KTransformers Documentation

KTransformers is a CPU-GPU heterogeneous computing project for large MoE model inference and LoRA fine-tuning. The documentation follows a task-first structure: install the right package path, choose inference or fine-tuning, run a model, then move into model-specific tutorials, technical background, hardware boundaries, and command references.

Current Public Surface

Task	Public packages	Primary entry
Inference serving	`kt-kernel sglang-kt`	`kt run` or `python -m sglang.launch_server` with `--kt-*` arguments
LoRA SFT	`ktransformers[sft]` through LLaMA-Factory	LLaMA-Factory training YAML with `use_kt: true` and an Accelerate KT config

Older local_chat.py, ktransformers/server/main.py, balance_serve, and kt_optimize_rule paths are historical unless a page explicitly marks them as revalidated.

Getting Started

Installation - choose the package set for inference or fine-tuning.
First inference server - start from kt run or manual SGLang-KT launch.
First LoRA SFT run - start from LLaMA-Factory KT examples.

Inference

Inference overview - serving paths and method selection.
Launch a server - choose between registry-driven and manual launches.
Sending requests - use cURL, Python requests, or the OpenAI client.
OpenAI-compatible API - endpoint and client behavior for SGLang-KT.
Popular model usage - where to start for DeepSeek, Kimi, MiniMax, Qwen, and GLM.

Fine-Tuning

Fine-tuning overview - LoRA SFT as a first-class KTransformers workflow.
LoRA SFT with LLaMA-Factory - current public SFT entry and config shape.
SFT backends and precision - AMXBF16, AMXINT8, and AMXINT4.
Weight preparation - BF16, INT8, INT4, and DeepSeek V3 FP8 source checkpoints.
DeepSeek SFT and Qwen SFT - current model tutorial status.

Advanced Features

Server arguments - KT-specific launch parameters and tuning rules.
Precision and quantization - BF16, FP8, RAWINT4, AMXINT4, AMXINT8, MXFP4, and LLAMAFILE.
Expert placement - GPU expert count, deferred experts, and dynamic updates.
AMX backend - AMX architecture, weight conversion, and launch flow.
Layerwise Prefill - long-context prefill acceleration principles and tuning strategy.

Supported Models and Platforms

Support matrix - model, precision, backend, and validation status.
Text generation models - model-family entry points.
Model status policy - how support claims are written.
Hardware platform status - CPU/GPU requirements and known platform boundaries.

Technical Work

Technical work - system background separated from user task guides.
Heterogeneous inference - CPU-GPU MoE execution direction.
Local fine-tuning - local inference to local LoRA SFT.
Talks and slides - GOSIM 2026 and public decks.
GitHub docs migration map - what moves from GitHub docs and what stays historical.

Developer and Command Reference

Runtime smoke checklist - what must be verified before upgrading a support claim.
Benchmark and profiling - repeatable reporting expectations.
CLI reference - kt command surface.
Troubleshooting - common install, serving, and SFT failures.