LlamaFactory Finetuning CI tests

[sreeram-11]

What the tests validate

1. Create venv

2. Install required packages: huggingface_hub

3. Clone and install LLaMA Factory

4. Verify llamafactory-cli is available

5. Verify key example/config files exist

6. Patch the training YAML for a small CI run

   • Replace lora_rank 8 with lora_rank 6 (smaller LoRA rank, so fewer trainable parameters)
   • output_dir: saves/qwen3_lora_sft_ci (Writes output into a dedicated predictatble folder)
   • Force batch size per device to 1 (reduces memory usage)
   • Force gradient accumulation to 1 (simplifies the run and keeps it short)
   • Force training to 1 epoch (keeps CI fast)
   • Log every step (if CI fails, logs are easier to inspect)
   • Save checkpoints more frequently

7. Run a short LoRA fine-tuning job

8. Verify output artifacts

9. Export LoRA finetuned model

10. Verify export artifacts

Tests not recommended

1. Interactive llamafactory-cli chat
2. GUI / webui
3. Windows CI
   • Playbook is currently supported only on Linux