compressed_tensors.transform.transform_scheme

Bases: BaseModel

Scheme used to parameterize a particular transform type and specify how and where it should be applied to the model

Parameters:

• type –

  string indicating the particular transform type that should be created and applied. This should be one of the registered transform types (see Transforms.registered_names())

• apply –

  list of TransformationArgs containing the information about the modules that should be targeted by the specified transform

• randomize –

  True if uniquely randomized transform weights should be used, otherwise use identical transform weights where applicable

• requires_grad –

  True if weights include gradients for training

• head_dim –

  If set, the transform matrix will be block diagonal with each block being a square matrix of this size. The name head_dim was chosen because some rotations need to be block-diagonal with block size equal to the head_dim, but research has shown value in applying some rotations with smaller block size, irrespective of head_dim.

• precision –

  Precision at which this transform should be applied during online rotations. Fused (offline) rotations are always performed in float64