numr 0.12.0

WebGPU copies and gathers any element size at any offset, and runs past the 65535-workgroup and 256 MiB defaults.

Added

• Runtime::handles_take_byte_offsets — whether a storage handle plus a byte offset addresses device memory. The default is true. WebGPU returns false, and ensure_contiguous copies an offset view to fresh storage there.
• Runtime::computes_float_dtype — whether a backend computes in a float dtype. The default is true for every dtype. WebGPU returns true for F32 only.

Fixed

• WebGPU strided copy and index_select — element sizes that are not a multiple of 4 bytes and unaligned source offsets work. This covers U8, I8, Bool, F16, BF16 and quantized blocks. Before, the copy returned BackendLimitation.
• WebGPU offset views — a kernel binds the whole buffer from byte 0. A kernel handed a view at a nonzero offset returns an error, and ensure_contiguous copies such a view first. A rank-0 view copies as one element.
• WebGPU GEMV-BT — a matrix with more than 65535 output columns, such as an LLM vocabulary head, runs. Other grid dimensions above the dispatch limit return an error.
• WebGPU adapter limits — the device requests the adapter's own limits. A buffer can exceed the 256 MiB WebGPU default.
• CPU on other targets — the SIMD-only kernels compile on x86_64 and aarch64 only. wasm32 builds without dead-code and unused-import warnings.

Changed

• WebGPU adapter index — an index past the adapter list returns WgpuError::DeviceError. Before, it opened another adapter. A host that lists no adapters, such as a browser, still opens its default for index 0.
• wasm32 WebGPU — wgpu's fragile-send-sync-non-atomic-wasm feature is on for wasm32, so the WebGPU handles are Send and Sync there.

Removed

• runtime::wgpu::shaders::copy::STRIDED_COPY_SHADER — the strided copy builds its kernels through launch_strided_copy instead.

Full diff: v0.11.0...v0.12.0