Module kernels 

Kernel implementations: scalar reference, AVX2 SIMD, and CUDA PTX.

Each submodule provides three variants of its kernel:

• fn {name}_scalar(...) — Pure Rust scalar reference (ground truth)
• unsafe fn {name}_avx2(...) — AVX2 SIMD implementation
• fn {name}_ptx() -> &'static str — PTX assembly source string

Modules

absolute_position

Absolute position embeddings kernel.

activation

Activation kernels: ReLU, GELU, SiLU.

adamw

AdamW optimizer kernel.

alibi

ALiBi (Attention with Linear Biases) kernel.

attention

Scaled dot-product attention kernel.

batchnorm

Batch normalization kernel.

bias_add

Bias addition kernel.

cma_es

CMA-ES sampling kernel.

conv1d

1D Convolution kernel.

cross_entropy

Cross-entropy loss kernel with log-softmax.

dropout

Dropout kernel.

embedding

Embedding lookup kernel.

f16_convert

F16 (half-precision) conversion kernel.

flash_attention

Flash Attention: IO-aware tiled attention.

gated_delta_net

Gated Delta Net recurrence kernel.

gelu

GELU kernel (standalone module).

gqa

Grouped Query Attention kernel.

kmeans

K-means clustering kernel.

layernorm

Layer normalization kernel.

lbfgs

L-BFGS two-loop recursion kernel.

linear

Linear projection kernel.

matmul

Matrix multiplication kernel.

ops

Shared kernel primitives: dot product, softmax row, score matrix.

pagerank

PageRank iteration kernel.

rmsnorm

RMSNorm kernel: root mean square layer normalization.

rope

Rotary Position Embedding (RoPE) kernel.

sampling

Sampling algorithms kernel.

silu_standalone

Standalone SiLU kernel with explicit sigmoid.

softmax

Softmax kernel: numerically stable exponential normalization.

ssm

State-Space Model (SSM) scan kernel.

swiglu

SwiGLU gated MLP kernel.

tied_embeddings

Tied embeddings kernel (language model head).

transpose

Matrix transpose kernel: out-of-place B = A^T with AVX2 8×8 micro-kernel.

ulp

ULP (Unit in the Last Place) distance utilities for floating-point comparison.

Enums

Backend

Backend selector for kernel dispatch.