Loss is 0 after 1 step
#42
by
abipani
- opened
I finetuned 7b chat int4 model on 200 sample dataset. It started with loss 2.00 then went to 0 for rest of the steps.
I saved the model and when I trying to inference. It shows me below error.
next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
RuntimeError: probability tensor contains either inf, nan or element < 0
The finetuning procedure seems problematic. Quantized models need QLoRA. We have provided examples in https://github.com/QwenLM/Qwen#usage. Have a try!
jklj077
changed discussion status to
closed