Add shifted sparse attention (#973) [skip-ci]

Add `layers_to_transform` for `lora_config` (#1118)

Enable or disable bf16 support based on availability (#1116)

keep gate in fp32 for 16 bit loras (#1105)

add gptneox embeddings, fix phi2 inputs, also fix the casting (#1083)

update sharegpt conversations when chatml chat template is set (#1075) [skip ci]

fix: `train_on_inputs: true` ignored for sharegpt (#1045) [skip ci]

be more robust about checking embedding modules for lora finetunes (#1074) [skip ci]

attempt to also run e2e tests that needs gpus (#1070)

fix double eos token for chatml (#1054) [skip ci]

Phi2 rewrite (#1058)

streaming multipack for pretraining dataset (#959)

RL/DPO (#935)

bump transformers and update attention class map name (#1023)

Feat: Warns to add to modules_to_save when adding tokens or switching special_tokens (#787)

fix mistral prompt assembly (#982)

Fix prompt assembly for llama (#952)

Respect sequence_len in config for `type: llama2_chat` (#926)

support for mamba (#915)

Feat(wandb): Refactor to be more flexible (#767)

Feat: Add warmup_ratio (#893)

Phi update 202311 (#876)

add e2e tests for checking functionality of resume from checkpoint (#865)

use temp_dir kwarg instead

missing dunder-init

chore: lint

make sure to cleanup tmp output_dir for e2e tests

simplify by removing duplicate base_model_config (#772)

Fix: Warn when fullfinetune without adapter (#770)

convert exponential notation lr to floats (#771)

Fix: eval table conflict with eval_sample_packing (#769)

remove lora fused packing test (#758)

Implement fused modules (#747)

misc sharegpt fixes (#723)

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

add mistral e2e tests (#649)

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

use fastchat conversations template (#578)

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

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

misc fixes to add gptq tests (#621)

Support Sample packing for phi arch (#586)

E2e device cuda (#575)

e2e testing (#574)

Fix pretraining with iterable/streaming Dataset (#556)

workaround for md5 variations (#533)

recommend padding when using sample packing (#531)

fix test fixture b/c hf trainer tokenization changed (#464)

fix fixture for new tokenizer handling in transformers (#428)

simplify `load_tokenizer`