Closed xjdeng closed 1 year ago
Can you check if it works now?
Can you check if it works now?
Yep, it's able to get past that but now I'm getting a new error. Should I create a new issue?
Traceback (most recent call last): File “/content/drive/MyDrive/llm/text-generation-webui/modules/training.py”, line 190, in do_train lora_model = get_peft_model(shared.model, config) File “/usr/local/lib/python3.9/dist-packages/peft/mapping.py”, line 145, in get_peft_model return MODEL_TYPE_TO_PEFT_MODEL_MAPPING[peft_config.task_type](model, peft_config) File “/usr/local/lib/python3.9/dist-packages/peft/peft_model.py”, line 514, in init super().init(model, peft_config) File “/usr/local/lib/python3.9/dist-packages/peft/peft_model.py”, line 79, in init self.base_model = LoraModel(peft_config, model) File “/usr/local/lib/python3.9/dist-packages/peft/tuners/lora.py”, line 118, in init self._find_and_replace() File “/usr/local/lib/python3.9/dist-packages/peft/tuners/lora.py”, line 181, in _find_and_replace raise ValueError( ValueError: Target modules [‘q_proj’, ‘v_proj’] not found in the base model. Please check the target modules and try again.
It's possible that it's only related to distilgpt2.. i'll try it with another model in the meantime..
Edit: Tried it with Llama-7B and I ran out of ram (physical ram, not GPU ram) on my free colab instance with 12.7 GB. Is this normal? On an unrelated note, I've been unable to load a lot of models that have supposedly lower physical ram requirements (like GPT-4chan which should only need 4-5 GB and yet, I'm running out of ram trying to load them on colab).
Can you check if it works now?
It's working I'm training at the moment
Edit: Tried it with Llama-7B and I ran out of ram (physical ram, not GPU ram) on my free colab instance with 12.7 GB. Is this normal? On an unrelated note, I've been unable to load a lot of models that have supposedly lower physical ram requirements (like GPT-4chan which should only need 4-5 GB and yet, I'm running out of ram trying to load them on colab).
My machine used all 64gb of ram before training and about 10gb more vram while training 7b, its at 19gb of vram atm
Training works now with gpt-neo-125M in colab!
Describe the bug
I'm unable to train any LoRAs because it keeps thinking whatever input I have in the Dataset field is "None.txt" regardless of what I select there.
Is there an existing issue for this?
Reproduction
Download distilgpt2 to the models folder (I'm testing the LoRA training capabilities so I decided to start with a small model as a test. This may not be necessary and you can try it with the model of your choice.)
Download https://github.com/tloen/alpaca-lora/blob/main/alpaca_data_cleaned.json to the training/datasets/ folder
Launch with !python server.py --share --load-in-8bit --model distilgpt2 (or whatever model you have downloaded)
Go to the Training tab
Pick the alpaca_data_cleaned Dataset
Hit "Start LoRA Training"
Screenshot
Logs
System Info