Merge remote-tracking branch 'upstream/main'

Author: saireddythfc

.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
-    rev: 25.1.0
+-   repo: https://github.com/psf/black-pre-commit-mirror
+    rev: 25.11.0
     hooks:
     -   id: black
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: 'v0.12.8'
+    rev: 'v0.14.4'
     hooks:
         - id: ruff
           args: [--fix, --exit-non-zero-on-fix]

CHANGELOG.md

@@ -1,6 +1,36 @@
 # CHANGELOG
 
 
+## v0.1.1 (2025-08-27)
+
+### Bug Fixes
+
+- Make simulator workable
+  ([`0ceaa10`](https://github.com/EleutherAI/delphi/commit/0ceaa10bd7b1b7beb411c488ff348c06fb868a67))
+
+* bump vllm dependency to latest
+
+* change simulation scoring default to all at once for local models
+
+* use Role class instead of hardcoding
+
+* delete old oai simulator scorer
+
+* refactor
+
+* [pre-commit.ci] auto fixes from pre-commit.com hooks
+
+for more information, see https://pre-commit.ci
+
+* linter fix
+
+* replace lambda with function
+
+---------
+
+Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
+
+
 ## v0.1.0 (2025-06-17)
 
 ### Features

README.md

@@ -16,15 +16,15 @@ Install this library as a local editable installation. Run the following command
 
 To run the default pipeline from the command line, use the following command:
 
-`python -m delphi meta-llama/Meta-Llama-3-8B EleutherAI/sae-llama-3-8b-32x --n_tokens 10_000_000 --max_latents 100 --hookpoints layers.5 --scorers detection --filter_bos --name llama-3-8B`
+`python -m delphi EleutherAI/pythia-160m EleutherAI/Pythia-160m-SST-k32-32k --n_tokens 10_000_000 --max_latents 100 --hookpoints layers.5.mlp --scorers detection --filter_bos --name llama-3-8B`
 
 This command will:
 1. Cache activations for the first 10 million tokens of the default dataset, `EleutherAI/SmolLM2-135M-10B`.
 2. Generate explanations for the first 100 features of layer 5 using the default explainer model, `hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4`.
 3. Score the explanations using the detection scorer.
 4. Log summary metrics including per-scorer F1 scores and confusion matrices, and produce histograms of the scorer classification accuracies.
 
-The pipeline is highly configurable and can also be called programmatically (see the [end-to-end test](https://github.com/EleutherAI/delphi/blob/main/delphi/tests/e2e.py) for an example).
+The pipeline is highly configurable and can also be called programmatically (see the [end-to-end test](https://github.com/EleutherAI/delphi/blob/main/tests/e2e.py) for an example).
 
 To use experimental features, create a custom pipeline. You can take inspiration from the main pipeline in [delphi.\_\_main\_\_](https://github.com/EleutherAI/delphi/blob/main/delphi/__main__.py).
 

delphi/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.1.0"
+__version__ = "0.1.1"
 
 import logging
 

delphi/__main__.py

@@ -152,7 +152,7 @@ async def process_cache(
     if run_cfg.explainer_provider == "offline":
         llm_client = Offline(
             run_cfg.explainer_model,
-            max_memory=0.9,
+            max_memory=run_cfg.max_memory,
             # Explainer models context length - must be able to accommodate the longest
             # set of examples
             max_model_len=run_cfg.explainer_model_max_len,
@@ -262,13 +262,21 @@ def scorer_postprocess(result, score_dir, scorer_name=None):
         scorer_path.mkdir(parents=True, exist_ok=True)
 
         if scorer_name == "simulation":
-            scorer = OpenAISimulator(llm_client, tokenizer=tokenizer, all_at_once=False)
+            if isinstance(llm_client, Offline):
+                scorer = OpenAISimulator(
+                    llm_client, tokenizer=tokenizer, all_at_once=True
+                )
+            else:
+                scorer = OpenAISimulator(
+                    llm_client, tokenizer=tokenizer, all_at_once=False
+                )
         elif scorer_name == "fuzz":
             scorer = FuzzingScorer(
                 llm_client,
                 n_examples_shown=run_cfg.num_examples_per_scorer_prompt,
                 verbose=run_cfg.verbose,
                 log_prob=run_cfg.log_probs,
+                fuzz_type=run_cfg.fuzz_type,
             )
         elif scorer_name == "detection":
             scorer = DetectionScorer(

delphi/clients/offline.py

@@ -11,6 +11,7 @@
     destroy_distributed_environment,
     destroy_model_parallel,
 )
+from vllm.inputs import TokensPrompt
 
 from delphi import logger
 
@@ -103,6 +104,7 @@ async def process_func(
             prompt = self.tokenizer.apply_chat_template(
                 batch, add_generation_prompt=True, tokenize=True
             )
+            prompt = TokensPrompt(prompt_token_ids=prompt)
             prompts.append(prompt)
             if self.statistics:
                 non_cached_tokens = len(
@@ -121,7 +123,7 @@ async def process_func(
             None,
             partial(
                 self.client.generate,  # type: ignore
-                prompt_token_ids=prompts,
+                prompts,
                 sampling_params=self.sampling_params,
                 use_tqdm=False,
             ),

delphi/config.py

@@ -156,14 +156,18 @@ class RunConfig(Serializable):
     )
     """Scorer methods to score latent explanations. Options are 'fuzz', 'detection',
     'simulation' and 'surprisal_intervention'."""
+    fuzz_type: Literal["default", "active"] = "default"
+    """Type of fuzzing to use for the fuzz scorer. Default uses non-activating
+    examples and highlights n_incorrect tokens. Active uses activating examples
+    and highlights non-activating tokens."""
 
     name: str = ""
     """The name of the run. Results are saved in a directory with this name."""
 
     max_latents: int | None = None
     """Maximum number of features to explain for each sparse model."""
 
-    filter_bos: bool = False
+    filter_bos: bool = True
     """Whether to filter out BOS tokens from the cache."""
 
     log_probs: bool = False
@@ -187,6 +191,11 @@ class RunConfig(Serializable):
     )
     """Number of GPUs to use for explanation and scoring."""
 
+    max_memory: float = field(
+        default=0.7,
+    )
+    """Fraction of GPU memory to allocate to running explainer model."""
+
     seed: int = field(
         default=22,
     )

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/__init__.py

@@ -5,7 +5,9 @@
 from .embedding.example_embedding import ExampleEmbeddingScorer
 from .intervention.surprisal_intervention_scorer import SurprisalInterventionScorer
 from .scorer import Scorer
-from .simulator.oai_simulator import OpenAISimulator
+from .simulator.simulation.oai_simulator import (
+    RefactoredOpenAISimulator as OpenAISimulator,
+)
 from .surprisal.surprisal import SurprisalScorer
 
 __all__ = [