Safe to_str_tokens, fix memory issues

Author: Johnny Lin

sae_analysis/neuronpedia_runner.py

@@ -1,13 +1,13 @@
 import os
-from typing import Any, Optional, cast
+from typing import Any, Dict, List, Optional, Union, cast
 
 # set TOKENIZERS_PARALLELISM to false to avoid warnings
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 import time
 
 import torch
 from sae_vis.data_fetching_fns import get_feature_data
-from sae_vis.data_storing_fns import FeatureVisParams, to_str_tokens
+from sae_vis.data_storing_fns import FeatureVisParams
 from tqdm import tqdm
 
 import numpy as np
@@ -16,6 +16,8 @@
 
 from matplotlib import colors
 
+OUT_OF_RANGE_TOKEN = "<|outofrange|>"
+
 BG_COLOR_MAP = colors.LinearSegmentedColormap.from_list(
     "bg_color_map", ["white", "darkorange"]
 )
@@ -47,8 +49,9 @@ def __init__(
         n_features_at_a_time: int = 1024,
         buffer_tokens_left: int = 8,
         buffer_tokens_right: int = 8,
-        # start_batch
-        start_batch: int = 0,
+        # start and end batch
+        start_batch_inclusive: int = 0,
+        end_batch_inclusive: Optional[int] = None,
     ):
         self.sae_path = sae_path
         if init_session:
@@ -60,7 +63,8 @@ def __init__(
         self.buffer_tokens_right = buffer_tokens_right
         self.n_batches_to_sample_from = n_batches_to_sample_from
         self.n_prompts_to_select = n_prompts_to_select
-        self.start_batch = start_batch
+        self.start_batch = start_batch_inclusive
+        self.end_batch = end_batch_inclusive
 
         # Deal with file structure
         if not os.path.exists(neuronpedia_parent_folder):
@@ -107,6 +111,32 @@ def get_tokens(
     def round_list(self, to_round: list[float]):
         return list(np.round(to_round, 3))
 
+    def to_str_tokens_safe(
+        self, vocab_dict: Dict[int, str], tokens: Union[int, List[int], torch.Tensor]
+    ):
+        """
+        does to_str_tokens, except handles out of range
+        """
+        vocab_max_index = self.model.cfg.d_vocab - 1
+        # Deal with the int case separately
+        if isinstance(tokens, int):
+            if tokens > vocab_max_index:
+                return OUT_OF_RANGE_TOKEN
+            return vocab_dict[tokens]
+
+        # If the tokens are a (possibly nested) list, turn them into a tensor
+        if isinstance(tokens, list):
+            tokens = torch.tensor(tokens)
+
+        # Get flattened list of tokens
+        str_tokens = [
+            (vocab_dict[t] if t <= vocab_max_index else OUT_OF_RANGE_TOKEN)
+            for t in tokens.flatten().tolist()
+        ]
+
+        # Reshape
+        return np.reshape(str_tokens, tokens.shape).tolist()
+
     def run(self):
         """
         Generate the Neuronpedia outputs.
@@ -137,6 +167,16 @@ def run(self):
         feature_idx = np.array_split(feature_idx, n_subarrays)
         feature_idx = [x.tolist() for x in feature_idx]
 
+        print(f"==== Starting at batch: {self.start_batch}")
+        if self.end_batch is not None:
+            print(f"==== Ending at batch: {self.end_batch}")
+
+        if self.start_batch > len(feature_idx) + 1:
+            print(
+                f"Start batch {self.start_batch} is greater than number of batches + 1 {len(feature_idx)}, exiting"
+            )
+            exit()
+
         # write dead into file so we can create them as dead in Neuronpedia
         skipped_indexes = set(range(self.n_features)) - set(self.target_feature_indexes)
         skipped_indexes_json = json.dumps({"skipped_indexes": list(skipped_indexes)})
@@ -166,16 +206,20 @@ def run(self):
         }
         # pad with blank tokens to the actual vocab size
         for i in range(len(vocab_dict), self.model.cfg.d_vocab):
-            vocab_dict[i] = " "
+            vocab_dict[i] = OUT_OF_RANGE_TOKEN
 
         with torch.no_grad():
             feature_batch_count = 0
             for features_to_process in tqdm(feature_idx):
                 feature_batch_count = feature_batch_count + 1
 
                 if feature_batch_count < self.start_batch:
-                    print(f"Skipping batch: {feature_batch_count}")
+                    # print(f"Skipping batch - it's after start_batch: {feature_batch_count}")
+                    continue
+                if self.end_batch is not None and feature_batch_count > self.end_batch:
+                    # print(f"Skipping batch - it's after end_batch: {feature_batch_count}")
                     continue
+
                 print(f"Doing batch: {feature_batch_count}")
 
                 feature_vis_params = FeatureVisParams(
@@ -255,11 +299,11 @@ def run(self):
                         #     feature.left_tables_data.correlated_features_pearson
                         # )
 
-                    feature_output["neg_str"] = to_str_tokens(
+                    feature_output["neg_str"] = self.to_str_tokens_safe(
                         vocab_dict, feature.middle_plots_data.bottom10_token_ids
                     )
                     feature_output["neg_values"] = bottom10_logits
-                    feature_output["pos_str"] = to_str_tokens(
+                    feature_output["pos_str"] = self.to_str_tokens_safe(
                         vocab_dict, feature.middle_plots_data.top10_token_ids
                     )
                     feature_output["pos_values"] = top10_logits
@@ -320,11 +364,13 @@ def run(self):
                                 negContribs = []
                                 for i in range(len(sd.token_ids)):
                                     strs.append(
-                                        to_str_tokens(vocab_dict, sd.token_ids[i])
+                                        self.to_str_tokens_safe(
+                                            vocab_dict, sd.token_ids[i]
+                                        )
                                     )
                                     posContrib = {}
                                     posTokens = [
-                                        to_str_tokens(vocab_dict, j)
+                                        self.to_str_tokens_safe(vocab_dict, j)
                                         for j in sd.top5_token_ids[i]
                                     ]
                                     if len(posTokens) > 0:
@@ -335,7 +381,7 @@ def run(self):
                                     posContribs.append(posContrib)
                                     negContrib = {}
                                     negTokens = [
-                                        to_str_tokens(vocab_dict, j)
+                                        self.to_str_tokens_safe(vocab_dict, j)
                                         for j in sd.bottom5_token_ids[i]
                                     ]
                                     if len(negTokens) > 0: