Agnostic cloud gpu docker image and Jupyter lab (#1097)

Create preprocess CLI (#785)

Warn users to login to HuggingFace (#645)

refactor scripts/finetune.py into new cli modules (#550)

check for the existence of the default accelerate config that can create headaches (#561)

fix wandb so mypy doesn't complain (#562)

move is_llama_derived_model into normalize_config (#524)

Debug tokenization output: Add ability to output text only (no tokens), and/or specify num samples to see (#511)

split train from other cli options (#503)

customizable ascii art (#506)

tweak: use default config file when only one file is present (#501)

Refactor train cfg cli (#499)

fix: inference did not move the model to the correct device (#483)

ReLoRA implementation (with quantization) (#322)

feat(docs): improve user customized prompts (#443)

Ax art (#405)

add utils.data.prepare_dataset

use context manager to run things on rank0 before others (#397)

Feat(config): add max steps (#387)

save tokenizer before training starts (#380)

simplify `load_tokenizer`

improve GPU logging to break out pytorch cache and system mem

extract module for working with cfg

Attention mask and position id fixes for packing (#285)

Fix(save): Save as safetensors (#363)

feat(merge): save tokenizer on merge (#362)

Merge pull request #356 from tmm1/load_model-args

simplify load_model signature

log GPU memory usage

fix FSDP save of final model (#329)

add runpod envs to .bashrc, fix bnb env (#316)

misc fixes

Fixed pre-commit problems, fixed small bug in logging_config to handle LOG_LEVEL env var

Adding logging enhancement

Merge pull request #92 from OpenAccess-AI-Collective/flash-optimum

chore: Refactor inf_kwargs out

Merge branch 'main' into flash-optimum

Merge pull request #177 from NanoCode012/fix/landmark-patch

Merge pull request #159 from AngainorDev/patch-1

Fix strict and Lint

Fix set mem_id for inference and refactor

Set mem cache args on inference

fix formatting

pass a prompt in from stdin for inference

Update scripts/finetune.py

address PR feedback

add streaming dataset support for pretraining datasets

more tweaks to do pre-training with bettertransformers

experimental expansion of ctx len

add flash attn context for efficient training and attempt setting model to train mode: