Ludwig v0.10.4

What's Changed

• Small typo in dequantization script by @arnavgarg1 in #3993
• Docs: update visualize.py by @eltociear in #4001
• [MAINTENANCE] Most Recent Version of matplotlib breaks ptitprince and seaborn method calls. by @alexsherstinsky in #4007
• [MAINTENANCE] Make the implementation for the fix of the ViTEncoder to ensure that the transformers.ViTModel returns the output_attentions more elegant (and cuts on the amount of code) by @alexsherstinsky in #4008
• Fix mnist source by @mhabedank in #4011
• Support for freezing pretrained vision model layers with regex by @ethanreidel in #3981
• Add Phi-3 Support by @arnavgarg1 in #4014

New Contributors

• @eltociear made their first contribution in #4001
• @mhabedank made their first contribution in #4011

Full Changelog: v0.10.3...v0.10.4

Ludwig v0.10.3

What's Changed

• Replace Slack links with Discord links. by @alexsherstinsky in #3988
• Allow image bytes type during preprocessing by @vijayi1 in #3971
• Fix for 'upload_to_hf_hub()' path mismatch with 'save()' by @sanjaydasgupta in #3977
• Minor change to fix the incorrect response truncation by @amankhandelia in #3986
• Pin minimum transformers to 4.39 to reduce Llama/Gemma memory pressure by @arnavgarg1 in #3976
• Actually add support for RSLoRA and DoRA by @arnavgarg1 in #3984

New Contributors

• @amankhandelia made their first contribution in #3986

Full Changelog: v0.10.2...v0.10.3

Release v0.10.2

What's New

• Add support for RSLoRA and DoRA by @arnavgarg1 in #3948
  To enable, set the corresponding flag to true in config (can be used in conjunction):

    adapter:
        type:  lora
        use_rslora:  false
        use_dora: false

• Add support for eval batch size tuning for LLMs on local backend by @arnavgarg1 in #3957
  To enable, set "eval_batch_size" to "auto" in the trainer section:

    trainer:
        eval_batch_size:  auto

• Enable loading model weights from training checkpoint by @geoffreyangus in #3969
  To enable, pass from_checkpoint=True to LudwigModel.load():

LudwigModel.load(model_dir, from_checkpoint=True)

Full Changelog

• Save ludwig-config with model-weights in output directory by @sanjaydasgupta in #3965
• Add unit tests for image utils unet functions by @vijayi1 in #3921
• fix: Update imdb_genre_prediction dataset yaml to match dataset by @jeffreyftang in #3944
• Fix kube apt source by @noyoshi in #3952
• Temporarily disable expensive text metrics by @arnavgarg1 in #3954
• [MAINTENANCE] Comment Out PyTorch Nightly Test by @alexsherstinsky in #3955
• [BUGFIX] Fixing integration test failures. by @alexsherstinsky in #3959
• [MAINTENANCE] Use latest version of psutil library. by @alexsherstinsky in #3956

New Contributors

• @sanjaydasgupta made their first contribution in #3965

Full Changelog: v0.10.1...v0.10.2

Ludwig v0.10.1

What's Changed

• Fixed a critical bug in Gemma model fine-tuning that prevented the model from learning when to stop generation. This is accomplished by using eos token in target tensor for instruction-tuning by @geoffreyangus in #3945

Full Changelog: v0.10.0...v0.10.1

v0.10.0

What's Changed

• Add Phi-2 to model presets by @arnavgarg1 in #3912
• Add default LoRA target modules for Phi-2 by @arnavgarg1 in #3911
• Add support for prompt lookup decoding during generation by @arnavgarg1 in #3917
• Pin pyarrow to < 15.0.0 by @arnavgarg1 in #3918
• Add unet encoder-decoder and image output feature by @vijayi1 in #3913
• fix: Add Nested quantization check by @jeffkinnison in #3916
• fix typo in save_dequantized_base_model log statement by @arnavgarg1 in #3923
• Add example for base model dequantization/upscaling by @arnavgarg1 in #3924
• fix: Always return a list of quantization bits values from get_quantization by @jeffkinnison in #3926
• fix: set use_reentrant to True to fix Mixtral-7b bug by @geoffreyangus in #3928
• Disabling AdaptionPrompt till PEFT is fixed. by @alexsherstinsky in #3935
• Add default LoRA target modules for Gemma by @arnavgarg1 in #3936
• Pinning transformers to 4.38.1 or above in order to ensure support for Gemma by @alexsherstinsky in #3940
• Ludwig release version change by @alexsherstinsky in #3941

New Contributors

• @vijayi1 made their first contribution in #3913

Full Changelog: v0.9.3...v0.10.0

v0.9.3

What's Changed

• [MAINTENANCE] Use Trusted Publishers credentials instead of User/Password for uploading releases to PyPi by @alexsherstinsky in #3892
• Add support for official microsoft/phi-2 by @arnavgarg1 in #3880
• Ensure correct padding token for Phi and Pythia models by @arnavgarg1 in #3899
• Enable AdaLoRA tests for LLM adapter by @jeffkinnison in #3896
• Cast LLMEncoder output to torch.float32, freeze final layer at init. by @jeffkinnison in #3900
• Enable IA3 adapters in LLMEncoder by @jeffkinnison in #3902
• [Maintenance] Remove torch nightly pin by @arnavgarg1 in #3903
• Pin deepspeed to < 0.13 and pandas to < 2.2.0 by @arnavgarg1 in #3906
• Add batch size tuning for LLMs by @Infernaught in #3871

Full Changelog: v0.9.2...v0.9.3

v0.9.2: Fixes for OOM and other errors in Ludwig 0.9.1

What's Changed

• fix: Handle missing and unexpected keys during LLMEncoder state dict load by @jeffkinnison in #3841
• fix: Add name and description classmethods to IA3Config by @jeffkinnison in #3844
• Improve IA3 long description by @arnavgarg1 in #3845
• fix: Handle missing and unexpected keys during LLMEncoder state dict load, part 2 by @jeffkinnison in #3843
• Update description for max_new_tokens to explain the dynamic setting behavior in our docs by @arnavgarg1 in #3847
• Add default LoRA target modules for Mixtral and Mixtral instruct by @arnavgarg1 in #3852
• QOL: Fail config validation if a user tries to use ECD with a text output feature and an LLM encoder. by @justinxzhao in #3792
• Pin minimum transformers to 4.36 for Mixtral and Phi support by @arnavgarg1 in #3854
• Revert hack that leads to OOM during fine-tuning by @arnavgarg1 in #3858
• Add support for exporting models to Carton by @VivekPanyam in #3797
• [Maintenance] Bump minimum tokenizers to 0.15 by @arnavgarg1 in #3856
• fix: correct typo in FeatureCollection by @dennisrall in #3863
• Convert test main script in algorithm_utils to unit test by @dennisrall in #3864
• Allow hyperopt config to be loaded from a file by @arnavgarg1 in #3865
• fix: unify ludwig training set metadata and hf pad token by @geoffreyangus in #3860
• Add a utility to detect LLM usage in a config by @jeffkinnison in #3869
• Early stop training if model weights have nan or inf tensors by @arnavgarg1 in #3740
• Scrub credentials from model_hyperparameters.json and description.json by @Infernaught in #3866
• [Maintenance] Bump minimum torch version to 2.0.0 by @arnavgarg1 in #3873
• [Maintenance] Fix docker images by pinning ray==2.3.1, daft==0.1.20, unpinning proto, and using torch 2.1.1. by @justinxzhao in #3872
• [BUGFIX] Guard against UnicodeEncodeError when saving validation results in Google Colab environment by @alexsherstinsky in #3875
• Docker image fixes part 2: pin to torch==2.1.0, add dependency for urllib<2 by @arnavgarg1 in #3877
• Add custom prepare_for_trianing logic to ECD model for LLM encoder adapter initialization by @jeffkinnison in #3874
• qol: Fix some lints. by @justinxzhao in #3868
• [Maintenance] Docker Image Fix part 3: fix torchaudio 2.1.0 dependencies by installing libsox-dev and update API by @arnavgarg1 in #3879
• Add streaming support for zero shot inference by @arnavgarg1 in #3878
• [Maintenance] Remove torchdata pin for nightly install by @arnavgarg1 in #3855
• Add per-step token utilization to tensorboard and progress tracker. by @justinxzhao in #3867
• Set use_reentrant to False for gradient checkpointing by @arnavgarg1 in #3882
• [BUGFIX] Pinning torch nightly to January 13, 2024 to avoid AttributeError by @alexsherstinsky in #3885

New Contributors

• @VivekPanyam made their first contribution in #3797

Full Changelog: v0.9.1...v0.9.2

v0.9.1

What's Changed

• fix: Handle missing and unexpected keys during LLMEncoder state dict load by @jeffkinnison in #3841
• fix: Add name and description classmethods to IA3Config by @jeffkinnison in #3844
• Improve IA3 long description by @arnavgarg1 in #3845
• bump ludwig version by @geoffreyangus in #3846

Full Changelog: v0.9...v0.9.1

v0.9: Mixtral, Phi, Zephyr, and text classification for LLMs

What's Changed

• int: Rename original combiner_registry to combiner_config_registry, update decorator name by @ksbrar in #3516
• Add mechanic to override default values for generation during model.predict() by @justinxzhao in #3520
• [feat] Support for numeric date feature inputs by @jeffkinnison in #3517
• Add new sythesized response column for text output features during postprocessing by @arnavgarg1 in #3521
• Disable flaky twitter bots dataset loading test. by @justinxzhao in #3439
• Add test that verifies that the generation config passed in at model.predict() is used correctly. by @justinxzhao in #3523
• Move loss metric to same device as inputs by @Infernaught in #3522
• Add comment about batch size tuning by @arnavgarg1 in #3526
• Ensure user sets backend to local w/ quantization by @Infernaught in #3524
• README: Update LLM fine-tuning config by @arnavgarg1 in #3530
• Revert "Ensure user sets backend to local w/ quantization (#3524)" by @tgaddair in #3531
• Improve observability during LLM inference by @arnavgarg1 in #3536
• [bug] Pin pydantic to < 2.0 by @jeffkinnison in #3537
• [bug] Support preprocessing datetime.date date features by @jeffkinnison in #3534
• Remove obsolete prompt tuning example. by @justinxzhao in #3540
• Add Ludwig 0.8 notebook to the README by @arnavgarg1 in #3542
• Add effective_batch_size to auto-adjust gradient accumulation by @tgaddair in #3533
• Refactor evaluation metrics to support decoded generated text metrics like BLEU and ROUGE. by @justinxzhao in #3539
• Fix sequence generator test. by @justinxzhao in #3546
• Revert "Add Cosine Annealing LR scheduler as a decay method (#3507)" by @justinxzhao in #3545
• Set default max_sequence_length to None for LLM text input/output features by @arnavgarg1 in #3547
• Add skip_all_evaluation as a mechanic to skip all evaluation. by @justinxzhao in #3543
• Roll-forward with fixes: Fix interaction between scheduler.step() and gradient accumulation steps, refactor schedulers to use LambdaLR, and add cosine annealing LR scheduler as a decay method. by @justinxzhao in #3555
• fix: Move model to the correct device for eval by @jeffkinnison in #3554
• Report loss in tqdm to avoid log spam by @tgaddair in #3559
• Wrap each metric update in try/except. by @justinxzhao in #3562
• Move DDP model to device if it hasn't been wrapped yet by @tgaddair in #3566
• ensure that there are enough colors to match the score index in visua… by @thelinuxkid in #3560
• Pin Transformers to 4.31.0 by @arnavgarg1 in #3569
• Add test to show global_max_sequence_length can never exceed an LLMs context length by @arnavgarg1 in #3548
• WandB: Add metric logging support on eval end and epoch end by @arnavgarg1 in #3586
• schema: Add prompt validation check by @ksbrar in #3564
• Unpin Transformers for CodeLlama support by @arnavgarg1 in #3592
• Add support for Paged Optimizers (Adam, Adamw), 8-bit optimizers, and new optimizers: LARS, LAMB and LION by @arnavgarg1 in #3588
• FIX: Failure in TabTransformer Combiner Unit test by @jimthompson5802 in #3596
• fix: Move target tensor to model output device in check_module_parameters_updated by @jeffkinnison in #3567
• Allow user to specify huggingface link or local path to pretrained lora weights by @Infernaught in #3572
• Add codellama to tokenizer list for set_pad_token by @Infernaught in #3598
• Set default eval batch size to 2 for LLM fine-tuning by @arnavgarg1 in #3599
• [CI] Explicitly set eval batch size in determinism tests, introduce a new integration test group, and exclude slow tests. by @justinxzhao in #3590
• [CI] Run sudo apt-get update in GHAs. by @justinxzhao in #3608
• Store steps_per_epoch in Trainer by @hungcs in #3601
• Updated characters, underscore and comma preprocessors to be TorchScriptable. by @martindavis in #3602
• [CI] Deflake: Explicitly set eval batch size for mlflow test. by @justinxzhao in #3612
• Fix registration for char error rate. by @justinxzhao in #3604
• fix: Load 8-bit quantized models for eval after fine-tuning by @jeffkinnison in #3606
• Add Code Alpaca and Consumer Complaints Datasets by @connor-mccorm in #3611
• Add support for gradient checkpointing for LLM fine-tuning by @arnavgarg1 in #3613
• Bump min support transformers to 4.33.0 by @tgaddair in #3616
• [CI] Fix failing tests on master by @arnavgarg1 in #3617
• Eliminate short-circuiting for loading from local by @Infernaught in #3600
• Refactor integration tests into matrix by @tgaddair in #3618
• fix: Check underlying model device type when moving 8-bit quantized models to GPU at eval by @jeffkinnison in #3622
• Fixed range validation for text generation penalty parameters by @tgaddair in #3623
• Update comment for predict to update Ludwig docs by @Infernaught in #3535
• Avoid deprecation warnings on pandas Series.fillna by @carlogrisetti in #3631
• QoL: Default to using fast tokenizer for Llama models by @arnavgarg1 in #3625
• fixed typo in EfficientNet's model variant from v2_ to v2_s by @saad-palapa in #3628
• Add pytorch profiler and additional tensorboard logs for GPU memory usage. by @justinxzhao in #3607
• Pin minimum transformers version to 4.33.2 by @arnavgarg1 in #3637
• Add function to free GPU memory by @Infernaught in #3643
• ❗ Enable LLM fine-tuning tests when no quantization is specified by @arnavgarg1 in #3626
• Add check to ensure selected backend works with quantization for LLMs by @arnavgarg1 in #3646
• [CI] Use a torch-nightly-compatible version of torchaudio by @justinxzhao in #3644
• Set do_sample default to True by @Infernaught in #3641
• FIX: Failure in audio feature related test by @jimthompson5802 in #3651
• Remove unnecessary peft config updating by @Infernaught in #3642
• FIX: docker build error for ludwig-gpu by @jimthompson5802 in #3658
• Exclude getdaft on Windows by @carlogrisetti in #3629
• Add daft back for windows since the wheels are now officially published by @arnavgarg1 in #3663
• fix: The final batch of an epoch is skipped when batch size is 1 by @jeffkinnison in #3653
• Place metric functions for BLEU and Rogue on correct devices when using multiple GPUs by @arnavgarg1 in #3671
• Remove duplicate metrics by @Infernaught in #3670
• Increment epochs based on last_batch() instead of at the end of the train loop. by @justinxzhao in #3668
• [FEATURE] Support Merging LoRA Weights Into Base Model (Issue-3603) by @alexsherstinsky in #3649
• [FEATURE] Include Mistral-7B model in list of supported base models by @alexsherstinsky in #3674
• [MAINTENANCE] Partially reconcile type hints, fix some warnings, and fix comments in parts of the codebase. by @alexsherstinsky in #3673
• Improve error message for when an LLM base model can't be loaded. by @justinxzhao in #3675
• Fix eos_token and pad_token issue by @Infernaught in https://g...

v0.8.6

What's Changed

• Add consumer complaints generation dataset by @connor-mccorm in #3685
• Set the metadata only during first training run by @Infernaught in #3684
• Add ability to upload Ludwig models to Predibase. by @martindavis in #3687
• Log additional per-GPU information in model metadata files and GPU utilization on tensorboard. by @justinxzhao in #3712
• QoL: Only log generation config being used once at inference time by @arnavgarg1 in #3715
• [MAINTENANCE] Adding typehint annotations in backend and data components and fixing mypy errors. by @alexsherstinsky in #3709
• QoL: Limit top-level trainer logging messages such as saving model or resuming model training to main coordinator process by @arnavgarg1 in #3718
• Add sample_size as a global preprocessing parameter by @Infernaught in #3650
• QOL: Update recommended vscode settings. by @justinxzhao in #3717
• Add new fine-tuning notebooks to README by @arnavgarg1 in #3722
• Dynamically set max_new_tokens based on output feature length, GMSL and model window size by @arnavgarg1 in #3713
• Fix issue while logging cuda device utilization to tensorboard by @arnavgarg1 in #3727

Full Changelog: v0.8.5...v0.8.6