Miscellaneous#

Status#

class cutlass.operators.status.Status(error: Exception | None = None)#

Bases: object

A simple status class to wrap an optional exception.

__bool__() → bool#

Return True if this represents success, False otherwise.

Example

>>> if not (status := operator.supports(args)):
...     raise status.error
classmethod success() → Status#

Create a successful status.

classmethod fail(
error: str | Exception,
) → Status#

Create a failed status with an error.

Parameters:

error (str | Exception) – Error string or Exception describing the failure

raise_on_error() → None#

Raise the stored exception if this status represents a failure.

Configuration#

class cutlass.operators.config.GlobalOptions#

Bases: object

Singleton class that configures global options for CUTLASS Operator API.

This can be used to enable or disable certain features of the API. For example, the user can enable TVM-FFI support to allow for the use of framework-level tensors at run time.

Each option may also be initialized from an environment variable read the first time GlobalOptions() is instantiated. Programmatic changes via the property setters always override the environment value afterwards.

Supported environment variables:

  • CUTLASS_OPERATORS_USE_TVM_FFI –> GlobalOptions().use_tvm_ffi

Example:

GlobalOptions().use_tvm_ffi = True
property use_tvm_ffi: bool#

Check if TVM FFI support is enabled.

Default: True if tvm_ffi is installed.

Can be overridden at process start by setting the CUTLASS_OPERATORS_USE_TVM_FFI environment variable to a boolean value (1/0, true/false, yes/no, on/off).

When enabled, conversions from DLPack compatible tensors to cute.Tensor is via TVM FFI. Additionally, invoking the compiled Operator happens via TVM FFI. Both can offer significant (3x-10x) speedups.

Dependencies:

  • Required: tvm_ffi (pip install apache-tvm-ffi).

  • Optional: torch_c_dlpack_ext (pip install torch-c-dlpack-ext).

save() → dict#

Save the current options to a dictionary.

restore(inp: dict) → None#

Restore previously saved options from a dictionary.