Safe and Efficient Online Softmax Calculation

Understand the Performance Limitations and Gaps

Understanding CuTe Tiled MMA Using an Example

Same Performance as Group-Wise Weight-Only Quantization But with Better Accuracy

Calling cuBLAS GEMM API Correctly

CuTe Shared Memory Swizzling Abstractions

Matrix Transpose CUDA Kernel Implementation Using CuTe

Employing CUTLASS for Accelerated Computing

Mathematical Fundamentals to CUTLASS Computing

TensorRT Acceleration for PyTorch Native Eager Mode Quantization Models

General Matrix Multiplication CUDA Performance Optimization

Motivations for Different Tensor Layouts