Bump version to v0.24.0 + deprecations

composer/_version.py

@@ -3,4 +3,4 @@
 
 """The Composer Version."""
 
-__version__ = '0.24.0.dev0'
+__version__ = '0.24.0'

composer/callbacks/__init__.py

@@ -9,7 +9,6 @@
 from composer.callbacks.activation_monitor import ActivationMonitor
 from composer.callbacks.checkpoint_saver import CheckpointSaver
 from composer.callbacks.early_stopper import EarlyStopper
-from composer.callbacks.eval_output_logging_callback import EvalOutputLogging
 from composer.callbacks.export_for_inference import ExportForInferenceCallback
 from composer.callbacks.free_outputs import FreeOutputs
 from composer.callbacks.generate import Generate
@@ -36,7 +35,6 @@
     'CheckpointSaver',
     'MLPerfCallback',
     'EarlyStopper',
-    'EvalOutputLogging',
     'ExportForInferenceCallback',
     'ThresholdStopper',
     'ImageVisualizer',

composer/core/state.py

@@ -47,7 +47,6 @@
     ParallelismConfig,
     ParallelismType,
     TPConfig,
-    VersionedDeprecationWarning,
     batch_get,
     batch_set,
     dist,
@@ -617,7 +616,7 @@ def _validate_parallelism_configs(self):
         # Load monolith rank0 only
         if self.load_monolith_rank0_only:
             if self.tp_config is not None:
-                raise ValueError('load_fsdp_monolith_rank0_only is not compatible with tensor parallelism (TP).')
+                raise ValueError('load_monolith_rank0_only is not compatible with tensor parallelism (TP).')
             assert self.fsdp_config is not None
             error_message = ''
             if self.fsdp_config.sync_module_states == False:
@@ -900,21 +899,6 @@ def fsdp_state_dict_type(self):
     def fsdp_sharded_state_dict_enabled(self):
         return self.fsdp_config is not None and self.fsdp_enabled and self.fsdp_state_dict_type == 'sharded'
 
-    @property
-    def fsdp_device_mesh(self):
-        warnings.warn(VersionedDeprecationWarning('fsdp_device_mesh is deprecated. Use device_mesh instead.', '0.24'))
-        return self.device_mesh
-
-    @property
-    def load_fsdp_monolith_rank0_only(self):
-        warnings.warn(
-            VersionedDeprecationWarning(
-                'load_fsdp_monolith_rank0_only is deprecated. Use load_monolith_rank0_only instead.',
-                '0.24',
-            ),
-        )
-        return self.load_monolith_rank0_only
-
     @property
     def load_monolith_rank0_only(self):
         return (

composer/trainer/trainer.py

@@ -114,7 +114,6 @@
     Transform,
     VersionedDeprecationWarning,
     checkpoint,
-    create_fsdp_config,
     dist,
     ensure_tuple,
     export_with_logger,
@@ -1323,7 +1322,7 @@ def __init__(
                 if isinstance(parallelism_config['fsdp'], FSDPConfig):
                     parallelism_config_args['fsdp'] = parallelism_config['fsdp']
                 else:
-                    parallelism_config_args['fsdp'] = create_fsdp_config(parallelism_config['fsdp'])
+                    parallelism_config_args['fsdp'] = FSDPConfig(**parallelism_config['fsdp'])
             if 'tp' in parallelism_config and parallelism_config['tp'] is not None:
                 if isinstance(parallelism_config['tp'], TPConfig):
                     parallelism_config_args['tp'] = parallelism_config['tp']

composer/utils/__init__.py

@@ -76,7 +76,7 @@
     UCObjectStore,
     build_remote_backend,
 )
-from composer.utils.parallelism import FSDPConfig, ParallelismConfig, TPConfig, create_fsdp_config
+from composer.utils.parallelism import FSDPConfig, ParallelismConfig, TPConfig
 from composer.utils.remote_uploader import RemoteFilesExistingCheckStatus, RemoteUploader
 from composer.utils.retrying import retry
 from composer.utils.string_enum import StringEnum
@@ -153,7 +153,6 @@
     'KNOWN_COMPRESSORS',
     'STR_TO_DTYPE',
     'ParallelismType',
-    'create_fsdp_config',
     'FSDPConfig',
     'TPConfig',
     'ParallelismConfig',

composer/utils/parallelism.py

@@ -3,14 +3,11 @@
 
 """Parallelism configs."""
 
-import warnings
 from dataclasses import dataclass
 from typing import Any, Optional
 
 from torch.distributed._tensor.device_mesh import DeviceMesh
 
-from composer.utils.warnings import VersionedDeprecationWarning
-
 
 @dataclass
 class FSDPConfig:
@@ -45,42 +42,6 @@ class FSDPConfig:
     verbose: bool = False
 
 
-def create_fsdp_config(fsdp_config: dict[str, Any]):
-    """Modify fsdp_config to set default values for missing keys."""
-    fsdp_config = {**fsdp_config}  # Shallow copy to avoid modifying input
-    if 'process_group' in fsdp_config:
-        warnings.warn(
-            VersionedDeprecationWarning(
-                'process_group is deprecated. Please specify `data_parallel_shard_degree` and `data_parallel_replicate_degree` instead.',
-                remove_version='0.24',
-            ),
-        )
-
-    if 'device_mesh' in fsdp_config:
-        warnings.warn(
-            VersionedDeprecationWarning(
-                'device_mesh is deprecated. Please specify `data_parallel_shard_degree` and `data_parallel_replicate_degree` instead.',
-                remove_version='0.24',
-            ),
-        )
-        if 'data_parallel_shard_degree' in fsdp_config or 'data_parallel_replicate_degree' in fsdp_config:
-            raise ValueError(
-                'Cannot specify both `device_mesh` and `data_parallel_shard_degree` or `data_parallel_replicate_degree`. Please remove `device_mesh`.',
-            )
-        device_mesh = fsdp_config.pop('device_mesh')
-        if len(device_mesh) == 1:
-            fsdp_config['data_parallel_shard_degree'] = device_mesh[0]
-        elif len(device_mesh) == 2:
-            fsdp_config['data_parallel_replicate_degree'] = device_mesh[0]
-            fsdp_config['data_parallel_shard_degree'] = device_mesh[1]
-        else:
-            raise ValueError(
-                f'device_mesh must be of length 1 or 2 but received length {len(device_mesh)} with device mesh {device_mesh}.',
-            )
-
-    return FSDPConfig(**fsdp_config)
-
-
 @dataclass
 class TPConfig:
     """Configuration for tensor parallelism (TP)."""

docker/README.md

@@ -15,8 +15,8 @@ all dependencies for both NLP and Vision models. They are built on top of the
 <!-- BEGIN_COMPOSER_BUILD_MATRIX -->
 | Composer Version   | CUDA Support   | Docker Tag                                                                                                                                                       |
 |--------------------|----------------|------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| 0.23.5             | Yes            | `mosaicml/composer:latest`, `mosaicml/composer:0.23.5`                 |
-| 0.23.5             | No             | `mosaicml/composer:latest_cpu`, `mosaicml/composer:0.23.5_cpu` |
+| 0.24.0             | Yes            | `mosaicml/composer:latest`, `mosaicml/composer:0.24.0`                 |
+| 0.24.0             | No             | `mosaicml/composer:latest_cpu`, `mosaicml/composer:0.24.0_cpu` |
 <!-- END_COMPOSER_BUILD_MATRIX -->
 
 **Note**: For a lightweight installation, we recommended using a [MosaicML PyTorch Image](#pytorch-images) and manually

docker/build_matrix.yaml

@@ -194,36 +194,36 @@
   TORCHVISION_VERSION: 0.17.2
 - AWS_OFI_NCCL_VERSION: ''
   BASE_IMAGE: nvidia/cuda:12.4.1-cudnn-devel-ubuntu20.04
-  COMPOSER_INSTALL_COMMAND: mosaicml[all]==0.23.5
+  COMPOSER_INSTALL_COMMAND: mosaicml[all]==0.24.0
   CUDA_VERSION: 12.4.1
-  IMAGE_NAME: composer-0-23-5
+  IMAGE_NAME: composer-0-24-0
   MOFED_VERSION: latest-23.10
   NVIDIA_REQUIRE_CUDA_OVERRIDE: ''
   PYTHON_VERSION: '3.11'
   PYTORCH_NIGHTLY_URL: ''
   PYTORCH_NIGHTLY_VERSION: ''
   PYTORCH_VERSION: 2.4.0
   TAGS:
-  - mosaicml/composer:0.23.5
-  - ghcr.io/databricks-mosaic/composer:0.23.5
+  - mosaicml/composer:0.24.0
+  - ghcr.io/databricks-mosaic/composer:0.24.0
   - mosaicml/composer:latest
   - ghcr.io/databricks-mosaic/composer:latest
   TARGET: composer_stage
   TORCHVISION_VERSION: 0.19.0
 - AWS_OFI_NCCL_VERSION: ''
   BASE_IMAGE: ubuntu:20.04
-  COMPOSER_INSTALL_COMMAND: mosaicml[all]==0.23.5
+  COMPOSER_INSTALL_COMMAND: mosaicml[all]==0.24.0
   CUDA_VERSION: ''
-  IMAGE_NAME: composer-0-23-5-cpu
+  IMAGE_NAME: composer-0-24-0-cpu
   MOFED_VERSION: latest-23.10
   NVIDIA_REQUIRE_CUDA_OVERRIDE: ''
   PYTHON_VERSION: '3.11'
   PYTORCH_NIGHTLY_URL: ''
   PYTORCH_NIGHTLY_VERSION: ''
   PYTORCH_VERSION: 2.4.0
   TAGS:
-  - mosaicml/composer:0.23.5_cpu
-  - ghcr.io/databricks-mosaic/composer:0.23.5_cpu
+  - mosaicml/composer:0.24.0_cpu
+  - ghcr.io/databricks-mosaic/composer:0.24.0_cpu
   - mosaicml/composer:latest_cpu
   - ghcr.io/databricks-mosaic/composer:latest_cpu
   TARGET: composer_stage

docker/generate_build_matrix.py

@@ -244,7 +244,7 @@ def _main():
     composer_entries = []
 
     # The `GIT_COMMIT` is a placeholder and Jenkins will substitute it with the actual git commit for the `composer_staging` images
-    composer_versions = ['0.23.5']  # Only build images for the latest composer version
+    composer_versions = ['0.24.0']  # Only build images for the latest composer version
     composer_python_versions = [PRODUCTION_PYTHON_VERSION]  # just build composer against the latest
 
     for product in itertools.product(composer_python_versions, composer_versions, cuda_options):