Ratchet, Luminal, Ollama

• https://github.com/huggingface/ratchet

  • https://github.com/jafioti/luminal
    • https://github.com/ollama/ollama

• Just skim papers to know the domain/keywords/syllabus of efficient ML space : https://www.dropbox.com/scl/fo/gdbcm9bhnhwkk1a4gcpu7/ADU3sxKylI0DE5RcD8aFisw?rlkey=bvv1gtyjp8bowpi8nob07sw17&st=xczjastw&dl=0

  • Domains:
    • LORA
      • quantization
        • pruning
        • synthetic data
        • small models in the range of 1B-15B parameters
        • ml compilers
        • superoptimizers
        • compute operation optimization: GPU code etc etc

• Can Luminal’s composable compilers approach be built in combination with Ratchet’s unified graph execution (Ahead-Of-Time +Just-In-Time): jittor, relay?

• Dynamic linking in web environment -> connecting dots with copy and patch compilation paper