quantization

Modules

modelopt.torch.kernels.quantization.attention	Quantization-specific attention kernel pieces.
modelopt.torch.kernels.quantization.common	Shared composable Triton JIT fake-quantization functions.
modelopt.torch.kernels.quantization.conv	Implicit-GEMM CUDA kernel for quantized 3D convolution.
modelopt.torch.kernels.quantization.gemm	Triton quantization kernels.

Quantization kernels: conv (implicit GEMM) and gemm (tensor_quant + Triton FP4/FP8).