Due to some nuances with the phi code, please use deepspeed when training phi for full finetune.
accelerate launch -m axolotl.cli.train examples/phi/phi-ft.yml --deepspeed deepspeed_configs/zero1.json # OR python -m axolotl.cli.train examples/phi/phi-qlora.yml