KTransformers

Fine-Tuning Model Tutorials

Use this page to understand the current status of model-level Training routes. Exact Training TPS, commands, and blockers live on the corresponding model pages.

Current Tutorial Candidates

Model familyLLaMA-Factory exampleCurrent KT SFT scope
DeepSeek V2 Litedeepseek_v2_lora_sft_kt.yamlAMXBF16, AMXINT8, AMXINT4; needs smoke before production use.
DeepSeek V3-0324deepseek_v3_lora_sft_kt.yamlAMX SFT only; FP8 source checkpoints must be converted or prepared for the target backend.
Qwen3-235B-A22Bqwen3moe_lora_sft_kt.yamlAMX SFT; validate on the documented hardware configuration before production use.
Qwen3.5-397B-A17Bqwen3_5moe_lora_sft_kt.yamlUse AMXINT8 as the first documented path until BF16/INT4 runs are separately recorded.

Not Current Yet

TopicStatus
Kimi K2 / Kimi K2.5 SFTNot current public support; needs a current LLaMA-Factory path and smoke result.
DPOUnconfirmed for the current KT integration.
Old kt_optimize_rule tutorialsHistorical; need current LLaMA-Factory configs and smoke before use as current guidance.

Complete Tutorial Shape

Complete model tutorials record:

model checkpoint + source precision + target KT backend + conversion step + hardware configuration + launch command + validation result

For DeepSeek V3-family SFT, always explain all three AMX precision options: AMXBF16, AMXINT8, and AMXINT4.