The other issue is Pascal cards don't have tensor cores, so there much slower than those with them. You could try Unsloth for 2x faster llama fine-tuning - someone made P40s and P100s work. Although I would suggest upgrading to at least RTX 20x series.