Feature flags

The sceptre library ships default = [] — nothing is enabled by default, so every consumer picks its backend and capabilities explicitly.

Feature	Enables	Pick it when
ort	The native ONNX Runtime backend, provisioning-agnostic.	You’re linking libonnxruntime yourself and don’t need either provisioning helper below.
ort-bundled	ort plus a prebuilt ONNX Runtime fetched at build time (download-binaries, tls-rustls).	You want a zero-config native binary for a target with a prebuilt runtime available.
ort-dynamic	ort plus dlopen-based loading of libonnxruntime at runtime via ORT_DYLIB_PATH.	You want no build-time download — e.g. shipping a dylib next to the release binary — or you’re on a target with no prebuilt.
tract	The pure-Rust ONNX backend.	You’re targeting WASM or Android, want no native library dependency at all, or you’re on a target with no prebuilt and no libonnxruntime.
candle	Native-tensor backend that needs no ONNX Runtime: hand-written CRAFT and CRNN networks over the weights in the same ONNX files.	You want a build with no ONNX Runtime, or a GPU without a provider-specific runtime. Slower than ort — see Backends.
download	Runtime model download, caching, and SHA-256 verification from Hugging Face.	You want the library to fetch missing models itself rather than requiring them pre-provisioned.
mcp	The MCP (rmcp) server surface.	You’re running sceptre mcp or embedding the MCP server in your own binary.
ort-coreml	Compiles in ONNX Runtime’s CoreML execution provider.	You set model.accelerator = "coreml" (or --accelerator coreml) on macOS/iOS.
ort-directml	Compiles in ONNX Runtime’s DirectML execution provider.	You set model.accelerator = "directml" on Windows.
ort-cuda	Compiles in ONNX Runtime’s CUDA execution provider.	You set model.accelerator = "cuda" and your ONNX Runtime build carries CUDA.
candle-metal	Compiles in candle’s Metal device (implies candle).	You set model.accelerator = "metal" with --backend candle on macOS/iOS.
candle-cuda	Compiles in candle’s CUDA kernels (implies candle). Needs the CUDA toolkit at build time.	You set model.accelerator = "cuda" with --backend candle.
bench	Exposes the crate-internal detection/recognition hot paths through a bench seam module for the criterion benchmarks (cargo bench).	You’re running or writing the workspace’s own microbenchmarks; library consumers never need it.

The ort-* accelerator features only add the Rust-side provider registration; the ONNX Runtime build in use must also carry that provider. The pyke prebuilt for aarch64-apple-darwin is CoreML-enabled by construction, so on Apple Silicon ort-coreml is sufficient on its own.

ort vs. tract: which backend

See Backends for the full tradeoff; in short, ort for native CPU performance on desktop/server, tract for a pure-Rust WASM/Android build.

CLI defaults

The sceptre-cli binary’s own default feature set is ["ort-bundled", "download"], so cargo install sceptre-cli produces a self-contained binary: the ONNX Runtime is fetched and linked at build time, and models download at first use. The CLI re-exports eleven features, each forwarding to the identically-named library feature: ort-bundled, ort-dynamic, tract, candle, ort-coreml, ort-directml, ort-cuda, candle-metal, candle-cuda, download, and mcp (off by default, which adds the sceptre mcp subcommand). Compiling in an accelerator feature only makes the device available to select — the user still picks it at runtime with model.accelerator (or --accelerator, see Backends).

tract and candle are CLI features in their own right. Without them the matching --backend value is accepted by the argument parser but can never resolve, because no shipped build compiled the backend in — the run fails at model load with “not compiled in”.

Tip

ort-bundled and ort-dynamic are two provisioning strategies for the same ort backend, and enabling both is the supported way to override the default. Cargo features are additive, so --features ort-dynamic unions with the CLI’s ort-bundled default — and ort-dynamic still wins, because load-dynamic implies ort-sys/disable-linking, which the ort-sys build script checks before its prebuilt-download branch. You do not need --no-default-features to switch provisioning strategy.

Targets with no prebuilt ONNX Runtime

ort-bundled resolves a prebuilt runtime from a fixed target list. On a target that has none, the build fails with ort-sys’s own no prebuilt binaries available for target ... — at build time, so cargo install sceptre-cli never produces a binary at all:

• x86_64-apple-darwin (Intel macOS)
• every *-unknown-linux-musl target (Alpine)
• armv7-unknown-linux-gnueabihf
• riscv64gc-*
• *-unknown-freebsd
• i686-*, s390x-*, powerpc64le-*

On those targets pick ort-dynamic and supply a libonnxruntime, or drop ort entirely with --no-default-features --features tract,download and run with --backend tract. See Installation.