TODO

• gpu support

  • gpu weight loading, cpu is limited by system memory
  • hybrid cpu/gpu computing

• Implement the tokenizers : BPE / SentencePiece

• Model uncensoring in the fly

• a npmjs package

• cpu clustering

• gpu glsl shaders calculations support pas