blass

Basic Linear Algebra Subprogram Simplified

+Basic LLM Inference Engine

To-do list

Tensor

• More comprehensive constructors (for initializer_list for example)
• Getters
• Add stride and offset (easier to handle when data structure changes)
• Flatten, reshape, slicing, transpose
• Constructors
  • Fill
  • Random
• Basic math ops
  • Element-wise
  • Broadcasting
  • Matmul, convolution
• Optimizations for non-contiguous
  • Remove recursion
  • Coalesce (flattening)
  • Permute (find dimension of size 1)
  • Optimize blocking to make it usable
• Random module (distribution, seed, etc.)
• Model loading
  • Parsing GGUF file
  • Tokenizer library API
  • Implementing all NN layers
• Template functions for ops and strides
• Autograd

Perf measurement

• Record useful metrics
• Visualization

Graph compiler (to optimize autograd graph, and maybe inference for ONNX)

• Graph

Usage

To compile main

make

To run all tests & benchmarks

make tests
make bench

To run individual tests & benchmarks

make tests_elemwise

make bench_elemwise
make bench_matmul