CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.4.11] - TBD

• cleaned up some old issues and fixed a few bug in FSDP
• removing SSD offload to simplify the FSDP code

[0.4.8]/[0.4.9]/[0.4.10]

Added

• FSDP: Add pickle/unpickle support for SsdTensorHandle (and derived classes), verified that FSDP models w/ ssd_offload enabled can correctly call model.state_dict() and model.load_state_dict(...) and thus successfully checkpoint and recover parameters stored as SsdFlatParameters. [#964]
• FSDP Ssd Offload: [#974]
  • add setattr function for SsdTensorHandle/SsdFlatParameter to capture when .data is overridden and perform necessary checks/updates before letting the tensor metadata be updated.
  • There was a bug in pytorch core that caused re-assigning TensorSubclass .data to disable the torch_dispatch mechanism and effectively revert back to a normal tensor. Since ssd_offload feature uses torch_dispatch extensively and FSDP overrides .data, ssd_offload now can only be used if pytorch version is 1.12.0 or later (currently pytorch-nightly release). pytest tests are disabled and trying to import ssd_offload or use FSDP with ssd_offload enabled will raise ImportError Exceptions.
  • Enhance storage_state ON_CPU into ON_CPU_CLEAN and ON_CPU_DIRTY. ON_CPU_CLEAN indicates that the .tensor value and the value stored on disk are identical, and no writes are needed when calling .to_file(). ON_CPU_DIRTY indicates .tensor does not match value stored on disk, and a write is necessary. This is disabled in FSDP as it does not currently flush variables to disk as soon as optimizer modifies values.
  • Fix detection in SsdTensorHandle torch_dispatch when it is used in an in-place operator, or as an output, then calling mark_dirty().
  • Fix grad_fn on SsdFlatParameterViews so that both ssd_flat_param.view.grad_fn points to a split and view on ssd_flat_parameter so that when X.backwards() is called. ssd_flat_param.grad is properly updated in the backward pass.
  • Add unit tests to verify grad_fn and backward hooks are called appropriately on SsdFlatParameters/SsdFlatParameterViews
  • Implement SsdFlatParameter.from_tensor() direct_to_file option. This allows creating and SsdFlatParameter and directly writing it to disk, rather than creating a new tensor first. This prevents another copy of tensors to be instantiated and can save memory when creating very large SsdFlatParameters.
  • Add SsdFlatParameterViewProperty for overriding a parameter in an nn.Module similar to pytorch core's parameterization code path. This allows SsdFlatParameterViews to be updated as SsdFlatParameter.data is overridden by replacing Layer.weight with a property whose getter returns ssd_flat_param.views[view_id].
  • Added example SsdFlatParameterViewParameterization on how the existing parameterization method could be used instead of SsdFlatParameterViewProperty. But this method will result in memory inefficiencies due to parameterization always keeping a copy of the original parameter internally.

Fixed

• fixed some bugs in FSDP related to supporting data2vec EMA modules.

[0.4.6] - 2022-03-08

Added

• CosFace's LMCL is added to MEVO. This is a loss function that is suitable for large number of prediction target classes. It added normalization, class separation margins and feature vector scaling to the standard output projection + cross-entropy loss. MEVO supported this with its memory saving techniques so that peak GPU memory is much reduced. [#916]
• FSDP: Added skip_params_check_for_root flag to default_auto_wrap_policy which, if set, wraps the root module regardless of how many unwrapped params there were left after children were wrapped. [#930]
• FSDP: Add support for saving optimizer state when using expert replicas with FSDP.
• OSS: Add a new arg "forced_broadcast_object" to OSS init to apply "_broadcast_object" for rebuilding the sharded optimizer. [#937]
• FSDP: Add an arg disable_reshard_on_root for FSDP init [#878]

Fixed

• FSDP: fixed handling of internal states with state_dict and load_state_dict function so that they don't change lazy init state if training hasn't started. [#922]
• FSDP: added support of optimizer state handling when some of the parameters are not used. An example is that in a model with a EMA copy that doesn't get trained but still wants to be sharded. [#922]

[0.4.5] - 2022-01-14

Added

• Layer-wise Gradient Scaling [new feature][experimental] Layer-wise gradient scaling helps overcomes gradient overflow issues. When used in conjunction with mixed precision, it enables training larger models and makes the training process more stable, especially in deep networks [#879]
• FSDP: Added process_group_reduce_scatter parameter to allow users to pass in the process group that is used for reduce scatter operation. [#897]
• FSDP: Added state_dict_on_rank_0_only flag allow user choose to return full state dict on rank 0 and return empty dict non-rank 0 to prevent OOM [#844]

Changed

• FSDP: Enabled reduce_scatter operation to overlap with all_gather stream and computation stream in backward propagation. This change increases FSDP's throughput. To roll back this change, please pass ProcessGroupName.default to the process_group_reduce_scatter API parameter in FSDP [#897]

Fixed

• FSDP: Fixed the issue that the padding size of the input tensor of the reduce scatter is not equal to the reduce scatter process group size # [#907]

[0.4.4] - 2021-12-21

Fixed

• Inf/nan check on all gradient tensors in ShardedGradScaler [#890]
• Allow user to supress warning on trainable_parameter in sharded_ddp.py [#886]
• KeyError in certain FSDP wrapping scenarios [#881]

Added

• Support eval mode in MEVO [#884]
• A new benchmark for the MOE model [#866]

Changed

• Fixed a corner case of FSDP init order and losing one of the flags [#880]
• FSDP: Adding basic training support for SSD Offload, it now only supports flattened parameters. Renamed OffloadConfig.ssd_filepath_dir to more generic OffloadConfig.dir. SSD Offload remains an experimental feature. [#887]

[0.4.3] - 2021-11-18

Added

• Sharded Grad Scaler works with cpu offload in mixed and full precision. [#831]
• API for specifying SSD offload for params with FSDP. You can use a OffloadConfig to specify the type of offload and the file path for storing params on SSD. Note: This is an experimental feature. [#855]

Changed

• MEVO: fixed eval and checkpointing code paths [#851]
• Cleanup: Moving forward we would be testing all of our code with Python 3.9.7, CUDA 11.2 and the following three versions of PyTorch [#847]:
  • the most recent stable version
  • the most recent LTS version
  • a recent nightly build

[0.4.2] - 2021-11-08

Fixed

• FSDP: Fixed an pre-backward hook bug for certain type of models and FSDP config. [#833]

Added

• FSDP: Add support for SSD offload for eval workloads. This is a new experimental feature and should be used with caution.
• LayerwiseMemoryTracker[feature][experimental]: This is a new experimental tool to help track, visualize and suggest fix for memory issues occurring during the forward/backward pass of your models. [#808]
• FSDP: limited support of shared weights between FSDP wrappers. This allows large parameter and gradient memory to be sharded despite being needed from different layers due to weight sharing. [#836]
• OffloadModel: Fix node names to enable correct sharding in auto_shard.py [#830]
• OSS: Relaxed speed and memory constraints on OSS golden data due to regression when we bumped up the PyTorch version to 1.9. [#828] [#825]
• Chore: Update PyTorch version that we run benchmarks with. [#823]
• Chore: Update PyTorch version that we run test with. [#809]
• OffloadModel: Extend auto_shard.py to allow dealing with conditionals automatically when tracing with torch.fx. This will work for most cases except when the conditional is part of the root instance. [#817]
• [MEVO]: a custom layer to help big vocab trainings. Experimental. Docs is still TBD. [#840]
• SlowMoDistributedDataParallel[feature][experimental] - This is a distributed training wrapper which should be useful on clusters with slow network interconnects (eg Ethernet). This improves on performance as compared to Distributed Data Parallel in such clusters. [#378]

[0.4.1] - 2021-09-17

Fixed

• FSDP: We don't attach post backward hooks for params that don't require grad. However in the hook triggered after the post backward hook, we assert on the POST_BACKWARD state which can only be set in the post backward hook. Modified the assert to account for the fact that the root FSDP module can have child modules with params that require grad and it can contain params that don't require grad and hence can fail the previous assert. [#761]

• FSDP: Fixed a bug when multiple backward pass is called within an iteration, parameters' sharding state might be incorrect. [#775]

• activation checkpoint: Ensure outputs of checkpointed modules only require grad if either the input requires grad or if the parameters require grad. [#787]

• OSS: fix the broadcast_fp16 option, broken after a refactor, this flag was doing nothing (bugfix).[#795]

• OSS: update default device when refreshing the params, meaning that moving the model to GPU after the OSS wrap will not trigger warnings and slow the jobs (ease of use). [#786]

Added

• FSDP: Added support for returning the original names of parameters when named_parameters is called on the module. To retrieve the orginal names of the parameters along with the params, you need to call named_parameters under the summon_full_params context when using flattened params or original params. If you are using original params (i.e flatten_params=False), calling named_parameters outside of the summon_full_params context will still return the original param names along with the local shards. [#755]
• FSDP: Ensure gradient reduction accumulates into the unsharded gradient tensor within a backwards pass. This matters when an FSDP module is called multiple times within a forward pass, and reduction is not deferred using activation checkpoint forward counters, bucketing or some other mechanism. [#784]
• activation checkpoint: Added a context manager to disable checkpoint in case the same wrapped module needs to be checkpointed and not checkpointed in different parts of the module forward pass. [#772]
• FSDP: Added a toggle with an environment variable ENABLE_NCCL_BASE_COLLECTIVES=[0,1] to allow users enable/disable using new nccl base collecectives. By default, using new nccl base collectives is enabled. [#801]

[0.4.0] - 2021-07-31

Fixed

• FSDP: fixed final backward callback in certain activation checkpointed cases. Before this fix, if a model is activation checkpointed in a certain way, the final backward callback can fire incorrectly. That's due to autograd and reentrant backward graphs. With this fix, the final callback is always registered on the outer most root FSDP instance (i.e. the outer most backward graph), which result in reliably firing. This makes FSDP much more robust with respect to different models and activation checkpoints. [#753]

Added

• FSDP: support gradient accumulation without the no_sync context. This is useful in training with smaller number of GPU with same overall batch size as large number of GPUs. Compared with the no_sync context, this mode consumes less GPU memory but uses more networking bandwidth. [#752]

[0.3.9] - 2021-07-26

Fixed

• FSDP: fixed metadata saving and shard consolidation for MoE cases. When a model has shared parameters or mixture of expert layers, the handling of state dict metadata was broken. This release fixes that. [#746]
• OSS: fixed the buckets which would stay in fp16 if broadcast fp16 was required [#751]

Added

• FSDP: better performance; use _allgather_base and _reduce_scatter_base when they are available from pytorch nightly version (will be in 1.10 releases) [#729]
• FSDP: prepared FSDP internals for supporting multiple groups of flatten parameters (to support more general optimization) [#746]

[0.3.8] - 2021-07-12

Fixed

• checkpointing: Use dummy tensor to ensure backward pass is called. [#701]
• checkpointing: Ensure internal fwd counter is not incremented in eval mode. [#709]
• checkpointing: Use non-blocking CPU transfer to improve perf. [#719]
• FSDP: Fixed bug where buffers returned in state_dict() could still be half precision when mixed_precision is set to True. [#705]
• FSDP: Ensure requires_grad of FlatParameter is consistent with requires_grad of the original parameters. [#721]
• doc: Thoroughly improved the doc for FSDP. [#711]
• cleanup: Remove examples/ doc from the repo. [#712]
• cleanup: Future proof storage size test. [#735]
• cleanup: Migrate away from legacy torchtext iterators. [#713]
• chore: Updated torch 1.9 to release version. [#717]

Added

• FSDP: supporting multiple flatten parameter groups [#708] [#711]
• chore: Add the latest numpy version to requirements-test.txt to prevent mypy errors on certain PR commits [#732]

[0.3.7] - 2021-05-17

Fixed

• setup.py: hide CUDA extensions behind BUILD_CUDA_EXTENSIONS envvar [#634]
• checkpointing: rename and move the checkpoint_activations wrapper [#654]
• FSDP: fix local_state_dict potentially called child class's state_dict [#574]
• FSDP: fix extra process groups being created by default. Old behavior can cause excessive GPU memory usage [#678] [#681]
• FSDP: fix forward pass not overlapping compute and allgather [#671]
• FSDP: improved frozen weight support [#657]
• FSDP: workaround AMP autocast cache issue with clear_autocast_cache flag [#650]
• FSDP: Rename API arg cpu_offload to move_params_to_cpu to better reflect functionality. We will deprecate cpu_offload in an upcoming release [#676]
• MoE: several fixes [#666] [#667] [#668]
• SDP: re-expose the module property [#647]
• wrap: support wrapping based on wrapper_config [#685]

Added

• FSDP: added force_input_to_fp32 flag for SyncBatchNorm [#659]
• FSDP: better memory usage for reduce bucket [#633]
• FSDP: added local_metadata_dict to save sharding relating information [#683]
• FSDP: added consolidate_shard_weights to reconstruct the consolidated (non-sharded) model weights from saved sharded weights and metadata on the disk [#683]
• Experimental SyncBatchNorm [#662] [#680]

[0.3.6] - 2021-04-26

Added

• FSDP: Consolidate cpu_adam optimizer state dict (#607)

Fixed

• FSDP: handle model with multiple forward pass and checkpoint (#621)
• FSDP & SDP: check before calling _specify_ddp_gpu_num (#626)
• FSDP: relax checking root condition (#620)
• SDP: removing an assert which does not seem always accurate (#625)
• FSDP: changing FSDP init to by pass pg validation (#619)
• OSS: to 100% coverage (#618)

[0.3.5] - 2021-04-19

Added

• [offload] Add API, tutorial and smaller doc string changes. (#576)

Fixed

• FSDP: fixing training with freezing weights (#614)
• SDP: privatizing all the things (#611)
• FSDP: Make _get_default_cuda_device more robust to modules without params (#606)
• OffloadModel: Add prev codepath of using OffloadModel without activation checkpointing (#608)

[0.3.4] - 2021-04-13

Added

• FSDP: Add no broadcast optim state option (#560)

Fixed

• ShardedDDP: Properly handle .eval() mode (#587)
• ShardedDDP: Handle model being moved back to CPU prior to state consolidation (#573)
• FSDP: much faster state consolidation (#595)
• FSDP: Add gradient pre-dedivide to prevent overflow with large world sizes (#565)
• Offload: (experimental) Fix activation offloading to CPU (#588

[0.3.3] - 2021-04-1

Added

• FSDP: changed auto_wrap_bn utility function so that single FSDP group is optional (#556)
• FSDP: optimizer state load/save (#537)
• FSDP: fix weight init when using apply() (#543)
• Multiprocess Pipe: retired old implementation
• Experimental: xpipe

Fixed

• ShardedDDP deferred init (#558)

[0.3.2] - 2021-03-18

Added

• Experimental: Add spectrain support (#372)
• FSDP: enabled pytorch SyncBN (no asserting) (#527)
• FSDP: added auto_wrap_bn utility function (#531)

Fixed

• OSS: fix a compatibily problem with lightning wrt optimizer state dict (#510)
• FSDP: fixed a bug when part of autograd graph is traversed multiple times in mixed precision mode (#513)

[0.3.1] - 2021-03-09

Added

• FSDP docs (#455)
• enable_wrap and auto_wrap APIs (#446)
• Added experimental.nn.OffloadModel API for training large models on a single GPU.(#432)

Fixed

• OSS: fix a broken state dict when using non contiguous param groups
• Several SDP fixes around performance and corner cases
• Many FSDP fixes
• AdaScale & SDP/FSDP test added but not officially supported

[0.3.0] - 2021-02-22

Added

• FullyShardedDataParallel (FSDP) (#413)
• ShardedDDP fp16 grad reduction option (#402)
• Expose experimental algorithms within the pip package (#410)

Fixed

• Catch corner case when the model is too small with respect to the world size, and shards are empty (#406)
• Memory leak in checkpoint_wrapper (#412)

[0.1.7] - 2021-02-19

Fixed

• ShardedDDP and OSS handle model trainability changes during training (#369)
• ShardedDDP state dict load/save bug (#386)
• ShardedDDP handle train/eval modes (#393)
• AdaScale handling custom scaling factors (#401)

Added

• ShardedDDP manual reduce option for checkpointing (#389)

[0.1.6] - 2021-02-10

Added

• Checkpointing model wrapper (#376)
• Faster OSS, flatbuffers (#371)
• Small speedup in OSS clipgradnorm (#363)

Fixed

• Bug in ShardedDDP with 0.1.5 depending the init (KeyError / OSS)
• Much refactoring in Pipe (#357, #358, #360, #362, #370, #373)
• Better pip integration / resident pytorch (#375)

[0.1.5] - 2021-02-03

Added

• Pytorch compatibility for OSS checkpoints (#310)
• Elastic checkpoints for OSS, world size can vary in between save and loads (#310)
• Tensor views for OSS bucketing, reduced CPU use (#300)
• Bucket calls in ShardedDDP, for faster inter node communications (#327)
• FlattenParamWrapper, which flattens module parameters into a single tensor seamlessly (#317)
• AMPnet experimental support (#304)

Fixed

• ShardedDDP properly handles device changes via .to() (#353)
• Add a new interface for AdaScale, AdaScaleWrapper, which makes it compatible with OSS (#347)

[0.1.4] - 2021-01-07

Fixed

• Missing cu files in the pip package

[0.1.3] - 2021-01-04

Fixed

• Release numbering within python and from pypi

[0.1.2] - 2021-01-04

Added

• AdaScale: . Added gradient accumulation feature (#202) . Added support of torch.lr_scheduler (#229) . Added support for add_param_groups (#266) . Added support for scale != world_size (#266)

Fixed

• AdaScale: smoothing factor value fixed when using gradient accumulation (#235)
• Pipe: documentation on balancing functions (#243)
• ShardedDDP: handle typical NLP models
• ShardedDDP: better partitioning when finetuning

[0.1.1] - 2020-12-01

Fixed

• make sure pip package includes header files (#221)

[0.1.0] - 2020-12-01

Added

• ShardedDataParallel with autoreduce (#157)
• cpu support for Pipe (#188)
• ShardedOptim: Distributed Grad Scaler (for torch AMP) (#182)
• OSS-aware clip grads, bridge sharded states (#167)
• oss: add rank_local_state_dict staticmethod (#174)
• support for PyTorch 1.7.0 (#171)
• Add implementation of AdaScale (#139)

Fixed

• pip package install (#196, #200)

[0.0.3] - 2020-10-14

Added

• multi-process pipe

Fixed

• multiple OSS fixes
• MegaTron+OSS DDP fix

[0.0.2] - 2020-08-28

Added

• add ddp that works with oss with reduce() not all_reduce() (#19)
• support for PyTorch v1.6
• add mixed precision Adam (#40)
• Adam optimizer state scaling (#44)

Fixed

• properly restore a sharded optim state (#39)
• OSS restore state to proper device (#46)
• optim/oss: support optimizers with additional step kwargs (#53)
• optim/oss: fix state cast (#56)
• fix eval for oss_ddp (#55)
• optim/oss: work correctly with LRScheduler (#58)

[0.0.1] - 2020-07-31

• Initial release.