Expand description
2026-09-25: Shared foundation types and host-side helpers for the engine crates, one module each (listed below).
Owner: metrale-core. Invariants:
- The
cudafeature gates onlydevice::MetraleDeviceand theerror::MetraleError::CudaDrivervariant; the crate builds without it.
Modulesยง
- arch
- 2026-09-25: Decide whether PTX built for one SM architecture runs on a device, and name the target to rebuild when it does not.
- compute
- 2026-09-25: A build-time kernel compilation target: the
Vendorparsed fromHARDWARE.toml, theComputeTargettrait, and its nvcc implementation. - device
- 2026-09-25: GB10 hardware constants (
sm121) and the cudarc device handleMetraleDevice. - dtype
- 2026-09-25: Element types for weights and scales, and the quantization layout of a weight tensor.
- error
- fault
- 2026-09-25: Process-wide latch for a destroyed CUDA context, the probe-based verdict that sets it, and the exit status it maps to.
- mxfp4_
e8m0 - 2026-09-25: Native MXFP4 host unpack: packed E2M1 nibbles + E8M0 scales.
- numeric
- 2026-09-25: Host-side numeric conversions: the FP8 E4M3 decode table, the f32 -> BF16 cast and BF16 widening.
- safetensors
- 2026-09-25: The one place that turns a safetensors
data_offsetspair into a byte span. - scope
- 2026-09-25: Model teardown: ordered, fallible release of state that owns device memory.
- target
- 2026-09-25: Kernel target descriptor: the
(arch, model, quant)key of a compiled kernel set. - tensor
- 2026-09-25:
TensorRef, a non-owning description of a device tensor.