TorchLens Removed-Spellings Ledger¶

Every warning-producing compatibility shim listed below was REMOVED outright in the 2026-08 shim-removal pass (before launch the policy is to remove and rename, not to keep compatibility shims). The old spelling now raises (AttributeError/TypeError); the "New name" column is the spelling to use. Torch-version compatibility and artifact-format compatibility (tlspec floors, legacy-save loading) are NOT shims and are unaffected.

Top-Level Moved Names¶

Old name	New name	Status
`ActivationPostfunc`	none (no longer public)	removed
`Buffer`	`torchlens.types.Buffer`	removed
`FuncCallLocation`	`torchlens.types.FuncCallLocation`	removed
`GradientPostfunc`	none (no longer public)	removed
`GradFnAccessor`	none (no longer public)	removed
`GradFn`	`torchlens.types.GradFn`	removed
`GradFnCall`	`torchlens.types.GradFnCall`	removed
`LayerAccessor`	none (no longer public)	removed
`MetadataInvariantError`	`torchlens.errors.MetadataInvariantError`	removed
`MutatedReferenceError`	`torchlens.errors.MutatedReferenceError`	removed
`ModuleAccessor`	none (no longer public)	removed
`Module`	`torchlens.types.Module`	removed
`ModuleCall`	`torchlens.types.ModuleCall`	removed
`NodeSpec`	`torchlens.viz.NodeSpec`	removed
`Param`	`torchlens.types.Param`	removed
`PostTraceParamUnavailable`	`torchlens.errors.ParamReleasedError`	removed
`TraceState`	`torchlens.io.TraceState`	removed
`SaveLevel`	none (no longer public)	removed
`SiteTable`	none (no longer public)	removed
`SpecCompat`	`torchlens.types.SpecCompat`	removed
`StreamingOptions`	`torchlens.options.StorageOptions`	removed
`TargetManifestDiff`	`torchlens.types.TargetManifestDiff`	removed
`TensorLog`	none (no longer public)	removed
`TensorSliceSpec`	none (no longer public)	removed
`TorchLensPostfuncError`	`torchlens.errors.UserTransformError`	removed
`TrainingModeConfigError`	`torchlens.errors.TrainingModeConfigError`	removed
`VisualizationOptions`	`torchlens.options.VisualizationOptions`	removed
`build_render_audit`	none (no longer public)	removed
`check_metadata_invariants`	`torchlens.validation.check_metadata_invariants`	removed
`check_spec_compat`	`torchlens.intervention.check_spec_compat`	removed
`cleanup_tmp`	`torchlens.io.cleanup_tmp`	removed
`get_model_metadata`	`torchlens.trace(model, x, save="none")`	removed
`list_logs`	`torchlens.io.list_traces`	removed
`log_model_metadata`	`torchlens.trace(model, x, save="none")`	removed
`preview_fastlog`	none (no longer public)	removed
`rehydrate_nested`	`torchlens.io.materialize_nested`	removed
`render_lines_to_html`	none (no longer public)	removed
`reset_naming_counter`	`torchlens.io.reset_naming_counter`	removed
`resolve_sites`	`torchlens.Trace` indexing (`trace[selector]` returns every matching site as a `torchlens.Selection`)	removed
`save_intervention`	`torchlens.io.save_intervention`	removed
`suppress_mutate_warnings`	`torchlens.errors.quiet()`	removed
`unwrap_torch`	`torchlens.backends.torch.wrappers.unwrap_torch`	removed
`validate_batch_of_models_and_inputs`	none: call `torchlens.validate(model, x)` per model	removed
`wrap_torch`	`torchlens.backends.torch.wrappers.wrap_torch`	removed
`wrapped`	`torchlens.backends.torch.wrappers.wrapped`	removed

Top-Level Paper-Era Names¶

Old name	New name	Status
`log_forward_pass`	`trace`	removed
`validate_model_activations`	`validate(scope="forward")`	removed
`validate_saved_activations`	`validate(scope="saved")`	removed
`render_graph`	`Trace.draw()` (or `torchlens.draw`, the top-level spelling below)	removed
`render_model_graph`	`Trace.draw()` (or `torchlens.draw`, the top-level spelling below)	removed
`draw_model_graph`	`Trace.draw()` (or `torchlens.draw`, the top-level spelling below)	removed
`ModelHistory`	`Trace`	removed
`get_model_structure`	structure trace accessors	removed
`show_model_structure`	structure trace accessors	removed

Top-Level Convenience Aliases¶

Old name	New name	Status
`peek`	`pluck`	removed
`batched_extract`	`extract_dataset`	removed
`validate_forward_pass`	`torchlens.validation.validate_forward_pass`	removed
`validate_backward_pass`	`torchlens.validation.validate_backward_pass`	removed
`validate_saved_activations`	`torchlens.validate(scope="saved")`	removed
`summary`	`torchlens.summary`	removed
`show_model_graph`	`torchlens.draw`	removed
`draw_backward`	`torchlens.draw`	removed; pass `graph="backward"`
`draw_combined`	`torchlens.draw`	removed; pass `graph="combined"`
`show_bundle_graph`	`torchlens.Bundle.draw`	removed; call `bundle.draw(path)` (`graph=` replaces `signal=`, `node_overrides=` replaces `vis_node_overrides=`, `open_viewer=True` opens a viewer)
`load_intervention_spec`	`torchlens.io.load_intervention_spec`	removed
`ModuleInputSnapshot`	`torchlens.types.ModuleInputSnapshot`	removed
`PreHookEffect`	`torchlens.types.PreHookEffect`	removed
`TensorInputObservation`	`torchlens.types.TensorInputObservation`	removed

Capture And Option Keyword Aliases¶

Old name	New name	Status
`capture_output_structure`	`capture_container_structure`	removed
`show_containers="auto"`	`show_containers="collapsed"`	removed
`show_containers="cluster"`	`show_containers="nodes"`	removed
`layers_to_save`	`save=...`	removed
`random_seed`	grouped capture options	removed
`save_grads`	backward capture options	removed
`vis_node_mode`	`node_style`	removed
`vis_opt`	`view` (full alias chain `vis_opt` -> `vis_mode` -> `view`; the warning names `view`)	removed
flat `CaptureOptions` fields	grouped option fields	removed
`mark_layer_depths`	`capture.compute_input_output_distances`	removed
`num_context_lines`	`capture.source_context_lines`	removed
`mode`	`visualization.view`	removed
`max_module_depth`	`visualization.depth`	removed
`layout_engine`	`visualization.layout`	removed
`node_mode`	`visualization.node_style`	removed
`save_outs_to`	`streaming.path`	removed
`keep_outs_in_memory`	`streaming.retain_in_memory`	removed
`out_sink`	`streaming.on_save`	removed
`log_forward_pass(layers=...)`	`trace(save=...)`	removed
`log_forward_pass(save_function_args=...)`	`trace(save_args=...)`	removed
`log_forward_pass(save_gradients=...)`	`trace(save_grads=...)`	removed
`log_forward_pass(keep_unsaved_layers=...)`	no direct equivalent; see migration guide	removed
flat visualization option fields	grouped visualization option fields	removed

Recording And Fastlog Aliases¶

Old name	New name	Status
`record(keep_op=...)`	`record(save=...)`	REMOVED (predicate consolidation)
`record(keep_module=...)`	no equivalent; see note below	REMOVED (predicate consolidation)
`Recorder(keep_op=...)`	`Recorder(save=...)`	REMOVED (predicate consolidation)
`Recorder(keep_module=...)`	no equivalent; see note below	REMOVED (predicate consolidation)
`dry_run(keep_op=...)`	`dry_run(save=...)`	RENAMED (predicate consolidation; was never deprecation-warned)
`dry_run(keep_module=...)`	no equivalent; see note below	REMOVED (predicate consolidation)

keep_module note. Predicate-gated module-event selection was REMOVED, not migrated: save= routes to the op predicate only, and default_module= records ALL module enter/exit boundary events uniformly — it is not a per-module predicate. A module predicate passed via save= selects zero module events. The internal keep_module options slot and its two public simulators were deleted with it (Trace.preview_fastlog(keep_module=...) and RecordingTrace.repredicate(other_keep_module=...) no longer accept module predicates; both raise TypeError), so no surface simulates a configuration that capture cannot produce.

record(save=...) and dry_run(save=...) now default to None (previously an internal MISSING sentinel that existed only to arbitrate against the removed keep_op= alias); omitted and save=None are equivalent.

Trace And Conditional Aliases¶

Old name	New name	Status
`Trace.conditional_then_entry_edges`	`Trace.graph.conditional_arm_entry_edges`	removed
`Trace.conditional_elif_entry_edges`	`Trace.graph.conditional_arm_entry_edges`	removed
`Trace.conditional_else_entry_edges`	`Trace.graph.conditional_arm_entry_edges`	removed
`Trace.validate_saved_outs()`	`Trace.validate_forward_pass()`	removed
`Trace.replay()`	`Trace.do(site, edit)`	removed
`Trace.replay_from()`	`Trace.do(site, value)`	removed
`Trace.push()`	`Trace.do(site, edit)`: one verb stages and propagates	removed
`Trace.push_from()`	`Trace.do(site, value)`	removed
`Trace.attach_hooks()`, `Op.attach_hooks()`, `Layer.attach_hooks()`	`Trace.do(site, hook)`; `InterventionOptions(engine="set_only")` stages without propagating	removed
`Trace.set()`, `Layer.set()`, `Op.set()`	`Trace.do(site, value, intervention=InterventionOptions(engine="set_only"))`	removed
`Trace.remove()` and `with trace.attach_hooks(...)`	`Trace.clear_edits(site)`; the scoped form stays as `with trace.fork() as scoped: scoped.do(...)`, which detaches the last `do` hooks on exit	removed
`ReplayOptions(hooks=)`	a mapping of sites to edits on `Trace.do`	removed
`ReplayOptions(differentiable=)`	`InterventionOptions(differentiable=True)` on `Trace.do`	removed
`ReplayOptions(append=)`	`Trace.append(inputs, model=model)`	removed
`Trace.rerun()`	`Trace.run()`	removed

Record Members¶

Old name	New name	Status
`Op.gpu_kernels`	none in this release (artifacts keep the kernel annotation)	removed
`AtenOp.gpu_kernels`	none in this release (artifacts keep the kernel annotation)	removed
`Op.activation_memory`	`Op.cost.activation_memory`	removed
`Op.arg_expressions`	`Op.source.arg_expressions`	removed
`Op.arg_names`	`Op.call.arg_names`	removed
`Op.args_template`	`Op.call.args_template`	removed
`Op.arithmetic_intensity`	`Op.cost.arithmetic_intensity`	removed
`Op.atomic_module`	`Op.module_context.atomic_module`	removed
`Op.atomic_module_address`	`Op.module_context.atomic_module_address`	removed
`Op.atomic_module_call`	`Op.module_context.atomic_module_call`	removed
`Op.autograd_memory`	`Op.cost.autograd_memory`	removed
`Op.bool_value`	`Op.control_flow.bool_value`	removed
`Op.buffer_sink_ops`	`Op.graph.buffer_sink_ops`	removed
`Op.buffer_source`	`Op.graph.buffer_source`	removed
`Op.buffer_source_func_name`	`Op.graph.buffer_source_func_name`	removed
`Op.buffer_source_ops`	`Op.graph.buffer_source_ops`	removed
`Op.bytes_delta_at_call`	`Op.cost.bytes_delta_at_call`	removed
`Op.bytes_peak_at_call`	`Op.cost.bytes_peak_at_call`	removed
`Op.bytes_read`	`Op.cost.bytes_read`	removed
`Op.bytes_written`	`Op.cost.bytes_written`	removed
`Op.co_parents`	`Op.graph.co_parents`	removed
`Op.code_context`	`Op.source.frames`	removed
`Op.compute_record`	`Op.cost.compute_record`	removed
`Op.conditional_arm_children`	`Op.control_flow.arm_children`	removed
`Op.conditional_branch_depth`	`Op.control_flow.branch_depth`	removed
`Op.conditional_branch_stack`	`Op.control_flow.branch_stack`	removed
`Op.conditional_context_kind`	`Op.control_flow.context_kind`	removed
`Op.conditional_depth`	`Op.control_flow.depth`	removed
`Op.conditional_elif_children`	`Op.control_flow.elif_children`	removed
`Op.conditional_else_children`	`Op.control_flow.else_children`	removed
`Op.conditional_entry_children`	`Op.control_flow.entry_children`	removed
`Op.conditional_then_children`	`Op.control_flow.then_children`	removed
`Op.conditional_wrapper_kind`	`Op.control_flow.wrapper_kind`	removed
`Op.derived_grad`	`Op.autograd.derived_grad`	removed
`Op.dropped_edge_tensor_args`	`Op.call.dropped_edge_tensor_args`	removed
`Op.edge_substitutions`	`Op.graph.edge_substitutions`	removed
`Op.equivalence_class`	`Op.graph.equivalence_class`	removed
`Op.equivalent_ops`	`Op.graph.equivalent_ops`	removed
`Op.flops_backward`	`Op.cost.flops_backward`	removed
`Op.flops_forward`	`Op.cost.flops_forward`	removed
`Op.flops_total`	`Op.cost.flops_total`	removed
`Op.func_autocast_state`	`Op.call.autocast_state`	removed
`Op.func_call_id`	`Op.call.call_id`	removed
`Op.func_config`	`Op.call.config`	removed
`Op.func_duration`	`Op.cost.duration`	removed
`Op.func_id`	`Op.call.func_id`	removed
`Op.func_non_tensor_args`	`Op.call.non_tensor_args`	removed
`Op.func_qualname`	`Op.source.func_qualname`	removed
`Op.func_rng_states`	`Op.call.rng_states`	removed
`Op.fx_call_index`	`Op.source.fx_call_index`	removed
`Op.fx_qualpath`	`Op.source.fx_path`	removed
`Op.grad_dtype`	`Op.autograd.grad_dtype`	removed
`Op.grad_fn`	`Op.autograd.grad_fn`	removed
`Op.grad_fn_class_name`	`Op.autograd.grad_fn_class_name`	removed
`Op.grad_fn_class_qualname`	`Op.autograd.grad_fn_class_qualname`	removed
`Op.grad_fn_handle`	`Op.autograd.grad_fn_handle`	removed
`Op.grad_for`	`Op.autograd.grad_for`	removed
`Op.grad_shape`	`Op.autograd.grad_shape`	removed
`Op.grads`	`Op.autograd.grads`	removed
`Op.has_children`	`Op.graph.has_children`	removed
`Op.has_co_parents`	`Op.graph.has_co_parents`	removed
`Op.has_input_ancestor`	`Op.graph.has_input_ancestor`	removed
`Op.has_internal_source_ancestor`	`Op.graph.has_internal_source_ancestor`	removed
`Op.has_output_descendant`	`Op.graph.has_output_descendant`	removed
`Op.has_parents`	`Op.graph.has_parents`	removed
`Op.has_saved_args`	`Op.call.has_saved_args`	removed
`Op.has_siblings`	`Op.graph.has_siblings`	removed
`Op.in_conditionals`	`Op.control_flow.in_conditionals`	removed
`Op.injection_provenance`	`Op.call.injection_provenance`	removed
`Op.input_memory`	`Op.cost.input_memory`	removed
`Op.input_to_module_calls`	`Op.module_context.input_to_module_calls`	removed
`Op.input_to_modules`	`Op.module_context.input_to_modules`	removed
`Op.internal_source_ancestors`	`Op.graph.internal_source_ancestors`	removed
`Op.internal_source_parents`	`Op.graph.internal_source_parents`	removed
`Op.io_role`	`Op.graph.io_role`	removed
`Op.is_atomic_module`	`Op.module_context.is_atomic_module`	removed
`Op.is_buffer`	`Op.graph.is_buffer`	removed
`Op.is_compute_op`	`Op.graph.is_compute`	removed
`Op.is_final_output`	`Op.graph.is_final_output`	removed
`Op.is_in_conditional`	`Op.control_flow.is_in_conditional`	removed
`Op.is_in_conditional_body`	`Op.control_flow.is_in_conditional_body`	removed
`Op.is_in_conditional_evaluation`	`Op.control_flow.is_in_conditional_evaluation`	removed
`Op.is_internal_source`	`Op.graph.is_internal_source`	removed
`Op.is_internally_initialized`	`Op.call.is_internally_initialized`	removed
`Op.is_module_input`	`Op.module_context.is_module_input`	removed
`Op.is_module_output`	`Op.module_context.is_module_output`	removed
`Op.is_orphan`	`Op.graph.is_orphan`	removed
`Op.is_output_parent`	`Op.graph.is_output_parent`	removed
`Op.is_scalar_bool`	`Op.control_flow.is_scalar_bool`	removed
`Op.is_terminal_bool`	`Op.control_flow.is_terminal_bool`	removed
`Op.is_terminal_conditional_bool`	`Op.control_flow.is_terminal_conditional_bool`	removed
`Op.is_transform`	`Op.call.is_func_transform`	removed
`Op.kwargs_template`	`Op.call.kwargs_template`	removed
`Op.macs_forward`	`Op.cost.macs_forward`	removed
`Op.materialize_grad`	`Op.payload.materialize_grad`	removed
`Op.materialize_out`	`Op.payload.materialize_out`	removed
`Op.max_distance_from_input`	`Op.graph.max_distance_from_input`	removed
`Op.max_distance_to_output`	`Op.graph.max_distance_to_output`	removed
`Op.min_distance_from_input`	`Op.graph.min_distance_from_input`	removed
`Op.min_distance_to_output`	`Op.graph.min_distance_to_output`	removed
`Op.module_call_depth`	`Op.module_context.call_depth`	removed
`Op.module_call_stack`	`Op.module_context.call_stack`	removed
`Op.module_entry_arg_keys`	`Op.module_context.entry_arg_keys`	removed
`Op.non_tensor_kwargs`	`Op.call.non_tensor_kwargs`	removed
`Op.non_tensor_pos_args`	`Op.call.non_tensor_pos_args`	removed
`Op.num_args_total`	`Op.call.num_args_total`	removed
`Op.num_autograd_tensors`	`Op.autograd.num_autograd_tensors`	removed
`Op.num_inputs`	`Op.graph.num_inputs`	removed
`Op.num_kwargs`	`Op.call.num_kwargs`	removed
`Op.num_pos_args`	`Op.call.num_pos_args`	removed
`Op.output_of_module_calls`	`Op.module_context.output_of_module_calls`	removed
`Op.output_of_modules`	`Op.module_context.output_of_modules`	removed
`Op.param_memory`	`Op.cost.param_memory`	removed
`Op.parent_arg_positions`	`Op.call.parent_arg_positions`	removed
`Op.parent_param_ops`	`Op.graph.parent_param_ops`	removed
`Op.parent_params`	`Op.graph.parent_params`	removed
`Op.recurrent_ops`	`Op.graph.recurrent_ops`	removed
`Op.root_ancestors`	`Op.graph.root_ancestors`	removed
`Op.saved_args`	`Op.call.saved_args`	removed
`Op.saved_kwargs`	`Op.call.saved_kwargs`	removed
`Op.siblings`	`Op.graph.siblings`	removed
`Op.terminal_bool_for`	`Op.control_flow.terminal_bool_for`	removed
`Op.tl_authored_root`	`Op.call.tl_authored_root`	removed
`Op.transform_chain`	`Op.call.func_transform_chain`	removed
`Op.transform_config`	`Op.call.func_transform_config`	removed
`Op.transform_fn_name`	`Op.call.func_transform_fn_name`	removed
`Op.transform_fn_qualname`	`Op.call.func_transform_fn_qualname`	removed
`Op.transform_fn_source`	`Op.call.func_transform_fn_source`	removed
`Op.transform_kind`	`Op.call.func_transform_kind`	removed
`Op.transformed_activation_memory`	`Op.cost.transformed_activation_memory`	removed
`Op.transformed_grad`	`Op.payload.transformed_grad`	removed
`Op.transformed_grad_dtype`	`Op.payload.transformed_grad_dtype`	removed
`Op.transformed_grad_shape`	`Op.payload.transformed_grad_shape`	removed
`Op.transformed_out`	`Op.payload.transformed_out`	removed
`Op.transformed_out_dtype`	`Op.payload.transformed_out_dtype`	removed
`Op.transformed_out_shape`	`Op.payload.transformed_out_shape`	removed
`Op.unattributed_tensor_args`	`Op.call.unattributed_tensor_args`	removed
`Op.var_names`	`Op.source.var_names`	removed
`Layer.activation_memory`	`Layer.cost.activation_memory`	removed
`Layer.arg_names`	`Layer.call.arg_names`	removed
`Layer.autograd_memory`	`Layer.cost.autograd_memory`	removed
`Layer.bool_value`	`Layer.control_flow.bool_value`	removed
`Layer.buffer_pass`	`Layer.graph.buffer_pass`	removed
`Layer.buffer_source`	`Layer.graph.buffer_source`	removed
`Layer.buffer_source_func_name`	`Layer.graph.buffer_source_func_name`	removed
`Layer.child_ops_per_layer`	`Layer.graph.child_ops_per_layer`	removed
`Layer.children_per_pass`	`Layer.graph.children_per_pass`	removed
`Layer.code_context`	`Layer.source.frames`	removed
`Layer.conditional_arm_children`	`Layer.control_flow.arm_children`	removed
`Layer.conditional_branch_stack_ops`	`Layer.control_flow.branch_stack_ops`	removed
`Layer.conditional_depth`	`Layer.control_flow.depth`	removed
`Layer.conditional_elif_children`	`Layer.control_flow.elif_children`	removed
`Layer.conditional_else_children`	`Layer.control_flow.else_children`	removed
`Layer.conditional_entry_children`	`Layer.control_flow.entry_children`	removed
`Layer.conditional_role_stacks`	`Layer.control_flow.role_stacks`	removed
`Layer.conditional_then_children`	`Layer.control_flow.then_children`	removed
`Layer.equivalence_class`	`Layer.graph.equivalence_class`	removed
`Layer.equivalent_ops`	`Layer.graph.equivalent_ops`	removed
`Layer.flops_backward`	`Layer.cost.flops_backward`	removed
`Layer.flops_forward`	`Layer.cost.flops_forward`	removed
`Layer.flops_total`	`Layer.cost.flops_total`	removed
`Layer.func_config`	`Layer.call.config`	removed
`Layer.func_duration`	`Layer.cost.duration`	removed
`Layer.func_rng_states`	`Layer.call.rng_states`	removed
`Layer.fx_call_index`	`Layer.source.fx_call_index`	removed
`Layer.fx_qualpath`	`Layer.source.fx_path`	removed
`Layer.grad_fn`	`Layer.autograd.grad_fn`	removed
`Layer.grad_fn_class_name`	`Layer.autograd.grad_fn_class_name`	removed
`Layer.grad_fn_class_qualname`	`Layer.autograd.grad_fn_class_qualname`	removed
`Layer.grad_fn_handle`	`Layer.autograd.grad_fn_handle`	removed
`Layer.has_children`	`Layer.graph.has_children`	removed
`Layer.has_co_parents`	`Layer.graph.has_co_parents`	removed
`Layer.has_input_ancestor`	`Layer.graph.has_input_ancestor`	removed
`Layer.has_parents`	`Layer.graph.has_parents`	removed
`Layer.has_siblings`	`Layer.graph.has_siblings`	removed
`Layer.in_conditionals`	`Layer.control_flow.in_conditionals`	removed
`Layer.io_role`	`Layer.graph.io_role`	removed
`Layer.is_atomic_module`	`Layer.module_context.is_atomic_module`	removed
`Layer.is_buffer`	`Layer.graph.is_buffer`	removed
`Layer.is_compute_layer`	`Layer.graph.is_compute`	removed
`Layer.is_final_output`	`Layer.graph.is_final_output`	removed
`Layer.is_in_conditional`	`Layer.control_flow.is_in_conditional`	removed
`Layer.is_in_conditional_body`	`Layer.control_flow.is_in_conditional_body`	removed
`Layer.is_in_conditional_evaluation`	`Layer.control_flow.is_in_conditional_evaluation`	removed
`Layer.is_internal_sink`	`Layer.graph.is_internal_sink`	removed
`Layer.is_internal_source`	`Layer.graph.is_internal_source`	removed
`Layer.is_module_input`	`Layer.module_context.is_module_input`	removed
`Layer.is_orphan`	`Layer.graph.is_orphan`	removed
`Layer.is_scalar_bool`	`Layer.control_flow.is_scalar_bool`	removed
`Layer.is_terminal_bool`	`Layer.control_flow.is_terminal_bool`	removed
`Layer.macs_forward`	`Layer.cost.macs_forward`	removed
`Layer.module_call_depth`	`Layer.module_context.call_depth`	removed
`Layer.num_args_total`	`Layer.call.num_args_total`	removed
`Layer.num_autograd_tensors`	`Layer.autograd.num_autograd_tensors`	removed
`Layer.num_kwargs`	`Layer.call.num_kwargs`	removed
`Layer.num_pos_args`	`Layer.call.num_pos_args`	removed
`Layer.output_of_module_calls`	`Layer.module_context.output_of_module_calls`	removed
`Layer.output_of_modules`	`Layer.module_context.output_of_modules`	removed
`Layer.parent_arg_positions`	`Layer.call.parent_arg_positions`	removed
`Layer.parent_ops_per_layer`	`Layer.graph.parent_ops_per_layer`	removed
`Layer.parents_per_pass`	`Layer.graph.parents_per_pass`	removed
`Layer.saved_args`	`Layer.call.saved_args`	removed
`Layer.saved_kwargs`	`Layer.call.saved_kwargs`	removed
`Layer.site_peers`	`Layer.graph.site_peers`	removed
`Layer.terminal_bool_for`	`Layer.control_flow.terminal_bool_for`	removed
`Layer.total_activation_memory`	`Layer.cost.total_activation_memory`	removed
`Layer.total_autograd_memory`	`Layer.cost.total_autograd_memory`	removed
`Layer.total_flops_backward`	`Layer.cost.total_flops_backward`	removed
`Layer.total_flops_forward`	`Layer.cost.total_flops_forward`	removed
`Layer.total_flops_total`	`Layer.cost.total_flops`	removed
`Layer.total_func_duration`	`Layer.cost.total_func_duration`	removed
`Layer.total_macs_forward`	`Layer.cost.total_macs_forward`	removed
`Layer.total_macs_total`	`Layer.cost.total_macs`	removed
`Layer.total_param_memory`	`Layer.cost.total_param_memory`	removed
`Layer.transformed_activation_memory`	`Layer.cost.transformed_activation_memory`	removed
`Layer.transformed_grad`	`Layer.payload.transformed_grad`	removed
`Layer.transformed_grad_dtype`	`Layer.payload.transformed_grad_dtype`	removed
`Layer.transformed_grad_shape`	`Layer.payload.transformed_grad_shape`	removed
`Layer.transformed_out`	`Layer.payload.transformed_out`	removed
`Layer.transformed_out_dtype`	`Layer.payload.transformed_out_dtype`	removed
`Layer.transformed_out_shape`	`Layer.payload.transformed_out_shape`	removed
`Op.address`	`Op.buffer_address`	removed
`Op.activation_transform`	`trace.options.activation_transform` (a capture setting)	removed
`Op.detach_saved_activations`	`trace.options.detach_saved_activations` (a capture setting)	removed
`Op.output_device`	`trace.options.output_device` (a capture setting)	removed
`Op.grad_fn_object_id`	`Op.autograd.grad_fn`	removed
`Layer.address`	`Layer.buffer_address`	removed
`Layer.activation_transform`	`trace.options.activation_transform` (a capture setting)	removed
`Layer.detach_saved_activations`	`trace.options.detach_saved_activations` (a capture setting)	removed
`Layer.output_device`	`trace.options.output_device` (a capture setting)	removed
`Layer.grad_fn_object_id`	`Layer.autograd.grad_fn`	removed
`Op.out_ref`	`Op.out` / `Op.payload.materialize_out()`	removed
`Op.grad_ref`	`Op.grad` / `Op.payload.materialize_grad()`	removed
`Op.grad_fn_cls`	`Op.autograd.grad_fn_class_name`	removed

Module Paths¶

A submodule is a public import path only when its parent lists it in __all__ or it is taught (torchlens.backends.torch, torchlens.backends.registry, torchlens.intervention.errors, torchlens.intervention.hooks, torchlens.intervention.stochastic and the per-framework backend namespaces torchlens.backends.jax, mlx, paddle, tf and tinygrad). Reading an old spelling as an attribute of its parent (tl.intervention.types) raises the typed facade_redirect error; an import statement of an old path fails with ModuleNotFoundError.

Old name	New name	Status
`torchlens.agent.cli`	none public; import its names from `torchlens.agent` (private home `torchlens.agent._cli`)	private
`torchlens.attribution.graph`	none public; import its names from `torchlens.attribution` (private home `torchlens.attribution._graph`)	private
`torchlens.attribution.onebackward`	none public; import its names from `torchlens.attribution` (private home `torchlens.attribution._onebackward`)	private
`torchlens.backends.default_specs`	none public; import its names from `torchlens.backends` (private home `torchlens.backends._default_specs`)	private
`torchlens.bridge.hf_attention`	none public; import its names from `torchlens.bridge` (private home `torchlens.bridge._hf_attention`)	private
`torchlens.errors.episode`	none public; import its names from `torchlens.errors` (private home `torchlens.errors._episode`)	private
`torchlens.errors.runnable`	none public; import its names from `torchlens.errors` (private home `torchlens.errors._runnable`)	private
`torchlens.intervention.audit`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._audit`)	private
`torchlens.intervention.binding`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._binding`)	private
`torchlens.intervention.bundle`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._bundle`)	private
`torchlens.intervention.edge_semantics`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._edge_semantics`)	private
`torchlens.intervention.edge_substitution`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._edge_substitution`)	private
`torchlens.intervention.handles`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._handles`)	private
`torchlens.intervention.helpers`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._helpers`)	private
`torchlens.intervention.injection`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._injection`)	private
`torchlens.intervention.masked_edit`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._masked_edit`)	private
`torchlens.intervention.model_door`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._model_door`)	private
`torchlens.intervention.param_substitution`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._param_substitution`)	private
`torchlens.intervention.predicates`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._predicates`)	private
`torchlens.intervention.regions`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._regions`)	private
`torchlens.intervention.replay`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._replay`)	private
`torchlens.intervention.rerun`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._rerun`)	private
`torchlens.intervention.runtime`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._runtime`)	private
`torchlens.intervention.save`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._save`)	private
`torchlens.intervention.site_keys`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._site_keys`)	private
`torchlens.intervention.spec`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._spec`)	private
`torchlens.intervention.spec_compat`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._spec_compat`)	private
`torchlens.intervention.sweep`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._sweep`)	private
`torchlens.intervention.types`	none public; import its names from `torchlens.intervention` (private home `torchlens.intervention._types`)	private
`torchlens.recording.cleanup`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._cleanup`)	private
`torchlens.recording.exceptions`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._exceptions`)	private
`torchlens.recording.options`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._options`)	private
`torchlens.recording.storage_disk`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._storage_disk`)	private
`torchlens.recording.storage_ram`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._storage_ram`)	private
`torchlens.recording.types`	none public; import its names from `torchlens.recording` (private home `torchlens.recording._types`)	private
`torchlens.validation.backward`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._backward`)	private
`torchlens.validation.consolidated`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._consolidated`)	private
`torchlens.validation.core`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._core`)	private
`torchlens.validation.diagnostics`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._diagnostics`)	private
`torchlens.validation.exemptions`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._exemptions`)	private
`torchlens.validation.invariants`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._invariants`)	private
`torchlens.validation.status`	none public; import its names from `torchlens.validation` (private home `torchlens.validation._status`)	private
`torchlens.viz.batch_summary`	none public; import its names from `torchlens.viz` (private home `torchlens.viz._batch_summary`)	private
`torchlens.viz.node_plots`	none public; import its names from `torchlens.viz` (private home `torchlens.viz._node_plots`)	private
`torchlens.intervention.compose` (the module)	`torchlens.intervention.compose` is the public function or object; the module is `torchlens.intervention._compose`	private
`torchlens.intervention.population` (the module)	`torchlens.intervention.population` is the public function or object; the module is `torchlens.intervention._population`	private
`torchlens.receptive_field.rules` (the module)	`torchlens.receptive_field.rules` is the public function or object; the module is `torchlens.receptive_field._builtin_rules`	private
`torchlens.recording.dry_run` (the module)	`torchlens.recording.dry_run` is the public function or object; the module is `torchlens.recording._dry_run_door`	private
`torchlens.recording.recover` (the module)	`torchlens.recording.recover` is the public function or object; the module is `torchlens.recording._recover`	private
`torchlens.viz.feature_maps` (the module)	`torchlens.viz.feature_maps` is the public function or object; the module is `torchlens.viz._feature_maps`	private

Intervention, Sweep, And Observer Aliases¶

Old name	New name	Status
`replay()`	`do(site, edit)`	removed
`replay_from()`	`do(site, value)`	removed
`rerun()`	`run()`	removed
`Bundle.replay()`	`Bundle.do(site, edit)`	removed
`Bundle.rerun()`	`Bundle.run()`	removed
`torchlens.bundle.AmbiguousLabelError`	`torchlens.errors.AmbiguousLabelError`	removed
`Bundle.measure_members(metric=, order=)`	`Bundle.map(fn)`: live members by name, released candidates read `"unmeasured"`	removed
`intervening()`	`without_op()`	removed
`sweep(param=...)`	`sweep(at=...)`	removed
`record_span()`	`span()`	removed
`get_model_metadata()`	`torchlens.trace(model, x, save="none")`	removed

Gradient Edits (gradient sites)¶

The site, not a keyword or a gradient-only helper, says which tensor an edit applies to: tl.grad(site) addresses the gradient of the site's output. Removed with no alias:

Old name	New name	Status
`tl.zero_grad()`	`tl.zero_ablate()` on `tl.grad(site)`	removed
`tl.scale_grad(f)`	`tl.scale(f)` on `tl.grad(site)`	removed
`tl.clamp_grad(lo, hi)`	`tl.clamp(low=lo, high=hi)` on `tl.grad(site)`	removed
`tl.add_grad_noise(std, seed=)`	`tl.add_noise(std, seed=)` on `tl.grad(site)`	removed
`tl.clip_grad(max_norm)`	`tl.clip_norm(max_norm)` on `tl.grad(site)`	removed
`tl.hook_grad(fn)`	`fn` on `tl.grad(site)`	removed
`direction=` on `Trace.do`, `tl.when`, `tl.observers.tap`, `tl.stats.narrate`, `tl.sweep`	the site: `tl.grad(site)`, or `tl.grad(site) \\| site` for both	removed
`tl.stats.aggregate(target=)`	metric keys are sites (`tl.grad(site)` for a gradient)	removed
`tl.observers.span(direction=)`	`span(signal=)`	removed
`show_bundle_graph(direction=)`	`show_bundle_graph(signal=)`	removed
`FireRecord.direction`, `TapRecord.direction`, `HookContext.direction`	`.signal`	removed
`Edit.kind`, `Edit.direction`	`Edit.signal`; `Rule.signal` (derived)	removed
`Rule.direction`, `Rule(direction=)`	the target decides the pass: `Rule(tl.grad(site), edit)` for the gradient, `Rule(site, edit)` for the activation; `Rule.signal` (derived)	removed
`HookContext.ctx`, `HookContext.run_ctx`	`HookContext.state`, `HookContext.run_state`	removed
`torchlens.intervention.sites` (`sites()`, `SiteCollection`, `SiteSpec`)	pass selectors to `torchlens.experiment.sweep_sites(...)` directly	removed

Experiment Ledger¶

The opt-in semantic record (hypothesis, observations, verdicts in a .tlledger file) was removed on 1 Nov 2026. The Bundle is the record that stays: torchlens.experiment.sweep_sites returns one whose per-candidate effect table (bundle.effects()) and operation chronology (bundle.history) persist with the artifact. Each old spelling raises the typed facade_redirect refusal naming that record; there is no alias.

Old name	New name	Status
`tl.experiment.ledger(path, hypothesis=, metric=)`	your own record beside the Bundle (`bundle.effects()`, `bundle.history`)	removed
`tl.experiment.ExperimentLedger`	none (the Bundle is the record)	removed
`tl.experiment.LedgerEntry`	none	removed
`tl.experiment.active_ledger()`	none	removed
`tl.experiment.read_ledger(path)`	`tl.load(path).effects()` on the saved Bundle	removed
MCP `torchlens_ledger_overview` / `torchlens_ledger_entry` / `torchlens_ledger_evidence`	`torchlens_overview` on the saved Bundle artifact	removed
`ledger(on_record_error=)` and the `ledger_*` refusal and warning codes	none	removed