Getting an error when trying to perform SFT on Tiny Llama

[survivebycoding]

We are getting this error when trying to execute SFT on TinyLlama:
[rank0]: if f.read(7) == "version": [rank0]: File "/usr/lib/python3.10/codecs.py", line 322, in decode [rank0]: (result, consumed) = self._buffer_decode(data, self.errors, final) [rank0]: UnicodeDecodeError: 'utf-8' codec can't decode byte 0xc1 in position 70: invalid start byte

However, we found no issue executing llama and mistral. If you have any idea regarding this issue, please let us know

[songmzhang]

It is due to the version of transformers library. We also found this issue in our early experiments. The version suggested by TinyLLaMA is 4.31, while we loaded it with transformers == 4.38 and also found this error.

We conjecture that it may be because some weights of LlamaForCausalLM were not initialized from the model checkpoint.

Here is our solution:

1. Create another environment with transformers == 4.31
2. Load the model checkpoint and re-save it into another directory (noted as tinyllama_new), i.e., model = AutoModelForCausalLM.from_pretrained(original_tinyllama_path) and model.save_pretrained(tinyllama_new, safe_serialization=False).
3. Switch to the original environment for this project and load the model checkpoint from tinyllama_new.

[songmzhang]

Thanks for reminding us of this issue and we will add this to README.md.

[survivebycoding]

original_tinyllama_path - is this the path in our system where we have downloaded tiny llama?

[songmzhang]

original_tinyllama_path - is this the path in our system where we have downloaded tiny llama?

Yes.

[survivebycoding]

The new llama folder created does not have the same files.

Getting the error"
[OSError: Can't load tokenizer for '/tinyllama/tinyllama-1.1b-3T'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure '/tinyllama/tinyllama-1.1b-3T' is the correct path to a directory containing all relevant files for a LlamaTokenizerFast tokenizer.]

[songmzhang]

The new llama folder created does not have the same files.

Getting the error" [OSError: Can't load tokenizer for '/tinyllama/tinyllama-1.1b-3T'. If you were trying to load it from 'https://huggingface.co/models', make sure you don't have a local directory with the same name. Otherwise, make sure '/tinyllama/tinyllama-1.1b-3T' is the correct path to a directory containing all relevant files for a LlamaTokenizerFast tokenizer.]

You need to copy tokenizer files (i.e., special_tokens_map.json, tokenizer_config.json, tokenizer.json, tokenizer.model) to the new directory to load the tokenizer.

[survivebycoding]

Full finetuning eval scripts

are under gpt2 folder of scripts and for lora its under tiny llama scripts? Or is lora only applicable to tiny llama?

This part is a bit confusing, maybe you can clarify a bit what ckpt_path and lora_adapter_path is.... also, what is usually given in eval_batch_size?

[survivebycoding]

In vanilla_KD_tinyllama [TEACHER_PEFT_PATH="path_to_teacher_sft_lora_ckpt"]
is the path to the folder called epoch9 right?

[songmzhang]

Full finetuning eval scripts are under gpt2 folder of scripts and for lora its under tiny llama scripts? Or is lora only applicable to tiny llama?

This part is a bit confusing, maybe you can clarify a bit what ckpt_path and lora_adapter_path is.... also, what is usually given in eval_batch_size?

No, run_eval_lora.sh is not only applicable to TinyLLaMA. The scripts and README.md are just for re-implementing the experiments in our paper. Actually, you can also evaluate other fully fine-tuned checkpoints besides GPT2 with run_eval.sh.

Here, CKPT_PATH means the path of the full fine-tuned checkpoint. For example, in your case, it is the path of "epoch9_step...".

Similarly, LORA_ADAPTER_PATH is the path of the LoRA adapter, whose name has the same format as the full fine-tuned ckpt like "epoch9_step...".

For EVAL_BATCH_SIZE, it depends on the available GPU memory and the amount of the model parameters. For example, we use 32 for GPT2-base and 8 for TinyLLaMA.

[songmzhang]

In vanilla_KD_tinyllama [TEACHER_PEFT_PATH="path_to_teacher_sft_lora_ckpt"] is the path to the folder called epoch9 right?

Yes.