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 family | LLaMA-Factory example | Current KT SFT scope |
|---|---|---|
| DeepSeek V2 Lite | deepseek_v2_lora_sft_kt.yaml | AMXBF16, AMXINT8, AMXINT4; needs smoke before production use. |
| DeepSeek V3-0324 | deepseek_v3_lora_sft_kt.yaml | AMX SFT only; FP8 source checkpoints must be converted or prepared for the target backend. |
| Qwen3-235B-A22B | qwen3moe_lora_sft_kt.yaml | AMX SFT; validate on the documented hardware configuration before production use. |
| Qwen3.5-397B-A17B | qwen3_5moe_lora_sft_kt.yaml | Use AMXINT8 as the first documented path until BF16/INT4 runs are separately recorded. |
Not Current Yet
| Topic | Status |
|---|---|
| Kimi K2 / Kimi K2.5 SFT | Not current public support; needs a current LLaMA-Factory path and smoke result. |
| DPO | Unconfirmed for the current KT integration. |
Old kt_optimize_rule tutorials | Historical; 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.