* let hf trainer handle torch compile
* remove torch compile checks, include option for backend
* suppress torch errors to get further
* require min torch version of 2.1.0 for torch compile to work
---------
Co-authored-by: Aman Karmani <aman@tmm1.net>
* update readme to point to direct link to runpod template, cleanup install instrucitons
* default install flash-attn and auto-gptq now too
* update readme w flash-attn extra
* fix version in setup
* Add Metharme tokenizing strategy
This strategy accounts for how the Metharme JSONLs are formatted as well as adds duplicated EOS tokens which can help trim model output length.
I haven't gotten the chance to test this yet, and probably won't have the chance for quite a bit, so I'm committing this now.
* Redo Metharme tokenizing strategy
lol
* fix: oops
* Rearrange a conditional
* chore: reformat code in accordance with linter
* chore: Make lint not freak out
* chore: fix lint
---------
Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>
* support user defined prompters, pretokenized datasets in config, local parquet, local arrow files
* fix user defined dataset types
* fix for system prompts
* fix tests
* fix checks for parquet and arrow
* aha moment that d.data_files isn't used
* add documentation for ds_type to add support for parquet and arrow
* flash attn pip
* add packaging
* add packaging to apt get
* install flash attn in dockerfile
* remove unused whls
* add wheel
* clean up pr
fix packaging requirement for ci
upgrade pip for ci
skip build isolation for requiremnents to get flash-attn working
install flash-attn seperately
* install wheel for ci
* no flash-attn for basic cicd
* install flash-attn as pip extras
---------
Co-authored-by: Ubuntu <mgh@mgh-vm.wsyvwcia0jxedeyrchqg425tpb.ax.internal.cloudapp.net>
Co-authored-by: mhenrichsen <some_email@hey.com>
Co-authored-by: Mads Henrichsen <mads@BrbartiendeMads.lan>
Co-authored-by: Wing Lian <wing.lian@gmail.com>
* fix attetion mask with packing
* set position ids and use block diagonal attn mask
* fix expand mask for multiple batch items, make sure we pad position_ids
* don't move masks to cpu
* use multi pack dataloader w random sampler
* add position_ids back
* more fixes for dataloader integration
* est total tokens, fix field loop
* more fixes, position_ids seems broken
* more fixes for sample packing
* use distributed sampler, avoid accelerate prepare
* use accelerator prepare for dataloader
* fix for position_ids w packing
* Update src/axolotl/utils/dataloader.py
* validation for sample packing and doc
* more fixes for 4k and optimizations
* optimized expand mask fn
* better handling of variance in multipack dataloader length and trainer hanging when it runs out of data
* fix rounding of len of batches to int
* better handling so that all devices have the same dataloader len
* fix step calc for packing
* pass sample packing efficiency to training args
* add a test for the mask expansion for sequence packing
* only process eval dataset for packing if not None
* don't split batches when packing
* weighted CE losses
* weighted CEL fixes
* limit packing to sequences of max seq len
* seq_len_multiple for packing
* make sure the chunk size is an int
* sample_packing_seq_len_multiplier config
* use cumulative seq len with var len flash attn v2 w packing
* properly calculate max len
* fix flash-attn, xformers, packing, support chatml
* fix chatml system prompt for openorca, legacy tokenizer opts
* add chatml
* add unit tests for cum seq lens, add ability to build cu_seq_lens from positional ids, fix prompt test
* fix test and pylint checks
* more packing and dataset optimizations and fixes
* filter w multiple cpus
* more fixes and optimizations
* fixes and go back to distributed sampler since batch sampler won't work
* fix counts by accounting for num devices
* fix steps calculation
* previous accelerate is still most performant
* add numba to requirements.
* use custom distributed checks
* fix sampler to prevent overfit w new epochs
* let's not cleanup the cached datasets
* calculate cum seq lens with pos_ids instead of mask, simplify packing params, fix distributed barrier
* speed optimizations and set accelerate fsdp env vars
* optimize dataset concatenation?
* more optimizations for dataset handling
* fix import for annotation
* manual pre-commit fixes
* another sum optimization and bug fix for calc steps
* fix packing estimations
* fix formatting
* pylint problems
* add back flash attention branch for handling unpacked sequences seperately
* Address PR feedback
* add optional sample packing config params to readme
