tracel-ai/burn · error

BURN_DEVICE=cuda requested, but the 'cuda' feature is not en

Error message

BURN_DEVICE=cuda requested, but the 'cuda' feature is not enabled.

What it means

Environment/feature mismatch guard at backend selection: `ReadDevice::default()` parses `BURN_DEVICE`; the `cuda` arm returns a CUDA device only when the `cuda` cargo feature is compiled in, otherwise it panics. It fires at runtime when the env var requests a backend the binary was not built with — the input at fault is the `BURN_DEVICE` value versus the crate's enabled features.

Source

Thrown at crates/burn-dispatch/src/device.rs:235

    #[allow(unreachable_code)]
    fn default() -> Self {
        // TODO: which priority?
        // Single override e.g. `BURN_DEVICE=vulkan` forces Vulkan or panics if not available.
        // Priority list e.g. `BURN_DEVICE_PRIORITY=cuda,vulkan,cpu` sets the order.
        // Both could be tied into `burn.toml` config
        // For now we just use `BURN_DEVICE` on CI to force a single device

        #[cfg(feature = "std")]
        {
            if let Ok(device_str) = std::env::var("BURN_DEVICE") {
                match device_str.to_lowercase().as_str() {
                    // Every cubecl runtime is the one `Cube` backend; the name here
                    // picks the runtime the device names, and the wgpu spellings all
                    // reach wgpu, whose compiler is chosen for it at runtime.
                    "cuda" => {
                        #[cfg(feature = "cuda")]
                        return Self::Cube(CubeDevice::Cuda(Default::default()));
                        panic!(
                            "BURN_DEVICE=cuda requested, but the 'cuda' feature is not enabled."
                        );
                    }
                    "rocm" => {
                        #[cfg(feature = "rocm")]
                        return Self::Cube(CubeDevice::Hip(Default::default()));
                        panic!(
                            "BURN_DEVICE=rocm requested, but the 'rocm' feature is not enabled."
                        );
                    }
                    "metal" | "vulkan" | "webgpu" | "wgpu" => {
                        #[cfg(any(
                            feature = "metal",
                            feature = "vulkan",
                            feature = "webgpu",
                            feature = "wgpu"
                        ))]
                        return Self::Cube(CubeDevice::Wgpu(Default::default()));

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Rebuild with the `cuda` feature enabled (e.g. `burn = { features = ["cuda"] }`).
2. Remove/unset `BURN_DEVICE` or set it to a backend the build supports.
3. Check CI/deployment scripts that export BURN_DEVICE and align them with the compiled feature set.

When it happens

Trigger: Thrown at crates/burn-dispatch/src/device.rs:235 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/afdabb80fa107898. Report an issue: GitHub.