ~111k lines of TypeScript -> ~21k of Rust. 222 MB of Node -> one 14.5 MB binary.
Paired vs Claude Code, same model: 479k input tokens per task against 4.07M. 73 credits per solve vs 343.
One binary: TUI, CLI, Zed, VS Code.
github.com/automataIA/sir-bo…
Built a Vulkan engine for ONNX in Rust: any GPU, no CUDA, 7 MB binary against the 4 GB of vendor libs. Alpha/early stage
I measured it against WebGPU EP and i win 7/11 ()
"It ain't much, but it's honest work"
github.com/automataIA/onnx-v…
