Expand description
CPU kernels grouped by operation family.
Modulesยง
- attention
- Prefill/decode attention kernels and related attention variants.
- audio
- Audio feature extraction helpers used by speech models.
- conv
- Small convolution and layout-specific convolution helpers.
- fft
- Small FFT helpers, interleaved complex transforms, and normalization.
- fused
- Inference-oriented fused operations that combine common model steps.
- matmul
- GEMM, batched GEMM, ragged batched GEMM, and matvec variants.
- moe
- Mixture-of-experts routing and alignment helpers.
- normalization
- RMS norm, layer norm, group norm, sparsemax, and related reductions.
- positional
- RoPE and positional embedding variants.
- quantization
- Quantize, dequantize, int4 unpacking, fp8 helpers, and quantized matmul.
- recurrent
- Recurrent/state-space preprocessing helpers.
- shape
- Copies, transposes, padding, slicing, tiling, and layout transforms.