Module kernels 

GPU kernels for BitNet quantization operations.

This module provides CubeCL-based GPU kernels for efficient ternary weight x activation matrix multiplication.

Kernels

• absmean_quantize - Quantize weights to ternary {-1, 0, +1}
• ternary_dequantize - Convert ternary back to float
• ternary_matmul_gpu - Optimized ternary matmul (no multiply ops!)
• packed_ternary_matmul - 2-bit packed weights for reduced bandwidth
• bitlinear_forward - Fused LayerNorm + ternary matmul

Feature Gate

Requires the cuda feature to be enabled:

[dependencies]
bitnet-quantize = { version = "0.1", features = ["cuda"] }

Functions

cuda_available

Check if CUDA kernels are available.