modalities.running_env package

Subpackages

• modalities.running_env.fsdp package
  • Submodules
  • modalities.running_env.fsdp.device_mesh module
    • DeviceMeshConfig
      • DeviceMeshConfig.context_parallel_degree
      • DeviceMeshConfig.data_parallel_replicate_degree
      • DeviceMeshConfig.data_parallel_shard_degree
      • DeviceMeshConfig.device_type
      • DeviceMeshConfig.enable_loss_parallel
      • DeviceMeshConfig.model_config
      • DeviceMeshConfig.pipeline_parallel_degree
      • DeviceMeshConfig.tensor_parallel_degree
      • DeviceMeshConfig.world_size
    • ParallelismDegrees
      • ParallelismDegrees.CP
      • ParallelismDegrees.DP_REPLICATE
      • ParallelismDegrees.DP_SHARD
      • ParallelismDegrees.PP
      • ParallelismDegrees.TP
    • get_device_mesh()
  • modalities.running_env.fsdp.fsdp_auto_wrapper module
    • FSDPAutoWrapFactoryIF
      • FSDPAutoWrapFactoryIF.get_auto_wrap_policy()
    • FSDPAutoWrapFactoryTypes
      • FSDPAutoWrapFactoryTypes.FSDPTransformerAutoWrapPolicyFactory
    • FSDPTransformerAutoWrapPolicyFactory
      • FSDPTransformerAutoWrapPolicyFactory.get_auto_wrap_policy()
  • modalities.running_env.fsdp.reducer module
    • Reducer
      • Reducer.reduce()
  • Module contents

Submodules

modalities.running_env.cuda_env module

class modalities.running_env.cuda_env.CudaEnv(process_group_backend)

Bases: object

Context manager to set the CUDA environment for distributed training.

Initializes the CudaEnv context manager with the process group backend.

Args:

process_group_backend (ProcessGroupBackendType): Process group backend to be used for distributed training.

Parameters:

process_group_backend (ProcessGroupBackendType)

modalities.running_env.env_utils module

class modalities.running_env.env_utils.FSDP2MixedPrecisionSettings(**data)

Bases: BaseModel

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.

Parameters:

• param_dtype (PyTorchDtypes)

• reduce_dtype (PyTorchDtypes)

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

param_dtype: PyTorchDtypes

reduce_dtype: PyTorchDtypes

class modalities.running_env.env_utils.MixedPrecisionSettings(value)

Bases: LookupEnum

BF_16 = MixedPrecision(param_dtype=torch.bfloat16, reduce_dtype=torch.bfloat16, buffer_dtype=torch.bfloat16, keep_low_precision_grads=False, cast_forward_inputs=False, cast_root_forward_inputs=True, _module_classes_to_ignore=(<class 'torch.nn.modules.batchnorm._BatchNorm'>,))

BF_16_WORKING = MixedPrecision(param_dtype=torch.float32, reduce_dtype=torch.bfloat16, buffer_dtype=torch.bfloat16, keep_low_precision_grads=False, cast_forward_inputs=False, cast_root_forward_inputs=True, _module_classes_to_ignore=(<class 'torch.nn.modules.batchnorm._BatchNorm'>,))

FP_16 = MixedPrecision(param_dtype=torch.float16, reduce_dtype=torch.float16, buffer_dtype=torch.float16, keep_low_precision_grads=False, cast_forward_inputs=False, cast_root_forward_inputs=True, _module_classes_to_ignore=(<class 'torch.nn.modules.batchnorm._BatchNorm'>,))

FP_32 = MixedPrecision(param_dtype=torch.float32, reduce_dtype=torch.float32, buffer_dtype=torch.float32, keep_low_precision_grads=False, cast_forward_inputs=False, cast_root_forward_inputs=True, _module_classes_to_ignore=(<class 'torch.nn.modules.batchnorm._BatchNorm'>,))

MIXED_PRECISION_MEGATRON = MixedPrecision(param_dtype=torch.bfloat16, reduce_dtype=torch.float32, buffer_dtype=None, keep_low_precision_grads=False, cast_forward_inputs=False, cast_root_forward_inputs=True, _module_classes_to_ignore=(<class 'torch.nn.modules.batchnorm._BatchNorm'>,))

NO_MIXED_PRECISION = None

class modalities.running_env.env_utils.PyTorchDtypes(value)

Bases: LookupEnum

BF_16 = torch.bfloat16

FP_16 = torch.float16

FP_32 = torch.float32

modalities.running_env.env_utils.has_bfloat_support()

modalities.running_env.env_utils.is_running_with_torchrun()

Module contents