nemo_automodel.components.utils.compile_utils

Module Contents

Classes

Name	Description
`CompileConfig`	Configuration for torch.compile.

Functions

Name	Description
`_resolve_compile_kwargs`	Translate a CompileConfig into torch.compile keyword arguments.
`apply_flash_attention_compile_fix`	Apply the Flash Attention + torch.compile compatibility fix.
`build_compile_config`	Build a compile config from configuration.
`compile_model`	Compile the model with Flash Attention compatibility.
`compile_module_inplace`	In-place variant of `compile_model` using `nn.Module.compile()`.
`configure_torch_dynamo`	Configure torch._dynamo settings for compilation.
`create_compile_config_from_dict`	Create a CompileConfig from a dictionary.
`enable_torch_dynamo_scalar_outputs`	Enable torch.dynamo to capture scalar outputs for better Flash Attention + torch.compile compatibility.
`patch_prepare_fa2_from_position_ids`	Apply a simple targeted patch to fix the prepare_fa2_from_position_ids function

Data

_FLASH_ATTENTION_FIX_APPLIED

logger

API

class nemo_automodel.components.utils.compile_utils.CompileConfig(
    enabled: bool = False,
    mode: str = 'default',
    fullgraph: bool = False,
    dynamic: bool = False,
    backend: typing.Optional[str] = None,
    options: typing.Optional[typing.Dict[str, typing.Any]] = None,
    dynamo_cache_size_limit: int = 256
)

Dataclass

Configuration for torch.compile.

options

Optional[Dict[str, Any]] = options or {}

nemo_automodel.components.utils.compile_utils.CompileConfig.to_dict() -> typing.Dict[str, typing.Any]

Convert to dictionary.

nemo_automodel.components.utils.compile_utils._resolve_compile_kwargs(
    config: nemo_automodel.components.utils.compile_utils.CompileConfig
) -> typing.Dict[str, typing.Any]

Translate a CompileConfig into torch.compile keyword arguments.

torch.compile is keyword-only, rejects unknown kwargs, and treats mode and options as mutually exclusive — inductor options must go through the options dict, not be flattened into the call. A non-empty CompileConfig.options therefore replaces mode, and combining options with a non-default mode is an explicit error.

nemo_automodel.components.utils.compile_utils.apply_flash_attention_compile_fix()

Apply the Flash Attention + torch.compile compatibility fix.

This enables scalar output capture and patches the key function that causes issues. Note: This function is focused solely on Flash Attention compatibility. For dynamo configuration (cache size, etc.), use configure_torch_dynamo() separately.

nemo_automodel.components.utils.compile_utils.build_compile_config(
    cfg: typing.Optional[typing.Dict[str, typing.Any]]
) -> nemo_automodel.components.utils.compile_utils.CompileConfig

Build a compile config from configuration.

Parameters:

cfg

Optional[Dict[str, Any]]

Configuration dictionary for compilation.

Returns: CompileConfig

CompileConfig instance.

nemo_automodel.components.utils.compile_utils.compile_model(
    model: torch.nn.Module,
    config: nemo_automodel.components.utils.compile_utils.CompileConfig
) -> torch.nn.Module

Compile the model with Flash Attention compatibility.

Parameters:

model

nn.Module

The model to compile.

config

CompileConfig

Compile configuration.

Returns: nn.Module

The compiled model.

nemo_automodel.components.utils.compile_utils.compile_module_inplace(
    module: torch.nn.Module,
    config: nemo_automodel.components.utils.compile_utils.CompileConfig
) -> None

In-place variant of compile_model using nn.Module.compile().

Keeps the module object identity and state-dict keys unchanged (no OptimizedModule wrapper, no _orig_mod. key prefix), for callers that track the module by reference or save its state dict directly. Compilation is lazy: errors surface at the first forward, as with torch.compile.

nemo_automodel.components.utils.compile_utils.configure_torch_dynamo(
    cache_size_limit: int = 256,
    capture_scalar_outputs: bool = True
)

Configure torch._dynamo settings for compilation.

Parameters:

cache_size_limit

intDefaults to 256

Cache size limit for dynamo compilation

capture_scalar_outputs

boolDefaults to True

Whether to capture scalar outputs for Flash Attention compatibility

nemo_automodel.components.utils.compile_utils.create_compile_config_from_dict(
    config_dict: typing.Dict[str, typing.Any]
) -> nemo_automodel.components.utils.compile_utils.CompileConfig

Create a CompileConfig from a dictionary.

Parameters:

config_dict

Dict[str, Any]

Dictionary containing compile configuration.

Returns: CompileConfig

CompileConfig instance.

nemo_automodel.components.utils.compile_utils.enable_torch_dynamo_scalar_outputs()

Enable torch.dynamo to capture scalar outputs for better Flash Attention + torch.compile compatibility.

nemo_automodel.components.utils.compile_utils.patch_prepare_fa2_from_position_ids()

Apply a simple targeted patch to fix the prepare_fa2_from_position_ids function for torch.compile compatibility.

This is the key function that needs the fix for the max_length computation.

nemo_automodel.components.utils.compile_utils._FLASH_ATTENTION_FIX_APPLIED = apply_flash_attention_compile_fix()

nemo_automodel.components.utils.compile_utils.logger = logging.getLogger(__name__)