v0.34.1 Patchfix

Bug fixes

• Fixes an issue where processed DataLoaders could no longer be pickled in #3074 thanks to @byi8220
• Fixes an issue when using FSDP where default_transformers_cls_names_to_wrap would separate _no_split_modules by characters instead of keeping it as a list of layer names in #3075

Full Changelog: v0.34.0...v0.34.1

v0.34.0: StatefulDataLoader Support, FP8 Improvements, and PyTorch Updates!

Dependency Changes

• Updated Safetensors Requirement: The library now requires safetensors version 0.4.3.
• Added support for Numpy 2.0: The library now fully supports numpy 2.0.0

Core

New Script Behavior Changes

• Process Group Management: PyTorch now requires users to destroy process groups after training. The accelerate library will handle this automatically with accelerator.end_training(), or you can do it manually using PartialState().destroy_process_group().
• MLU Device Support: Added support for saving and loading RNG states on MLU devices by @huismiling
• NPU Support: Corrected backend and distributed settings when using transfer_to_npu, ensuring better performance and compatibility.

DataLoader Enhancements

• Stateful DataDataLoader: We are excited to announce that early support has been added for the StatefulDataLoader from torchdata, allowing better handling of data loading states. Enable by passing use_stateful_dataloader=True to the DataLoaderConfiguration, and when calling load_state() the DataLoader will automatically be resumed from its last step, no more having to iterate through passed batches.
• Decoupled Data Loader Preparation: The prepare_data_loader() function is now independent of the Accelerator, giving you more flexibility towards which API levels you would like to use.
• XLA Compatibility: Added support for skipping initial batches when using XLA.
• Improved State Management: Bug fixes and enhancements for saving/loading DataLoader states, ensuring smoother training sessions.
• Epoch Setting: Introduced the set_epoch function for MpDeviceLoaderWrapper.

FP8 Training Improvements

• Enhanced FP8 Training: Fully Sharded Data Parallelism (FSDP) and DeepSpeed support now work seamlessly with TransformerEngine FP8 training, including better defaults for the quantized FP8 weights.
• Integration baseline: We've added a new suite of examples and benchmarks to ensure that our TransformerEngine integration works exactly as intended. These scripts run one half using 🤗 Accelerate's integration, the other with raw TransformersEngine, providing users with a nice example of what we do under the hood with accelerate, and a good sanity check to make sure nothing breaks down over time. Find them here
• Import Fixes: Resolved issues with import checks for the Transformers Engine that has downstream issues.
• FP8 Docker Images: We've added new docker images for TransformerEngine and accelerate as well. Use docker pull huggingface/accelerate@gpu-fp8-transformerengine to quickly get an environment going.

torchpippy no more, long live torch.distributed.pipelining

• With the latest PyTorch release, torchpippy is now fully integrated into torch core, and as a result we are exclusively supporting the PyTorch implementation from now on
• There are breaking examples and changes that comes from this shift. Namely:
  • Tracing of inputs is done with a shape each GPU will see, rather than the size of the total batch. So for 2 GPUs, one should pass in an input of [1, n, n] rather than [2, n, n] as before.
  • We no longer support Encoder/Decoder models. PyTorch tracing for pipelining no longer supports encoder/decoder models, so the t5 example has been removed.
  • Computer vision model support currently does not work: There are some tracing issues regarding resnet's we are actively looking into.
• If either of these changes are too breaking, we recommend pinning your accelerate version. If the encoder/decoder model support is actively blocking your inference using pippy, please open an issue and let us know. We can look towards adding in the old support for torchpippy potentially if needed.

Fully Sharded Data Parallelism (FSDP)

• Environment Flexibility: Environment variables are now fully optional for FSDP, simplifying configuration. You can now fully create a FullyShardedDataParallelPlugin yourself manually with no need for environment patching:

from accelerate import FullyShardedDataParallelPlugin
fsdp_plugin = FullyShardedDataParallelPlugin(...)

• FSDP RAM efficient loading: Added a utility to enable RAM-efficient model loading (by setting the proper environmental variable). This is generally needed if not using accelerate launch and need to ensure the env variables are setup properly for model loading:

from accelerate.utils import enable_fsdp_ram_efficient_loading, disable_fsdp_ram_efficient_loading
enable_fsdp_ram_efficient_loading()

• Model State Dict Management: Enhanced support for unwrapping model state dicts in FSDP, making it easier to manage distributed models.

New Examples

• Configuration and Models: Improved configuration handling and introduced a configuration zoo for easier experimentation. You can learn more here. This was largely inspired by the axolotl library, so very big kudos to their wonderful work
• FSDP + SLURM Example: Added a minimal configuration example for running jobs with SLURM and using FSDP

Bug Fixes

• Fix bug of clip_grad_norm_ for xla fsdp by @hanwen-sun in #2941
• Explicit check for step when loading the state by @muellerzr in #2992
• Fix find_tied_params for models with shared layers by @qubvel in #2986
• clear memory after offload by @SunMarc in #2994
• fix default value for rank size in cpu threads_per_process assignment logic by @rbrugaro in #3009
• Fix batch_sampler maybe None error by @candlewill in #3025
• Do not import transformer_engine on import by @oraluben in #3056
• Fix torchvision to be compatible with torch version in CI by @SunMarc in #2982
• Fix gated test by @muellerzr in #2993
• Fix typo on warning str: "on the meta device device" -> "on the meta device" by @HeAndres in #2997
• Fix deepspeed tests by @muellerzr in #3003
• Fix torch version check by @muellerzr in #3024
• Fix fp8 benchmark on single GPU by @muellerzr in #3032
• Fix typo in comment by @zmoki688 in #3045
• Speed up tests by shaving off subprocess when not needed by @muellerzr in #3042
• Remove skip_first_batches support for StatefulDataloader and fix all the tests by @muellerzr in #3068

New Contributors

• @byi8220 made their first contribution in #2957
• @alex-jw-brooks made their first contribution in #2959
• @XciD made their first contribution in #2981
• @hanwen-sun made their first contribution in #2941
• @HeAndres made their first contribution in #2997
• @yitongh made their first contribution in #2966
• @qubvel made their first contribution in #2986
• @rbrugaro made their first contribution in #3009
• @candlewill made their first contribution in #3025
• @siddk made their first contribution in #3047
• @oraluben made their first contribution in #3056
• @tmm1 made their first contribution in #3055
• @zmoki688 made their first contribution in #3045

Full Changelog:

• Require safetensors>=0.4.3 by @byi8220 in #2957
• Fix torchvision to be compatible with torch version in CI by @SunMarc in #2982
• Enable Unwrapping for Model State Dicts (FSDP) by @alex-jw-brooks in #2959
• chore: Update runs-on configuration for CI workflows by @XciD in #2981
• add MLU devices for rng state saving and loading. by @huismiling in #2940
• remove .md to allow proper linking by @nbroad1881 in #2977
• Fix bug of clip_grad_norm_ for xla fsdp by @hanwen-sun in #2941
• Fix gated test by @muellerzr in #2993
• Explicit check for step when loading the state by @muellerzr in #2992
• Fix typo on warning str: "on the meta device device" -> "on the meta device" by @HeAndres in #2997
• Support skip_first_batches for XLA by @yitongh in #2966
• clear memory aft...

v0.33.0: MUSA backend support and bugfixes

MUSA backend support and bugfixes

Small release this month, with key focuses on some added support for backends and bugs:

• Support MUSA (Moore Threads GPU) backend in accelerate by @fmo-mt in #2917
• Allow multiple process per device by @cifkao in #2916
• Add torch.float8_e4m3fn format dtype_byte_size by @SunMarc in #2945
• Properly handle Params4bit in set_module_tensor_to_device by @matthewdouglas in #2934

What's Changed

• [tests] fix bug in torch_device by @faaany in #2909
• Fix slowdown on init with device_map="auto" by @muellerzr in #2914
• fix: bug where multi_gpu was being set and warning being printed even with num_processes=1 by @HarikrishnanBalagopal in #2921
• Better error when a bad directory is given for weight merging by @muellerzr in #2852
• add xpu device check before moving tensor directly to xpu device by @faaany in #2928
• Add huggingface_hub version to setup.py by @nullquant in #2932
• Correct loading of models with shared tensors when using accelerator.load_state() by @jkuntzer in #2875
• Hotfix PyTorch Version Installation in CI Workflow for Minimum Version Matrix by @yhna940 in #2889
• Fix import test by @muellerzr in #2931
• Consider pynvml available when installed through the nvidia-ml-py distribution by @matthewdouglas in #2936
• Improve test reliability for Accelerator.free_memory() by @matthewdouglas in #2935
• delete CCL env var setting by @Liangliang-Ma in #2927
• feat(ci): add pip caching in CI by @SauravMaheshkar in #2952

New Contributors

• @HarikrishnanBalagopal made their first contribution in #2921
• @fmo-mt made their first contribution in #2917
• @nullquant made their first contribution in #2932
• @cifkao made their first contribution in #2916
• @jkuntzer made their first contribution in #2875
• @matthewdouglas made their first contribution in #2936
• @Liangliang-Ma made their first contribution in #2927
• @SauravMaheshkar made their first contribution in #2952

Full Changelog: v0.32.1...v0.33.0

v0.32.0: Profilers, new hooks, speedups, and more!

Core

• Utilize shard saving from the huggingface_hub rather than our own implementation (#2795)
• Refactor logging to use logger in dispatch_model (#2855)
• The Accelerator.step number is now restored when using save_state and load_state (#2765)
• A new profiler has been added allowing users to collect performance metrics during model training and inference, including detailed analysis of execution time and memory consumption. These can then be generated in Chrome's tracing tool. Read more about it here (#2883)
• Reduced import times for doing import accelerate and any other major core import by 68%, now should be only slightly longer than doing import torch (#2845)
• Fixed a bug in get_backend and added a clear_device_cache utility (#2857)

Distributed Data Parallelism

• Introduce DDP communication hooks to have more flexibility in how gradients are communicated across workers, overriding the standard allreduce. (#2841)
• Make log_line_prefix_template optional the notebook_launcher (#2888)

FSDP

• If the output directory doesn't exist when using accelerate merge-weights, one will be automatically created (#2854)
• When merging weights, the default is now .safetensors (#2853)

XPU

• Migrate to pytorch's native XPU backend on torch>=2.4 (#2825)
• Add @require_triton test decorator and enable test_dynamo work on xpu (#2878)
• Fixed load_state_dict not working on xpu and refine xpu safetensors version check (#2879)

XLA

• Added support for XLA Dynamo backends for both training and inference (#2892)

Examples

• Added a new multi-cpu SLURM example using accelerate launch (#2902)

Full Changelog

• Use shard saving from huggingface_hub by @SunMarc in #2795
• doc: fix link by @imba-tjd in #2844
• Revert "Slight rename" by @SunMarc in #2850
• remove warning hook addede during dispatch_model by @SunMarc in #2843
• Remove underlines between badges by @novialriptide in #2851
• Auto create dir when merging FSDP weights by @helloworld1 in #2854
• Add DDP Communication Hooks by @yhna940 in #2841
• Refactor logging to use logger in dispatch_model by @panjd123 in #2855
• xpu: support xpu backend from stock pytorch (>=2.4) by @dvrogozh in #2825
• Drop torch re-imports in npu and mlu paths by @dvrogozh in #2856
• Default FSDP weights merge to safetensors by @helloworld1 in #2853
• [tests] fix bug in test_tracking.ClearMLTest by @faaany in #2863
• [tests] use torch_device instead of 0 for device check by @faaany in #2861
• [tests] skip bnb-related tests instead of failing on xpu by @faaany in #2860
• Potentially fix tests by @muellerzr in #2862
• [tests] enable XPU backend for test_zero3_integration by @faaany in #2864
• Support saving and loading of step while saving and loading state by @bipinKrishnan in #2765
• Add Profiler Support for Performance Analysis by @yhna940 in #2883
• Speed up imports and add a CI by @muellerzr in #2845
• Make log_line_prefix_template Optional in Elastic Launcher for Backward Compatibility by @yhna940 in #2888
• Add XLA Dynamo backends for training and inference by @johnsutor in #2892
• Added a MultiCPU SLURM example using Accelerate Launch and MPIRun by @okhleif-IL in #2902
• make more cuda-only tests device-agnostic by @faaany in #2876
• fix mlu device longTensor bugs by @huismiling in #2887
• add require_triton and enable test_dynamo work on xpu by @faaany in #2878
• fix load_state_dict for xpu and refine xpu safetensor version check by @faaany in #2879
• Fix get_backend bug and add clear_device_cache function by @NurmaU in #2857

New Contributors

• @McPatate made their first contribution in #2836
• @imba-tjd made their first contribution in #2844
• @novialriptide made their first contribution in #2851
• @panjd123 made their first contribution in #2855
• @dvrogozh made their first contribution in #2825
• @johnsutor made their first contribution in #2892
• @okhleif-IL made their first contribution in #2902
• @NurmaU made their first contribution in #2857

Full Changelog: v0.31.0...v0.32.0

v0.31.0: Better support for sharded state dict with FSDP and Bugfixes

Core

• Set timeout default to PyTorch defaults based on backend by @muellerzr in #2758
• fix duplicate elements in split_between_processes by @hkunzhe in #2781
• Add Elastic Launch Support to notebook_launcher by @yhna940 in #2788
• Fix Wrong use of sync_gradients used to implement sync_each_batch by @fabianlim in #2790

FSDP

• Introduce shard-merging util for FSDP by @muellerzr in #2772
• Enable sharded state dict + offload to cpu resume by @muellerzr in #2762
• Enable config for fsdp activation checkpointing by @helloworld1 in #2779

Megatron

• Upgrade huggingface's megatron to nvidia's megatron when use MegatronLMPlugin by @zhangsheng377 in #2501

What's Changed

• Add feature to allow redirecting std streams into log files when using torchrun as the launcher. by @lyuwen in #2740
• Update modeling.py by adding try-catch section to skip the unavailable devices by @MeVeryHandsome in #2681
• Fixed the problem of incorrect conditional judgment statement when configuring enable_cpu_affinity by @statelesshz in #2748
• Fix stacklevel in logging to log the actual user call site (instead of the call site inside the logger wrapper) of log functions by @luowyang in #2730
• LOMO / FIX: Support multiple optimizers by @younesbelkada in #2745
• Fix max_memory assignment by @SunMarc in #2751
• Fix duplicate environment variable check in multi-cpu condition by @yhna940 in #2752
• Simplify CLI args validation and ensure CLI args take precedence over config file. by @Iain-S in #2757
• Fix sagemaker config by @muellerzr in #2753
• fix cpu omp num threads set by @jiqing-feng in #2755
• Revert "Simplify CLI args validation and ensure CLI args take precedence over config file." by @muellerzr in #2763
• Enable sharded cpu resume by @muellerzr in #2762
• Sets default to PyTorch defaults based on backend by @muellerzr in #2758
• optimize get_module_leaves speed by @BBuf in #2756
• fix minor typo by @TemryL in #2767
• Fix small edge case in get_module_leaves by @SunMarc in #2774
• Skip deepspeed test by @SunMarc in #2776
• Enable config for fsdp activation checkpointing by @helloworld1 in #2779
• Add arg from CLI to fix failing test by @muellerzr in #2783
• Skip tied weights disk offload test by @SunMarc in #2782
• Introduce shard-merging util for FSDP by @muellerzr in #2772
• FIX / FSDP : Guard fsdp utils for earlier PyTorch versions by @younesbelkada in #2794
• Upgrade huggingface's megatron to nvidia's megatron when use MegatronLMPlugin by @zhangsheng377 in #2501
• Fixup CLI test by @muellerzr in #2796
• fix duplicate elements in split_between_processes by @hkunzhe in #2781
• Add Elastic Launch Support to notebook_launcher by @yhna940 in #2788
• Fix Wrong use of sync_gradients used to implement sync_each_batch by @fabianlim in #2790
• Fix type in accelerator.py by @qgallouedec in #2800
• fix comet ml test by @SunMarc in #2804
• New template by @muellerzr in #2808
• Fix access error for torch.mps when using torch==1.13.1 on macOS by @SunMarc in #2806
• 4-bit quantization meta device bias loading bug by @SunMarc in #2805
• State dictionary retrieval from offloaded modules by @blbadger in #2619
• add cuda dep for a test by @SunMarc in #2820
• Remove out-dated xpu device check code in get_balanced_memory by @faaany in #2826
• Fix DeepSpeed config validation error by changing stage3_prefetch_bucket_size value to an integer by @adk9 in #2814
• Improve test speeds by up to 30% in multi-gpu settings by @muellerzr in #2830
• monitor-interval, take 2 by @muellerzr in #2833
• Optimize the megatron plugin by @zhangsheng377 in #2822
• fix fstr format by @Jintao-Huang in #2810

New Contributors

• @lyuwen made their first contribution in #2740
• @MeVeryHandsome made their first contribution in #2681
• @luowyang made their first contribution in #2730
• @Iain-S made their first contribution in #2757
• @BBuf made their first contribution in #2756
• @TemryL made their first contribution in #2767
• @helloworld1 made their first contribution in #2779
• @hkunzhe made their first contribution in #2781
• @adk9 made their first contribution in #2814
• @Jintao-Huang made their first contribution in #2810

Full Changelog: v0.30.1...v0.31.0

v0.30.1: Bugfixes

Patchfix

• Fix duplicate environment variable check in multi-cpu condition thanks to @yhna940 in #2752
• Fix issue with missing values in the SageMaker config leading to not being able to launch in #2753
• Fix CPU OMP num threads setting thanks to @jiqing-feng in #2755
• Fix FSDP checkpoint unable to resume when using offloading and sharded weights due to CUDA OOM when loading the optimizer and model #2762
• Fixed the problem of incorrect conditional judgment statement when configuring enable_cpu_affinity thanks to @statelesshz in #2748
• Fix stacklevel in logging to log the actual user call site (instead of the call site inside the logger wrapper) of log functions thanks to @luowyang in #2730
• Fix support for multiple optimizers when using LOMO thanks to @younesbelkada in #2745

Full Changelog: v0.30.0...v0.30.1

v0.30.0: Advanced optimizer support, MoE DeepSpeed support, add upcasting for FSDP, and more

Core

• We've simplified the tqdm wrapper to make it fully passthrough, no need to have tqdm(main_process_only, *args), it is now just tqdm(*args) and you can pass in is_main_process as a kwarg.
• We've added support for advanced optimizer usage:
  • Schedule free optimizer introduced by Meta by @muellerzr in #2631
  • LOMO optimizer introduced by OpenLMLab by @younesbelkada in #2695
• Enable BF16 autocast to everything during FP8 and enable FSDP by @muellerzr in #2655
• Support dataloader send_to_device calls to use non-blocking by @drhead in #2685
• allow gather_for_metrics to be more flexible by @SunMarc in #2710
• Add cann version info to command accelerate env for NPU by @statelesshz in #2689
• Add MLU rng state setter by @ArthurinRUC in #2664
• device agnostic testing for hooks&utils&big_modeling by @statelesshz in #2602

Documentation

• Through collaboration between @fabianlim (lead contribuitor), @stas00, @pacman100, and @muellerzr we have a new concept guide out for FSDP and DeepSpeed explicitly detailing how each interop and explaining fully and clearly how each of those work. This was a momumental effort by @fabianlim to ensure that everything can be as accurate as possible to users. I highly recommend visiting this new documentation, available here
• New distributed inference examples have been added thanks to @SunMarc in #2672
• Fixed some docs for using internal trackers by @brentyi in #2650

DeepSpeed

• Accelerate can now handle MoE models when using deepspeed, thanks to @pacman100 in #2662
• Allow "auto" for gradient clipping in YAML by @regisss in #2649
• Introduce a deepspeed-specific Docker image by @muellerzr in #2707. To use, pull the gpu-deepspeed tag docker pull huggingface/accelerate:cuda-deepspeed-nightly

Megatron

• Megatron plugin can support NPU by @zhangsheng377 in #2667

Big Modeling

• Add strict arg to load_checkpoint_and_dispatch by @SunMarc in #2641

Bug Fixes

• Fix up state with xla + performance regression by @muellerzr in #2634
• Parenthesis on xpu_available by @muellerzr in #2639
• Fix is_train_batch_min type in DeepSpeedPlugin by @yhna940 in #2646
• Fix backend check by @jiqing-feng in #2652
• Fix the rng states of sampler's generator to be synchronized for correct sharding of dataset across GPUs by @pacman100 in #2694
• Block AMP for MPS device by @SunMarc in #2699
• Fixed issue when doing multi-gpu training with bnb when the first gpu is not used by @SunMarc in #2714
• Fixup free_memory to deal with garbage collection by @muellerzr in #2716
• Fix sampler serialization failing by @SunMarc in #2723
• Fix deepspeed offload device type in the arguments to be more accurate by @yhna940 in #2717

Full Changelog

• Schedule free optimizer support by @muellerzr in #2631
• Fix up state with xla + performance regression by @muellerzr in #2634
• Parenthesis on xpu_available by @muellerzr in #2639
• add third-party device prefix to execution_device by @faaany in #2612
• add strict arg to load_checkpoint_and_dispatch by @SunMarc in #2641
• device agnostic testing for hooks&utils&big_modeling by @statelesshz in #2602
• Docs fix for using internal trackers by @brentyi in #2650
• Allow "auto" for gradient clipping in YAML by @regisss in #2649
• Fix is_train_batch_min type in DeepSpeedPlugin by @yhna940 in #2646
• Don't use deprecated Repository anymore by @Wauplin in #2658
• Fix test_from_pretrained_low_cpu_mem_usage_measured failure by @yuanwu2017 in #2644
• Add MLU rng state setter by @ArthurinRUC in #2664
• fix backend check by @jiqing-feng in #2652
• Megatron plugin can support NPU by @zhangsheng377 in #2667
• Revert "fix backend check" by @muellerzr in #2669
• tqdm: *args should come ahead of main_process_only by @rb-synth in #2654
• Handle MoE models with DeepSpeed by @pacman100 in #2662
• Fix deepspeed moe test with version check by @pacman100 in #2677
• Pin DS...again.. by @muellerzr in #2679
• fix backend check by @jiqing-feng in #2670
• Deprecate tqdm args + slight logic tweaks by @muellerzr in #2673
• Enable BF16 autocast to everything during FP8 + some tweaks to enable FSDP by @muellerzr in #2655
• Fix the rng states of sampler's generator to be synchronized for correct sharding of dataset across GPUs by @pacman100 in #2694
• Simplify test logic by @pacman100 in #2697
• Add source code for DataLoader Animation by @muellerzr in #2696
• Block AMP for MPS device by @SunMarc in #2699
• Do a pip freeze during workflows by @muellerzr in #2704
• add cann version info to command accelerate env by @statelesshz in #2689
• Add version checks for the import of DeepSpeed moe utils by @pacman100 in #2705
• Change dataloader send_to_device calls to non-blocking by @drhead in #2685
• add distributed examples by @SunMarc in #2672
• Add diffusers to req by @muellerzr in #2711
• fix bnb multi gpu training by @SunMarc in #2714
• allow gather_for_metrics to be more flexible by @SunMarc in #2710
• Add Upcasting for FSDP in Mixed Precision. Add Concept Guide for FSPD and DeepSpeed. by @fabianlim in #2674
• Segment out a deepspeed docker image by @muellerzr in #2707
• Fixup free_memory to deal with garbage collection by @muellerzr in #2716
• fix sampler serialization by @SunMarc in #2723
• Fix sampler failing test by @SunMarc in #2728
• Docs: Fix build main documentation by @SunMarc in #2729
• Fix Documentation in FSDP and DeepSpeed Concept Guide by @fabianlim in #2725
• Fix deepspeed offload device type by @yhna940 in #2717
• FEAT: Add LOMO optimizer by @younesbelkada in #2695
• Fix tests on main by @muellerzr in #2739

New Contributors

• @brentyi made their first contribution in #2650
• @regisss made their first contribution in #2649
• @yhna940 made their first contribution in #2646
• @Wauplin made their first contribution in #2658
• @ArthurinRUC made their first contribution in #2664
• @jiqing-feng made their first contribution in #2652
• @zhangsheng377 made their first contribution in #2667
• @rb-synth made their first contribution in #2654
• @drhead made their first contribution in #2685

Full Changelog: https://github.com/huggingface/acce...

v0.29.3: Patchfix

• Fixes issue with backend refactor not working on CPU-based distributed environments by @jiqing-feng: #2670
• Fixes issue where load_checkpoint_and_dispatch needs a strict argument
• by @SunMarc: #2641

Full Changelog: v0.29.2...v0.29.3

v0.29.2: Patchfix

• Fixes xpu missing parenthesis #2639
• Fixes XLA and performance degradation on init with the state #2634

v0.29.1: Patchfix

Fixed an import which would cause running accelerate CLI to fail if pytest wasn't installed