tdhook.weights.adapters#

Classes#

HookedModuleWithAdapters

Wrapper to enhance a module with hooking capabilities.

Adapters

ROME [28], sparse autoencoders [29] and transcoders [30].

Module Contents#

class tdhook.weights.adapters.HookedModuleWithAdapters(*args, adapters, **kwargs)[source]#

Bases: tdhook.modules.HookedModule

Wrapper to enhance a module with hooking capabilities.

Parameters:

adapters (Dict[str, torch.nn.Module])

adapters[source]#
class tdhook.weights.adapters.Adapters(adapters, cache_callback=None, relative=True, directions=None, cache=None, clear_cache=True)[source]#

Bases: tdhook.contexts.HookingContextFactory

ROME [28], sparse autoencoders [29] and transcoders [30].

Parameters:
  • adapters (Dict[str, Tuple[torch.nn.Module, str, str]])

  • cache_callback (Optional[Callable])

  • relative (bool)

  • directions (Optional[List[tdhook.hooks.HookDirection]])

  • cache (Optional[tensordict.TensorDict])

  • clear_cache (bool)

_hooked_module_class[source]#
_hooking_context_class[source]#
_adapters[source]#
_cache_callback = None[source]#
_relative = True[source]#
_directions = ['fwd'][source]#
_hook_module(module)[source]#
Parameters:

module (tdhook.modules.HookedModule)

Return type:

tdhook.hooks.MultiHookHandle