Skip to content

Releases: mosaicml/composer

v0.25.0

24 Sep 20:56
@dakinggg dakinggg
v0.25.0
0c4e110
Compare
Choose a tag to compare
v0.25.0 Latest
Latest

What's New

1. Torch 2.4.1 Compatibility (#3609)

We've added support for torch 2.4.1, including necessary patches to Torch.

Deprecations and breaking changes

1. Microbatch device movement (#3567)

Instead of moving the entire batch to device at once, we now move each microbatch to device. This saves memory for large inputs, e.g. multimodal data, when training with many microbatches.

This change may affect certain callbacks which run operations on the batch which require it to be moved to an accelerator ahead of time, such as the two changed in this PR. There shouldn't be too many of these callbacks, so we anticipate this change will be relatively safe.

2. DeepSpeed deprecation version (#3634)

We have update the Composer version that we will remove support for DeepSpeed to 0.27.0. Please reach out on GitHub if you have any concerns about this.

3. PyTorch legacy sharded checkpoint format

PyTorch briefly used a different sharded checkpoint format than the current one, which was quickly deprecated by PyTorch. We have continued to support loading legacy format checkpoints for a while, but we will likely be removing support for this format entirely in an upcoming release. We initially removed support for saving in this format in #2262, and the original feature was added in #1902. Please reach out if you have concerns or need help converting your checkpoints to the new format.

What's Changed

Full Changelog: v0.24.1...v0.25.0

Contributors

bigning, irenedea, and 5 other contributors
Assets 2
Loading

v0.24.1

27 Aug 22:37
@snarayan21 snarayan21
v0.24.1
3c7fefb
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.
Compare
Choose a tag to compare
v0.24.1

Bug Fixes

1. Disallow passing device_mesh to FSDPConfig (#3580)

Explicitly errors if device_mesh is passed to FSDPConfig. This completes the deprecation from v0.24.0 and also addresses cases where a user specified a device mesh but it was ignored, leading to training with the incorrect parallelism style (e.g., using FSDP instead of HSDP).

What's Changed

Full Changelog: v0.24.0...v0.24.1

Contributors

dependabot, snarayan21, and KevDevSha
Assets 2
Loading

v0.24.0

26 Aug 14:48
@snarayan21 snarayan21
v0.24.0
020b0ef
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.
Compare
Choose a tag to compare
v0.24.0

What's New

1. Torch 2.4 Compatibility (#3542, #3549, #3553, #3552, #3565)

Composer now supports Torch 2.4! We are tracking a few issues with the latest PyTorch we have raised with the PyTorch team related to checkpointing:

  • [PyTorch Issue] Distributed checkpointing using PyTorch DCP has issues with stateless optimizers, e.g. SGD. We recommend using composer.optim.DecoupledSGDW as a workaround.
  • [PyTorch Issue] Distributed checkpointing using PyTorch DCP broke backwards compatibility. We have patched this using the following planner, but this may break custom planner loading.

2. New checkpointing APIs (#3447, #3474, #3488, #3452)

We've added new checkpointing APIs to download, upload, and load / save, so that checkpointing is usable outside of a Trainer object. We will be fully migrating to these new APIs in the next minor release.

3: Improved Auto-microbatching (#3510, #3522)

We've fixed deadlocks with auto-microbatching with FSDP, bringing throughput in line with manually setting the microbatch size. This is achieved through enabling sync hooks wherever a training run might OOM to find the correct microbatch size, and disabling these hooks for the rest of training.

Bug Fixes

1. Fix checkpoint symlink uploads (#3376)

Ensures that checkpoint files are uploaded before the symlink file, fixing errors with missing or incomplete checkpoints.

2. Optimizer tracks same parameters after FSDP wrapping (#3502)

When only a subset of parameters should be tracked by the optimizer, FSDP wrapping will now not interfere.

What's Changed

Read more

Contributors

bigning, sashaDoubov, and 17 other contributors
Assets 2
Loading

v0.23.5

03 Jul 02:08
@XiaohanZhangCMU XiaohanZhangCMU
v0.23.5
56ccc2e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.
Compare
Choose a tag to compare
v0.23.5

What's New

1. Variable length dataloaders (#3416)

Adds support for dataloaders with rank-dependent lengths. The solution terminates iteration for dataloaders on all ranks when the first dataloader finishes.

Bug Fixed

1. Remove close flush for mosaicml logger (#3446)

Previously, the MosaicML Logger sporadically raised an error when the python interpreter was shutting down as it attempted to flush data on Event.CLOSE using futures, which cannot be scheduled at that time. Instead, we now only block on finishing existing data upload on Event.CLOSE, avoiding scheduling new futures.

What's Changed

Full Changelog: v0.23.4...v0.23.5

Contributors

eracah, JAEarly, and 11 other contributors
Assets 2
Loading

v0.23.4

21 Jun 15:09
@mvpatel2000 mvpatel2000
v0.23.4
ec8799d
Compare
Choose a tag to compare
v0.23.4

Bug Fixes

1. Patch PyTorch 2.3.1 (#3419)

Fixes missing import when monkeypatching device mesh functions in PyTorch 2.3.1. This is necessary for MoE training.

Full Changelog: v0.23.3...v0.23.4

Assets 2
Loading

v0.23.3

21 Jun 00:18
@karan6181 karan6181
v0.23.3
7c7f6de
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.
Compare
Choose a tag to compare
v0.23.3

New Features

1. Update mlflow logger to use the new API with time-dimension to view images in MLFlow (#3286)

We've enhanced the MLflow logger's log_image function to use the new API with time-dimension support, enabling images to be viewed in MLflow.

2. Add logging buffer time to MLFLow logger (#3401)

We've added the logging_buffer_seconds argument to the MLflow logger, which specifies how many seconds to buffer before sending logs to the MLflow tracking server.

Bug Fixes

1. Only require databricks-sdk when on Databricks platform (#3389)

Previously, MLFlow always imported the databricks-sdk. Now, we only require the sdk if on the databricks platform and using databricks secrets to access managed MLFlow.

2. Skip extra dataset state load during job resumption (#3393)

Previously, when loading a checkpoint with train_dataloader, the dataset_state would load first, and if train_dataloader was set again afterward, load_state_dict would be called with a None value. Now, we've added a check in the train_dataloader setter to skip this redundant load.

3. Fix auto-microbatching on CUDA 12.4 (#3400)

In CUDA 12.4, the out-of-memory error message has changed to CUDA error: out of memory. Previously, our logic hardcoded checks for CUDA out of memory when using device_train_microbatch_size="auto". Now, we check for both CUDA out of memory and CUDA error: out of memory.

4. Fix mlflow logging to Databricks workspace file paths which startswith /Shared/ prefix (#3410)

Previously, for MLflow logging, we prepended the path /Users/ to all user-provided logging paths on the Databricks platform, if not specified, including paths starting with /Shared/, which was incorrect since /Shared/ indicates a shared workspace. Now, the /Users/ prepend is skipped for paths starting with /Shared/.

What's Changed

New Contributors

Full Changelog: v0.23.2...v0.23.3

Contributors

bigning, eracah, and 12 other contributors
Assets 2
Loading

v0.23.2

08 Jun 03:11
@bigning bigning
v0.23.2
c7020f8
Compare
Choose a tag to compare
v0.23.2

Bug Fixes

  • Fix backward compatibility issue caused by missing eval metrics class

What's Changed:

  • Fix backward compatibility issue caused by missing eval metrics class by @bigning in #3385

Full Changelog: v0.23.1...release/v0.23.2

Contributors

bigning
Assets 2
Loading

v0.23.1

07 Jun 15:03
@mvpatel2000 mvpatel2000
v0.23.1
7a533cb
Compare
Choose a tag to compare
v0.23.1

What's New

1. PyTorch 2.3.1 Upgrade

Composer now supports PyTorch 2.3.1.

What's Changed

Full Changelog: v0.23.0...v0.23.1

Contributors

bigning, eracah, and 3 other contributors
Assets 2
Loading

v0.23.0

05 Jun 20:34
@bigning bigning
v0.23.0
bf2cb35
Compare
Choose a tag to compare
v0.23.0

What's New

1. Parallelism V2 + Tensor Parallel (#3335)

Composer now supports PyTorch's implementation of tensor parallelism. As part of this, we've revamped and simplified how Composer does distributed training. Previously, Composer accepted a fsdp_config attribute in the Trainer:

trainer = Trainer(model, fsdp_config = {'sharding_strategy': 'FULL_SHARD'})

As we generalize to more forms of parallelism, we've deprecated fsdp_config in favor of parallelism_config:

trainer = Trainer(
    model = model,
    ...
    parallelism_config = {
        'fsdp': {
            'sharding_strategy': 'FULL_SHARD',
            'data_parallel_shard_degree': 2,      # Size of shard dimension
            'data_parallel_replicate_degree': 2,  # Size of replicate dimension
        },
        'tp_config': {
            'tensor_parallel_degree': 2,          # Size of TP dimension
            'layer_plan': ...  # describes how to TP layers
        }
    }
)

As part of this change, we now default to using DTensor for parallelism with PyTorch FSDP. PyTorch has deprecated ShardedTensor, so this migrates to the new backend which avoids various checkpointing bugs.

See the docs for tensor parallel for more information. Note that tensor parallel is still experimental and may be subject to API breaking changes. All checkpointing features may also not work with this parallelism.

2. MLFLow API Simplification

Previously, MLFlow logger required a tracking URI and an absolute user path when using MLFlow with Databricks:

mlflow_logger = MLFlowLogger(
    tracking_uri = 'databricks',
    experiment_name = '/Users/xxx.yyy@zzz.com/my-first-project/'
)

trainer = Trainer(
    model = model,
    ...
    loggers = mlflow_logger,
)

Now, if you are using Databricks secrets as an environment variable, Composer will autopopulate tracking_uri and the experiment_name prefix:

trainer = Trainer(
    model = model,
    ...
    loggers = MLFlowLogger(experiment_name='my-first-project'),
)

3. Wallclock Save Interval

Composer now supports setting a save interval in wallclock time:

trainer = Trainer(
    model = model,
    ...
    save_interval='30m',
)

Note that most durations, such as max_duration, do not accept wallclock time, and the initial version of this feature is only limited to a subset of time features like save_interval.

Bug Fixes

  • Don't close the engine if it's already closed in #3143
  • Fix HF tests with Pin in #3248
  • Fix backwards compatibility tests in #3252
  • Fix unexpected remote checkpointing downloading in #3271
  • Fix HSDP with ShardDegree < 8 in #3313

What's Changed

Read more

Contributors

bigning, eracah, and 17 other contributors
Assets 2
Loading

v0.22.0

01 May 16:59
@snarayan21 snarayan21
v0.22.0
fe7964f
Compare
Choose a tag to compare
v0.22.0

What's New

🔥 Support for PyTorch v2.3.0

Composer now supports the recently-released PyTorch version 2.3.0! Please raise any issues with us so we can address them.

Bug Fixes

  • Fixing checks for device microbatch size for sequence parallelism in #3200
  • Fixing token logging in #3206
  • Search for run name in MLFlowLogger in #3215
  • Fix FQN names with activation checkpointing in #3210
  • Strict weight matching for checkpoint loading in #3219

What's Changed

Full Changelog: v0.21.3...v0.22.0

Contributors

bfontain, hanlint, and 6 other contributors
Assets 2
Loading