Know the machine

The probe crate answers what a program needs before it picks a device. It works on its own, without embed or the model.

cargo run --release --example machine

Omarchy system information
Linux with AMD Vulkan
RX 9060 XT (RADV GFX1200) · 16 GB · 32 cores · 62 GB

embeddings  WebGPU
speech      Vulkan
video       VAAPI

adapter     AMD Radeon RX 9060 XT (RADV GFX1200), Vulkan, DiscreteGpu
adapter     llvmpipe (LLVM 22.1.8, 256 bits), Vulkan, Cpu
disk        382 GB free of 1999 GB

That is examples/machine.rs:

use std::path::Path;

let machine = probe::probe_machine();
println!("{}", machine.title());
println!("{}", machine.headline());
println!("{}", machine.spec_line());

println!("embeddings  {}", machine.filing.label());
println!("speech      {}", machine.listening.label());
println!("video       {}", machine.video.label());

for adapter in probe::probe_gpu().adapters {
    println!("adapter     {}, {:?}, {:?}", adapter.name, adapter.backend, adapter.device_type);
}
if let Some(disk) = probe::probe_disk(Path::new(".")) {
    println!("disk        {} GB free of {} GB", disk.free_bytes / 1_000_000_000, disk.total_bytes / 1_000_000_000);
}

What it reports

probe_machine() returns a Machine:

Field	What it holds
os	Linux, MacOs, Windows or Other
release	the system’s own name, such as the distribution, when it has one
vendor	the graphics vendor: Nvidia, Amd, Intel, Apple or None
card	the card’s name, when it is known
cores, ram_bytes, vram_bytes	the threads, the memory, and the card’s own memory when it has some
filing	the best lane for embeddings
listening	the best lane for speech models
video	the best lane for video decoding and encoding

Its headline(), spec_line() and title() are the three first lines above.

The lanes follow the vendor and the system:

	Embeddings	Speech	Video
NVIDIA	WebGPU (Linux)	CUDA	NVDEC and NVENC
AMD on Linux	WebGPU	Vulkan	VAAPI
Intel on Linux	WebGPU	Vulkan	Quick Sync
Apple	Core ML	Metal	VideoToolbox
no card	the processor	the processor	the processor

probe_gpu() lists every graphics adapter wgpu finds, with its backend and type, and runs once per process. probe_disk(path) gives the free and total bytes of the disk that holds path, or None when it can’t be read. probe_accel() names the video accelerator alone.

Pick a device turns filing into the device the model runs on.