sparsegpt

Utility functions of SparseGPT.

Classes

SparseGPTSearcher

SparseGPT-based sparse mask searching algorithm.

Functions

create_sgpt_mask	Create a sparse mask for the given tensor.
invert	Invert a Hessian matrix.
prepare	Prepare the inverse Hessian matrix.

class SparseGPTSearcher

Bases: BaseSparseSearcher

SparseGPT-based sparse mask searching algorithm.

after_search()

Remove Hessian artifacts from network.

before_search()

Register the forward hook to collect the hessian matrix.

config: dict[str, Any]

constraints: dict[str, str | float | dict | None]

property default_search_config: dict[str, Any]

Get the default config for the searcher.

deployment: dict[str, str] | None

dummy_input: Any | tuple

forward_loop: Callable[[Module], None] | None

model: Module

create_sgpt_mask(tensor, hessian, config)

Create a sparse mask for the given tensor.

Parameters:

• tensor (Tensor)

• hessian (Tensor)

• config (dict[str, Any])

Return type:

Tensor

invert(hessian)

Invert a Hessian matrix.

Parameters:

hessian (Tensor)

Return type:

Tensor

prepare(tensor, hessian, hessian_damp)

Prepare the inverse Hessian matrix.

Parameters:

• tensor (Tensor)

• hessian (Tensor)

• hessian_damp (float)

Return type:

tuple[Tensor, Tensor]