KTransformers

Extending Models

Before documenting a new model path, decide whether it is a registry model, a manual SGLang-KT tutorial, or an experimental page.

Minimum Documentation Checklist

  • model family and exact checkpoint
  • package set and version caveats
  • launch entry: kt run or manual python -m sglang.launch_server
  • --kt-method
  • --kt-weight-path expectation
  • CPU ISA/backend requirement
  • GPU memory and tensor parallel assumptions
  • parser or chat-template requirements
  • smoke result or explicit Needs smoke

Registry Candidate

A model is a registry candidate when:

  • the command can be represented by stable defaults;
  • aliases are clear;
  • tensor parallel limits and parser options are known;
  • kt run <alias> --dry-run produces the intended launch command.

Manual Tutorial Candidate

Keep a model as a manual tutorial when it needs:

  • special environment versions;
  • model-specific conversion steps;
  • non-default attention or quantization behavior;
  • hardware that has not been broadly validated.