change model loading mechanism

do not allow using base model directly, seems that it will break things

test new defaults

update ui

inference beams default to 2 since it’s better

ok

why

fix wrong Pillow version on colab

try

Revert "try to fix AttributeError: module 'PIL.Image' has no attribute 'Resampling'"

try to fix AttributeError: module 'PIL.Image' has no attribute 'Resampling'

update README.md

Merge branch 'dev-2'

rename project

add templates/user_and_ai.json

move to latest working version of peft

cache lora models

load base model on start

fix inference prompt preview not updated to the latest version

resolve generation canceling issue

report status while finetune starting

auto select prompt template for model

save train info with model

clear cache after training

update README.md

fix “AttributeError: module 'PIL.Image' has no attribute 'Resampling'”

cache loaded_base_model_with_lora

update requirements.lock.txt

latest peft makes model.save_pretrained in finetune.py save a 443 B adapter_model.bin which is clearly incorrect (normally adapter_model.bin should > 16 MB)

update requirements.lock.txt

update requirements.lock.txt

try to lock requirements

fix

Merge branch 'main' of github.com:zetavg/llama-lora

this should be set for training

Update README.md

some fixes

update README.md

Merge branch 'dev-2'

update README.md and LLaMA_LoRA.ipynb

update

update finetune

simulate actual behavior in ui dev mode

update default branch to main

fix possible error

fix “RuntimeError: expected scalar type Half but found Float” on lambdalabs and hf

update

add cancel timeout buttons and update ui

update ui

ui fix