Daniel Han comments

Results 781 comments of


                                            Daniel Han

GGUF breaks - llama-3

Update: Hi so I managed to test HF -> llama.cpp without Unsloth to remove Unsloth from the picture. 1. '\n\n' is tokenized as [1734, 1734], unless if I prompted it...

GGUF breaks - llama-3

It should be fixed!

Kaggle Notebook fails to train Llama3 on T4 GPU

Currently we do not support multi GPU - use our Kaggle Llama-3 notebook: https://www.kaggle.com/code/danielhanchen/kaggle-llama-3-8b-unsloth-notebook as is - it does not work on 2x T4s

Kaggle Notebook fails to train Llama3 on T4 GPU

Oh 2x Tesla T4 is the option!

Kaggle Notebook fails to train Llama3 on T4 GPU

Just select it

llama3 sft trainer configuration error when max_steps=None

Oh :( Sorry on the issue

RuntimeError: Triton Error [CUDA]: device kernel image is invalid

Ok great! sorry on the issue!

No module named 'triton.third_party'

Oh great you solved it!

Does unsloth/Phi-3-mini-128k-instruct model exist?

Ye sadly not 128K yet - on the roadmap though! It's sadly not RoPE but some other scaling mechanism

config.json file not found, fine tuning llama3 with unsloth, after saving the file to hugging face

Oh `adapter_config.json` is the `config.json` equivalent. If you're looking for Ooba inference or GGUF, please use our saving to 16bit instead