Merge branch 'EleutherAI:main' into gemma_device_fix

Author: kmaherx

.gitignore

@@ -183,3 +183,4 @@ results/
 statistics/
 .embedding_cache/
 wandb/
+uv.lock

.pre-commit-config.yaml

@@ -7,12 +7,12 @@ repos:
     -   id: trailing-whitespace
     -   id: end-of-file-fixer
     -   id: check-added-large-files
--   repo: https://github.com/psf/black
+-   repo: https://github.com/psf/black-pre-commit-mirror
     rev: 25.9.0
     hooks:
     -   id: black
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: 'v0.13.2'
+    rev: 'v0.13.3'
     hooks:
         - id: ruff
           args: [--fix, --exit-non-zero-on-fix]

delphi/pipeline.py

@@ -161,8 +161,9 @@ async def process_item(self, item: Any, semaphore: asyncio.Semaphore) -> Any:
         async with semaphore:
             result = item
             for pipe in self.pipes:
-                if result is not None:
-                    result = await pipe(result)
-                else:
-                    pass
+                if result is None:
+                    return None
+
+                result = await pipe(result)
+
         return result

delphi/scorers/classifier/intruder.py

@@ -1,6 +1,8 @@
 import asyncio
 import re
+from collections import defaultdict
 from dataclasses import dataclass
+from itertools import cycle
 from typing import Literal
 
 from beartype.typing import Sequence
@@ -136,12 +138,11 @@ def _get_quantiled_examples(
         """
         Get the quantiled examples.
         """
-        quantiles = {}
+        examples_grouped_by_quantiles = defaultdict(list)
         for example in examples:
-            if example.quantile not in quantiles:
-                quantiles[example.quantile] = []
-            quantiles[example.quantile].append(example)
-        return quantiles
+            examples_grouped_by_quantiles[example.quantile].append(example)
+
+        return examples_grouped_by_quantiles
 
     def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
         """
@@ -153,38 +154,39 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
         quantiled_intruder_sentences = self._get_quantiled_examples(record.test)
 
         intruder_sentences = record.not_active
-        for i, intruder in enumerate(intruder_sentences):
-            # select each quantile equally
-            quantile_index = i % len(quantiled_intruder_sentences.keys())
 
-            active_examples = quantiled_intruder_sentences[quantile_index]
+        # select each quantile equally by repeatedly cycling through them
+        quantile_iterator = cycle(quantiled_intruder_sentences.items())
+        for (active_quantile, all_active_examples), intruder in zip(
+            quantile_iterator, intruder_sentences
+        ):
             # if there are more examples than the number of examples to show,
             # sample which examples to show
-            examples_to_show = min(self.n_examples_shown - 1, len(active_examples))
-            example_indices = self.rng.sample(
-                range(len(active_examples)), examples_to_show
+            num_active_examples = min(
+                # - 1 because we are going to insert the intruder sentence
+                self.n_examples_shown - 1,
+                len(all_active_examples),
             )
-            active_examples = [active_examples[i] for i in example_indices]
-
-            # convert the examples to strings
+            active_examples = self.rng.sample(all_active_examples, num_active_examples)
 
-            # highlights the active tokens
+            # highlights the active tokens with <<>> markers
             majority_examples = []
-            active_tokens = 0
+            num_active_tokens = 0
             for example in active_examples:
-                text, _ = _prepare_text(
+                text, _str_tokens = _prepare_text(
                     example, n_incorrect=0, threshold=0.3, highlighted=True
                 )
                 majority_examples.append(text)
-                active_tokens += (example.activations > 0).sum().item()
-            active_tokens = int(active_tokens / len(active_examples))
+                num_active_tokens += (example.activations > 0).sum().item()
+
+            avg_active_tokens_per_example = num_active_tokens // len(active_examples)
             if self.type == "default":
                 # if example is contrastive, use the active tokens
                 # otherwise use the non-activating tokens
                 if intruder.activations.max() > 0:
                     n_incorrect = 0
                 else:
-                    n_incorrect = active_tokens
+                    n_incorrect = avg_active_tokens_per_example
                 intruder_sentence, _ = _prepare_text(
                     intruder,
                     n_incorrect=n_incorrect,
@@ -194,22 +196,15 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
             elif self.type == "internal":
                 # randomly select a quantile to be the intruder, make sure it's not
                 # the same as the source quantile
-                intruder_quantile_index = self.rng.randint(
-                    0, len(quantiled_intruder_sentences.keys()) - 1
-                )
-                while intruder_quantile_index == quantile_index:
-                    intruder_quantile_index = self.rng.randint(
-                        0, len(quantiled_intruder_sentences.keys()) - 1
-                    )
-                posible_intruder_sentences = quantiled_intruder_sentences[
-                    intruder_quantile_index
-                ]
-                intruder_index_selected = self.rng.randint(
-                    0, len(posible_intruder_sentences) - 1
-                )
-                intruder = posible_intruder_sentences[intruder_index_selected]
+                all_quantile_examples = list(quantiled_intruder_sentences.values())
+                all_quantile_examples.remove(all_active_examples)
+                possible_intruder_sentences = self.rng.choice(all_quantile_examples)
+
+                intruder = self.rng.choice(possible_intruder_sentences)
                 # here the examples are activating, so we have to convert them
                 # to non-activating examples
+                assert intruder.str_tokens is not None, "intruder has no str_tokens"
+
                 non_activating_intruder = NonActivatingExample(
                     tokens=intruder.tokens,
                     activations=intruder.activations,
@@ -224,23 +219,27 @@ def _prepare_and_batch(self, record: LatentRecord) -> list[IntruderSentence]:
                     highlighted=True,
                 )
                 intruder = non_activating_intruder
+            else:
+                raise ValueError("Invalid intruder scorer type")
 
             # select a random index to insert the intruder sentence
-            intruder_index = self.rng.randint(0, examples_to_show)
-            majority_examples.insert(intruder_index, intruder_sentence)
+            intruder_index = self.rng.randint(0, num_active_examples)
+            examples = (
+                majority_examples[:intruder_index]
+                + [intruder_sentence]
+                + majority_examples[intruder_index:]
+            )
 
-            activations = [example.activations.tolist() for example in active_examples]
-            tokens = [example.str_tokens for example in active_examples]
-            activations.insert(intruder_index, intruder.activations.tolist())
-            tokens.insert(intruder_index, intruder.str_tokens)
+            example_activations = [example.activations.tolist() for example in examples]
+            example_tokens = [example.str_tokens for example in examples]
 
             batches.append(
                 IntruderSentence(
-                    examples=majority_examples,
+                    examples=examples,
                     intruder_index=intruder_index,
-                    chosen_quantile=quantile_index,
-                    activations=activations,
-                    tokens=tokens,
+                    chosen_quantile=active_quantile,
+                    activations=example_activations,
+                    tokens=example_tokens,
                     intruder_distance=intruder.distance,
                 )
             )
@@ -275,7 +274,7 @@ def _build_prompt(
         """
 
         examples = "\n".join(
-            f"Example {i}: {example}" for i, example in enumerate(sample.examples)
+            f"Example {i}:{example}" for i, example in enumerate(sample.examples)
         )
 
         return self.prompt(examples=examples)
@@ -311,21 +310,11 @@ async def _generate(self, sample: IntruderSentence) -> IntruderResult:
         prompt = self._build_prompt(sample)
         try:
             response = await self.client.generate(prompt, **self.generation_kwargs)
+            interpretation, prediction = self._parse(response.text)
         except Exception as e:
-            logger.error(f"Error generating text: {e}")
-            response = None
-
-        if response is None:
+            logger.error(str(e))
             # default result is a error
             return IntruderResult()
-        else:
-
-            try:
-                interpretation, prediction = self._parse(response.text)
-            except Exception as e:
-                logger.error(f"Parsing selections failed: {e}")
-                # default result is a error
-                return IntruderResult()
 
         # check that the only prediction is the intruder
         correct = prediction == sample.intruder_index

delphi/scorers/classifier/sample.py

@@ -1,6 +1,8 @@
 import random
+from collections import deque
 from dataclasses import dataclass
-from typing import NamedTuple
+from itertools import groupby
+from typing import Callable, NamedTuple
 
 import torch
 
@@ -88,73 +90,95 @@ def _prepare_text(
     threshold: float,
     highlighted: bool,
 ) -> tuple[str, list[str]]:
+    assert n_incorrect >= 0, (
+        "n_incorrect must be 0 if highlighting correct example "
+        "or positive if creating false positives. "
+        f"Got {n_incorrect}"
+    )
+
     str_toks = example.str_tokens
     assert str_toks is not None, "str_toks were not set"
-    clean = "".join(str_toks)
+
     # Just return text if there's no highlighting
     if not highlighted:
+        clean = "".join(str_toks)
+
         return clean, str_toks
 
-    threshold = threshold * example.max_activation
+    abs_threshold = threshold * example.max_activation
 
     # Highlight tokens with activations above threshold
-    # if correct example
+    # if this is a correct example
     if n_incorrect == 0:
 
-        def threshold_check(i):
-            return example.activations[i] >= threshold
+        def is_above_activation_threshold(i: int) -> bool:
+            return example.activations[i] >= abs_threshold
 
-        return _highlight(str_toks, threshold_check), str_toks
+        return _highlight(str_toks, is_above_activation_threshold), str_toks
 
     # Highlight n_incorrect tokens with activations
-    # below threshold if incorrect example
-    below_threshold = torch.nonzero(example.activations <= threshold).squeeze()
+    # below threshold if this is an incorrect example
+    tokens_below_threshold = torch.nonzero(
+        example.activations <= abs_threshold
+    ).squeeze()
 
     # Rare case where there are no tokens below threshold
-    if below_threshold.dim() == 0:
-        logger.error("Failed to prepare example.")
+    if tokens_below_threshold.dim() == 0:
+        logger.error(
+            f"Tried to prepare false-positive example with {n_incorrect} tokens "
+            "incorrectly highlighted, but no tokens were below activation threshold."
+        )
         return DEFAULT_MESSAGE, str_toks
 
     random.seed(22)
 
-    n_incorrect = min(n_incorrect, len(below_threshold))
+    num_tokens_to_highlight = min(n_incorrect, tokens_below_threshold.shape[0])
 
     # The activating token is always ctx_len - ctx_len//4
-    # so we always highlight this one, and if  n_incorrect > 1
-    # we highlight n_incorrect-1 random ones
+    # so we always highlight this one, and if num_tokens_to_highlight > 1
+    # we highlight num_tokens_to_highlight - 1 random ones
     token_pos = len(str_toks) - len(str_toks) // 4
-    if token_pos in below_threshold:
+    if token_pos in tokens_below_threshold:
         random_indices = [token_pos]
-        if n_incorrect > 1:
+
+        num_remaining_tokens_to_highlight = num_tokens_to_highlight - 1
+        if num_remaining_tokens_to_highlight > 0:
+            remaining_tokens_below_threshold = tokens_below_threshold.tolist()
+            remaining_tokens_below_threshold.remove(token_pos)
+
             random_indices.extend(
-                random.sample(below_threshold.tolist(), n_incorrect - 1)
+                random.sample(
+                    remaining_tokens_below_threshold,
+                    num_remaining_tokens_to_highlight,
+                )
             )
     else:
-        random_indices = random.sample(below_threshold.tolist(), n_incorrect)
+        random_indices = random.sample(
+            tokens_below_threshold.tolist(), num_tokens_to_highlight
+        )
 
     random_indices = set(random_indices)
 
-    def check(i):
+    def is_false_positive(i):
         return i in random_indices
 
-    return _highlight(str_toks, check), str_toks
+    return _highlight(str_toks, is_false_positive), str_toks
+
 
+def _highlight(tokens: list[str], check: Callable[[int], bool]) -> str:
+    result: deque[str] = deque()
 
-def _highlight(tokens, check):
-    result = []
+    tokens_grouped_by_check_fn = groupby(
+        enumerate(tokens), key=lambda item: check(item[0])
+    )
 
-    i = 0
-    while i < len(tokens):
-        if check(i):
-            result.append(L)
+    for should_highlight, token_group in tokens_grouped_by_check_fn:
+        highlighted_tokens = deque(token for _token_index, token in token_group)
 
-            while i < len(tokens) and check(i):
-                result.append(tokens[i])
-                i += 1
+        if should_highlight:
+            highlighted_tokens.appendleft(L)
+            highlighted_tokens.append(R)
 
-            result.append(R)
-        else:
-            result.append(tokens[i])
-            i += 1
+        result.extend(highlighted_tokens)
 
     return "".join(result)

pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
     "blobfile",
     "bitsandbytes",
     "flask",
-    "vllm",
+    "vllm>=0.10.2",
     "aiofiles",
     "sentence_transformers",
     "anyio>=4.8.0",