Module cpu 

CPU backend using Accelerate (macOS) / portable fallback (Linux). Context = () — all ops execute immediately, no batching needed.

Modules

vnext_ops

CPU operators over admitted host buffers and original physical weights. Operator registration is supplied by the CPU composition, not the legacy backend or the test reference device.

vnext_runtime

Host execution for the production plan runtime. CPU commands run in order under a runtime-wide submission lock; errors after execution starts produce failed, quiescent fences, never permission to retry a partially written batch.

Structs

CpuBackend

CpuGptqStore

CPU-side GPTQ store — dequantized f32 weights in row-major [n, k] layout. Trades memory for simplicity: repack once at load, then run normal GEMM.

Enums

CpuQuantStore

CPU-side container for any GGUF k-quant flavour. Each variant holds the dense fp32 weights post-eager-dequant — CPU isn’t the bench target so we don’t pay the complexity of on-the-fly dequant here; the variant tag exists so gemm_quant can route consistently.