Fix: Higher vram usage for mistral and sample_packing (#691)

Adding qlora config for Mistral (#675)

flash_attention + sample packing for stablelm 3b (#671)

Fix: ValueError when FA + Mistral when padding_side=right (#681)

Feat: Set WORKDIR to /workspace/axolotl (#679)

Fix: Future deprecation warning with use_auth_token (#680)

Fix(tokenizer): Set rstrip,lstrip,norm to False (#678)

Fix(version): Update FA to work with Mistral SWA (#673)

chore: Clean up repetitive model kwargs (#670)

Feat: Allow usage of native Mistral FA when no sample_packing (#669)

Feat: Add config yaml to section for reprod in bug-report.yaml (#667)

refactor to set eval_batch_size earlier if unset, so we can warn if mismatched (#662)

remove patch fix for phi (#664)

prepared dataset caching, other misc fixes (#665)

make sure we also run CI tests when requirements.txt changes (#663)

removed duplicate on requirements.txt (#661)

add mistral e2e tests (#649)

Make dataset_processes configurable (#651)

Fix bug when using pretokenized datasets (#652)

add support for defined train split (#654)

don't strip the prompt for check since we don't strip to tokenize anymore (#650)

fix for flash attn w mistral w/o sammple packing (#648)

Update mistral/README.md (#647)

Feat: Add example for Mistral (#644)

Fix(cfg): Add validation for save_strategy and eval_strategy (#633)

Mistral flash attn packing (#646)

Warn users to login to HuggingFace (#645)

Fix bug in dataset loading (#284)

Correct typos in datasets.py (#639)

skip some flash attn patches unless explicitly enabled (#643)

use fastchat conversations template (#578)

update for recent transformers updates (#636)

attention_mask not needed for training (#642)

eval_table isn't quite stable enough to be in default llama configs (#637)

Feat: Add support for upstream FA2 (#626)

Added quotes to the pip install -e command to fix an incompatibility with shells that do glob expansion like zsh (#632)

Merge pull request #629 from OpenAccess-AI-Collective/chore/-change-default-model

Fix: Fail bf16 check when running on cpu during merge (#631)

default model changed

Feat(doc): Add eval_sample_packing to doc (#625)

tweak: improve base builder for smaller layers (#500)

better handling and logging of empty sharegpt turns (#603)

create a model card with axolotl badge (#624)

update README w deepspeed info (#605)

chore(callback): Remove old peft saving code (#510)

misc fixes to add gptq tests (#621)

split completion text to sequence_len (#616)

run eval on the first step to get a baseline (#617)

let MAX_JOBS use the default since we're not resource constrained on our self-hosted runners (#427)

skip the gpu memory checks if the device is set to 'auto' (#609)