[Minor] Allow intervene on generated tokens

Author: PinetreePantry

pyvene/models/intervenable_base.py

@@ -99,6 +99,7 @@ def __init__(self, config, model, backend, **kwargs):
         # token, or on a combination of both.
         self._is_generation = False
         self._intervene_on_prompt = None
+        self._intervene_on_generated = False
         self._key_getter_call_counter = {}
         self._key_setter_call_counter = {}
         self._intervention_pointers = {}
@@ -724,6 +725,7 @@ def _cleanup_states(self, skip_activation_gc=False):
         Clean up all old in memo states of interventions
         """
         self._is_generation = False
+        self._intervene_on_generated = False
         if not skip_activation_gc:
             self.activations.clear()
             self.hot_activations.clear()
@@ -767,7 +769,7 @@ def _intervention_getter(
         for key_i, key in enumerate(keys):
             intervention = self.interventions[key]
             (module_hook, hook_type) = self.intervention_hooks[key]
-            if self._is_generation:
+            if self._is_generation or self._intervene_on_generated:
                 raise NotImplementedError("Generation is not implemented for ndif backend")
 
             if hook_type == CONST_INPUT_HOOK:
@@ -818,7 +820,7 @@ def _intervention_setter(
                     0 for _ in range(len(unit_locations_base[0]))
                 ]  # batch_size
 
-            if self._is_generation:
+            if self._is_generation or self._intervene_on_generated:
                 raise NotImplementedError("Generation is not implemented for ndif backend")
 
             if hook_type == CONST_INPUT_HOOK:
@@ -916,7 +918,7 @@ def _sync_forward_with_parallel_intervention(
         **kwargs,
     ):
         # torch.autograd.set_detect_anomaly(True)
-        all_set_handlers = HandlerList([])
+        all_handlers_to_remove = HandlerList([])
         unit_locations_sources = unit_locations["sources->base"][0]
         unit_locations_base = unit_locations["sources->base"][1]
 
@@ -1141,6 +1143,7 @@ def _cleanup_states(self, skip_activation_gc=False):
         Clean up all old in memo states of interventions
         """
         self._is_generation = False
+        self._intervene_on_generated = False
         self._remove_forward_hooks()
         self._reset_hook_count()
         if not skip_activation_gc:
@@ -1392,7 +1395,7 @@ def _intervention_getter(
             module_hook = self.intervention_hooks[key]
 
             def hook_callback(model, args, kwargs, output=None):
-                if self._is_generation:
+                if self._is_generation and not self._intervene_on_generated:
                     pass
                     # for getter, there is no restriction.
                     # is_prompt = self._key_getter_call_counter[key] == 0
@@ -1685,7 +1688,7 @@ def _wait_for_forward_with_parallel_intervention(
         subspaces: Optional[List] = None,
     ):
         # torch.autograd.set_detect_anomaly(True)
-        all_set_handlers = HandlerList([])
+        all_handlers_to_remove = HandlerList([])
         unit_locations_sources = unit_locations["sources->base"][0]
         unit_locations_base = unit_locations["sources->base"][1]
 
@@ -1708,7 +1711,11 @@ def _wait_for_forward_with_parallel_intervention(
                     )
                     group_get_handlers.extend(get_handlers)
                 _ = self.model(**sources[group_id])
-                group_get_handlers.remove()
+                if not self._is_generation or not self._intervene_on_generated:
+                    group_get_handlers.remove()
+                else:
+                    # For generation, we use one getter for all setters
+                    all_handlers_to_remove.extend(group_get_handlers)
         else:
             # simply patch in the ones passed in
             self.activations = activations_sources
@@ -1740,8 +1747,8 @@ def _wait_for_forward_with_parallel_intervention(
                         else None,
                     )
                     # for setters, we don't remove them.
-                    all_set_handlers.extend(set_handlers)
-        return all_set_handlers
+                    all_handlers_to_remove.extend(set_handlers)
+        return all_handlers_to_remove
 
     def _wait_for_forward_with_serial_intervention(
         self,
@@ -1750,7 +1757,8 @@ def _wait_for_forward_with_serial_intervention(
         activations_sources: Optional[Dict] = None,
         subspaces: Optional[List] = None,
     ):
-        all_set_handlers = HandlerList([])
+        set_handlers_to_remove = HandlerList([])
+        get_handlers_to_remove = HandlerList([])
         for group_id, keys in self._intervention_group.items():
             if sources[group_id] is None:
                 continue  # smart jump for advance usage only
@@ -1782,11 +1790,15 @@ def _wait_for_forward_with_serial_intervention(
             if activations_sources is None:
                 # this is when previous setter and THEN the getter get called
                 _ = self.model(**sources[group_id])
-                get_handlers.remove()
+                if not self._is_generation or not self._intervene_on_generated:
+                    get_handlers.remove()
+                else:
+                    # For generation, we use one getter for all setters
+                    get_handlers_to_remove.extend(get_handlers)
                 # remove existing setters after getting the curr intervened reprs
-                if len(all_set_handlers) > 0:
-                    all_set_handlers.remove()
-                    all_set_handlers = HandlerList([])
+                if len(set_handlers_to_remove) > 0:
+                    set_handlers_to_remove.remove()
+                    set_handlers_to_remove = HandlerList([])
 
             for key in keys:
                 # skip in case smart jump
@@ -1807,8 +1819,9 @@ def _wait_for_forward_with_serial_intervention(
                         else None,
                     )
                     # for setters, we don't remove them.
-                    all_set_handlers.extend(set_handlers)
-        return all_set_handlers
+                    set_handlers_to_remove.extend(set_handlers)
+        all_handlers_to_remove = set_handlers_to_remove + get_handlers_to_remove
+        return all_handlers_to_remove
 
     def forward(
         self,
@@ -1923,7 +1936,7 @@ def forward(
         try:
             # intervene
             if self.mode == "parallel":
-                set_handlers_to_remove = (
+                all_handlers_to_remove = (
                     self._wait_for_forward_with_parallel_intervention(
                         sources,
                         unit_locations,
@@ -1932,7 +1945,7 @@ def forward(
                     )
                 )
             elif self.mode == "serial":
-                set_handlers_to_remove = (
+                all_handlers_to_remove = (
                     self._wait_for_forward_with_serial_intervention(
                         sources,
                         unit_locations,
@@ -1950,7 +1963,7 @@ def forward(
 
             counterfactual_outputs = self.model(**base, **model_kwargs)
 
-            set_handlers_to_remove.remove()
+            all_handlers_to_remove.remove()
 
             self._output_validation()
 
@@ -2000,6 +2013,7 @@ def generate(
         intervene_on_prompt: bool = False,
         subspaces: Optional[List] = None,
         output_original_output: Optional[bool] = False,
+        intervene_on_generated: bool = False,
         **kwargs,
     ):
         """
@@ -2035,6 +2049,7 @@ def generate(
         self._cleanup_states()
 
         self._intervene_on_prompt = intervene_on_prompt
+        self._intervene_on_generated = intervene_on_generated
         self._is_generation = True
 
         if not intervene_on_prompt and unit_locations is None:
@@ -2061,11 +2076,11 @@ def generate(
             # returning un-intervened output
             base_outputs = self.model.generate(**base, **kwargs)
 
-        set_handlers_to_remove = None
+        all_handlers_to_remove = None
         try:
             # intervene
             if self.mode == "parallel":
-                set_handlers_to_remove = (
+                all_handlers_to_remove = (
                     self._wait_for_forward_with_parallel_intervention(
                         sources,
                         unit_locations,
@@ -2074,7 +2089,7 @@ def generate(
                     )
                 )
             elif self.mode == "serial":
-                set_handlers_to_remove = (
+                all_handlers_to_remove = (
                     self._wait_for_forward_with_serial_intervention(
                         sources,
                         unit_locations,
@@ -2099,9 +2114,10 @@ def generate(
         except Exception as e:
             raise e
         finally:
-            if set_handlers_to_remove is not None:
-                set_handlers_to_remove.remove()
+            if all_handlers_to_remove is not None:
+                all_handlers_to_remove.remove()
             self._is_generation = False
+            self._intervene_on_generated = False
             self._cleanup_states(
                 skip_activation_gc = \
                     (sources is None and activations_sources is not None) or \