Skip to content

Runners

High-level runner functions and their settings for common workflows.

PretrainedSAE pydantic-model

Bases: BaseModelConfig

Fields:

  • device (str)
  • dtype (dtype)
  • pretrained_name_or_path (str)
  • fold_activation_scale (bool)
  • strict_loading (bool)

TrainSAESettings

Bases: BaseSettings

Settings for training a Sparse Autoencoder (SAE).

sae instance-attribute 

sae: BaseSAEConfig | PretrainedSAE

Configuration for the SAE model architecture and parameters, or the path to a pretrained SAE.

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

initializer class-attribute instance-attribute 

initializer: InitializerConfig | None = None

Configuration for model initialization. Should be None for a pretrained SAE.

trainer instance-attribute 

trainer: TrainerConfig

Configuration for training process

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. Mixcoder requires a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

train_sae 

train_sae(settings: TrainSAESettings) -> None

Train a SAE model.

Parameters:

Name Type Description Default
settings TrainSAESettings

Configuration settings for SAE training

 required
Source code in src/lm_saes/runners/train.py 
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
def train_sae(settings: TrainSAESettings) -> None:
    """Train a SAE model.

    Args:
        settings: Configuration settings for SAE training
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.data_parallel_size, settings.model_parallel_size),
            mesh_dim_names=("data", "model"),
        )
        if settings.model_parallel_size > 1 or settings.data_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    logger.info("Loading model and datasets")
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    logger.info("Initializing SAE")

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
            resume=settings.wandb.wandb_resume,
            id=settings.wandb.wandb_run_id,
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    assert settings.initializer is None or not isinstance(settings.initializer, str), (
        "Cannot use an initializer for a pretrained SAE"
    )
    if isinstance(settings.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            settings.sae.pretrained_name_or_path,
            device_mesh=device_mesh,
            fold_activation_scale=settings.sae.fold_activation_scale,
            strict_loading=settings.sae.strict_loading,
            device=settings.sae.device,
            dtype=settings.sae.dtype,
        )
    elif settings.initializer is not None:
        initializer = Initializer(settings.initializer)
        sae = initializer.initialize_sae_from_config(
            settings.sae,
            activation_stream=activations_stream,
            device_mesh=device_mesh,
            wandb_logger=wandb_logger,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(settings.sae, device_mesh=device_mesh)

    if settings.trainer.from_pretrained_path is not None:
        trainer = Trainer.from_checkpoint(
            sae,
            settings.trainer.from_pretrained_path,
        )
        trainer.wandb_logger = wandb_logger
    else:
        trainer = Trainer(settings.trainer)

    logger.info(f"SAE initialized: {type(sae).__name__}")

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting training")

    sae.cfg.save_hyperparameters(settings.trainer.exp_result_path)
    end_of_stream = trainer.fit(
        sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger
    )
    logger.info("Training completed, saving model")
    if end_of_stream:
        trainer.save_checkpoint(
            sae=sae,
            checkpoint_path=settings.trainer.exp_result_path,
        )
    else:
        sae.save_pretrained(
            save_path=settings.trainer.exp_result_path,
        )
        if is_primary_rank(device_mesh) and mongo_client is not None:
            assert settings.sae_name is not None and settings.sae_series is not None, (
                "sae_name and sae_series must be provided when saving to MongoDB"
            )
            mongo_client.create_sae(
                name=settings.sae_name,
                series=settings.sae_series,
                path=str(Path(settings.trainer.exp_result_path).absolute()),
                cfg=sae.cfg,
            )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed")

    logger.info("SAE training completed successfully")

TrainCLTSettings

Bases: BaseSettings

Settings for training a Cross Layer Transcoder (CLT). CLT works with multiple layers and their corresponding hook points.

sae instance-attribute 

sae: CLTConfig | PretrainedSAE

Configuration for the CLT model architecture and parameters, or the path to a pretrained CLT.

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

initializer class-attribute instance-attribute 

initializer: InitializerConfig | None = None

Configuration for model initialization

trainer instance-attribute 

trainer: TrainerConfig

Configuration for training process

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. CLT requires a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

train_clt 

train_clt(settings: TrainCLTSettings) -> None

Train a Cross Layer Transcoder (CLT) model.

Parameters:

Name Type Description Default
settings TrainCLTSettings

Configuration settings for CLT training

 required
Source code in src/lm_saes/runners/train.py 
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
def train_clt(settings: TrainCLTSettings) -> None:
    """Train a Cross Layer Transcoder (CLT) model.

    Args:
        settings: Configuration settings for CLT training
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.data_parallel_size, settings.model_parallel_size),
            mesh_dim_names=("data", "model"),
        )
        if settings.model_parallel_size > 1 or settings.data_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    logger.info("Loading model and datasets")
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
            resume=settings.wandb.wandb_resume,
            id=settings.wandb.wandb_run_id,
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    logger.info("Initializing CLT")
    assert settings.initializer is None or not isinstance(settings.initializer, str), (
        "Cannot use an initializer for a pretrained CLT"
    )
    if isinstance(settings.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            settings.sae.pretrained_name_or_path,
            device_mesh=device_mesh,
            fold_activation_scale=settings.sae.fold_activation_scale,
            strict_loading=settings.sae.strict_loading,
            device=settings.sae.device,
            dtype=settings.sae.dtype,
        )
    elif settings.initializer is not None:
        initializer = Initializer(settings.initializer)
        sae = initializer.initialize_sae_from_config(
            settings.sae,
            activation_stream=activations_stream,
            device_mesh=device_mesh,
            wandb_logger=wandb_logger,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(settings.sae, device_mesh=device_mesh)

    n_params = sum(p.numel() for p in sae.parameters())
    logger.info(f"CLT initialized with {n_params / 1e9:.2f}B parameters")

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting CLT training")
    if settings.trainer.from_pretrained_path is not None:
        trainer = Trainer.from_checkpoint(
            sae,
            settings.trainer.from_pretrained_path,
        )
        trainer.wandb_logger = wandb_logger
    else:
        trainer = Trainer(settings.trainer)
    sae.cfg.save_hyperparameters(settings.trainer.exp_result_path)
    end_of_stream = trainer.fit(
        sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger
    )

    logger.info("Training completed, saving CLT model")
    if end_of_stream:
        trainer.save_checkpoint(
            sae=sae,
            checkpoint_path=settings.trainer.exp_result_path,
        )
    else:
        sae.save_pretrained(
            save_path=settings.trainer.exp_result_path,
        )
        if is_primary_rank(device_mesh) and mongo_client is not None:
            assert settings.sae_name is not None and settings.sae_series is not None, (
                "sae_name and sae_series must be provided when saving to MongoDB"
            )
            mongo_client.create_sae(
                name=settings.sae_name,
                series=settings.sae_series,
                path=str(Path(settings.trainer.exp_result_path).absolute()),
                cfg=sae.cfg,
            )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed")

    logger.info("CLT training completed successfully")

TrainCrossCoderSettings

Bases: BaseSettings

Settings for training a CrossCoder. The main difference to TrainSAESettings is that the activation factory is a list of ActivationFactoryConfig, one for each head.

sae instance-attribute 

sae: CrossCoderConfig | PretrainedSAE

Configuration for the CrossCoder model architecture and parameters, or the path to a pretrained CrossCoder.

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

initializer class-attribute instance-attribute 

initializer: InitializerConfig | None = None

Configuration for model initialization

trainer instance-attribute 

trainer: TrainerConfig

Configuration for training process

activation_factories instance-attribute 

activation_factories: list[ActivationFactoryConfig]

Configuration for generating activations

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. Mixcoder requires a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

train_crosscoder 

train_crosscoder(settings: TrainCrossCoderSettings) -> None

Train a CrossCoder.

Parameters:

Name Type Description Default
settings TrainCrossCoderSettings

Configuration settings for SAE training

 required
Source code in src/lm_saes/runners/train.py 
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
def train_crosscoder(settings: TrainCrossCoderSettings) -> None:
    """Train a CrossCoder.

    Args:
        settings: Configuration settings for SAE training
    """
    # Set up logging
    setup_logging(level="INFO")

    assert isinstance(settings.sae, CrossCoderConfig), "CrossCoderConfig is required for training a CrossCoder"
    assert all(
        len(activation_factory.hook_points) == len(settings.activation_factories[0].hook_points)
        for activation_factory in settings.activation_factories
    ), "Number of hook points of activation factories must be the same"
    assert (
        len(settings.activation_factories) * len(settings.activation_factories[0].hook_points) == settings.sae.n_heads
    ), "Total number of hook points must match the number of heads in the CrossCoder"
    head_parallel_size = len(settings.activation_factories)

    device_mesh = init_device_mesh(
        device_type=settings.device_type,
        mesh_shape=(settings.data_parallel_size, head_parallel_size, settings.model_parallel_size),
        mesh_dim_names=("data", "head", "model"),
    )

    logger.info(
        f"Device mesh initialized with {settings.sae.n_heads} heads, {head_parallel_size} head parallel size, {settings.data_parallel_size} data parallel size, {settings.model_parallel_size} model parallel size"
    )

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    logger.info("Loading model and datasets")
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory_mesh = device_mesh[
        "data", "model"
    ]  # Remove the head dimension, since each activation factory should only be responsible for a subset of the heads.

    logger.info("Setting up activation factory for CrossCoder")
    activation_factory = ActivationFactory(
        settings.activation_factories[device_mesh.get_local_rank("head")], device_mesh=activation_factory_mesh
    )

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
            resume=settings.wandb.wandb_resume,
            id=settings.wandb.wandb_run_id,
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    logger.info("Initializing CrossCoder")
    assert settings.initializer is None or not isinstance(settings.initializer, str), (
        "Cannot use an initializer for a pretrained CrossCoder"
    )
    if isinstance(settings.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            settings.sae.pretrained_name_or_path,
            device_mesh=device_mesh,
            fold_activation_scale=settings.sae.fold_activation_scale,
            strict_loading=settings.sae.strict_loading,
            device=settings.sae.device,
            dtype=settings.sae.dtype,
        )
    elif settings.initializer is not None:
        initializer = Initializer(settings.initializer)
        sae = initializer.initialize_sae_from_config(
            settings.sae,
            activation_stream=activations_stream,
            device_mesh=device_mesh,
            wandb_logger=wandb_logger,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(settings.sae, device_mesh=device_mesh)

    logger.info("CrossCoder initialized")

    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting CrossCoder training")
    if settings.trainer.from_pretrained_path is not None:
        trainer = Trainer.from_checkpoint(
            sae,
            settings.trainer.from_pretrained_path,
        )
        trainer.wandb_logger = wandb_logger
    else:
        trainer = Trainer(settings.trainer)

    sae.cfg.save_hyperparameters(settings.trainer.exp_result_path)
    end_of_stream = trainer.fit(
        sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger
    )

    logger.info("Training completed, saving CrossCoder")
    if end_of_stream:
        trainer.save_checkpoint(
            sae=sae,
            checkpoint_path=settings.trainer.exp_result_path,
        )
    else:
        sae.save_pretrained(
            save_path=settings.trainer.exp_result_path,
        )
        if is_primary_rank(device_mesh) and mongo_client is not None:
            assert settings.sae_name is not None and settings.sae_series is not None, (
                "sae_name and sae_series must be provided when saving to MongoDB"
            )
            mongo_client.create_sae(
                name=settings.sae_name,
                series=settings.sae_series,
                path=str(Path(settings.trainer.exp_result_path).absolute()),
                cfg=settings.sae,
            )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed")

    logger.info("CrossCoder training completed successfully")

TrainLorsaSettings

Bases: BaseSettings

Settings for training a Lorsa (Low-Rank Sparse Autoencoder) model.

sae instance-attribute 

sae: LorsaConfig | PretrainedSAE

Configuration for the Lorsa model architecture and parameters, or the path to a pretrained Lorsa.

sae_name instance-attribute 

sae_name: str

Name of the Lorsa model. Use as identifier for the Lorsa model in the database.

sae_series instance-attribute 

sae_series: str

Series of the Lorsa model. Use as identifier for the Lorsa model in the database.

initializer class-attribute instance-attribute 

initializer: InitializerConfig | None = None

Configuration for model initialization

trainer instance-attribute 

trainer: TrainerConfig

Configuration for training process

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. LORSA may require a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

train_lorsa 

train_lorsa(settings: TrainLorsaSettings) -> None

Train a LORSA (Low-Rank Sparse Autoencoder) model.

Parameters:

Name Type Description Default
settings TrainLorsaSettings

Configuration settings for LORSA training

 required
Source code in src/lm_saes/runners/train.py 
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
def train_lorsa(settings: TrainLorsaSettings) -> None:
    """Train a LORSA (Low-Rank Sparse Autoencoder) model.

    Args:
        settings: Configuration settings for LORSA training
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.data_parallel_size, settings.model_parallel_size),
            mesh_dim_names=("data", "model"),
        )
        if settings.model_parallel_size > 1 or settings.data_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    logger.info("Loading model and datasets")
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    logger.info("Initializing lorsa")

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
            resume=settings.wandb.wandb_resume,
            id=settings.wandb.wandb_run_id,
        )
        if settings.wandb is not None and (device_mesh is None or device_mesh.get_rank() == 0)
        else None
    )

    assert settings.initializer is None or not isinstance(settings.initializer, str), (
        "Cannot use an initializer for a pretrained Lorsa"
    )
    if isinstance(settings.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            settings.sae.pretrained_name_or_path,
            device_mesh=device_mesh,
            fold_activation_scale=settings.sae.fold_activation_scale,
            strict_loading=settings.sae.strict_loading,
            device=settings.sae.device,
            dtype=settings.sae.dtype,
        )
    elif settings.initializer is not None:
        initializer = Initializer(settings.initializer)
        sae = initializer.initialize_sae_from_config(
            settings.sae,
            activation_stream=activations_stream,
            device_mesh=device_mesh,
            wandb_logger=wandb_logger,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(settings.sae, device_mesh=device_mesh)

    n_params = sum(p.numel() for p in sae.parameters())
    logger.info(f"lorsa initialized with {n_params / 1e9:.2f}B parameters")

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting LORSA training")
    if settings.trainer.from_pretrained_path is not None:
        trainer = Trainer.from_checkpoint(
            sae,
            settings.trainer.from_pretrained_path,
        )
        trainer.wandb_logger = wandb_logger
    else:
        trainer = Trainer(settings.trainer)

    sae.cfg.save_hyperparameters(settings.trainer.exp_result_path)
    end_of_stream = trainer.fit(
        sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger
    )

    logger.info("Training completed, saving LORSA model")
    if end_of_stream:
        trainer.save_checkpoint(
            sae=sae,
            checkpoint_path=settings.trainer.exp_result_path,
        )
    else:
        sae.save_pretrained(
            save_path=settings.trainer.exp_result_path,
        )
        if is_primary_rank(device_mesh) and mongo_client is not None:
            assert settings.sae_name is not None and settings.sae_series is not None, (
                "sae_name and sae_series must be provided when saving to MongoDB"
            )
            mongo_client.create_sae(
                name=settings.sae_name,
                series=settings.sae_series,
                path=str(Path(settings.trainer.exp_result_path).absolute()),
                cfg=sae.cfg,
            )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed")

    logger.info("LORSA training completed successfully")

TrainMOLTSettings

Bases: BaseSettings

Settings for training a Mixture of Linear Transforms (MOLT). MOLT is a more efficient alternative to transcoders that sparsely replaces MLP computation in transformers.

sae instance-attribute 

sae: MOLTConfig | PretrainedSAE

Configuration for the MOLT model architecture and parameters

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

initializer class-attribute instance-attribute 

initializer: InitializerConfig | None = None

Configuration for model initialization. Should be None for a pretrained MOLT.

trainer instance-attribute 

trainer: TrainerConfig

Configuration for training process

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. MOLT requires a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

train_molt 

train_molt(settings: TrainMOLTSettings) -> None

Train a Mixture of Linear Transforms (MOLT) model.

Parameters:

Name Type Description Default
settings TrainMOLTSettings

Configuration settings for MOLT training

 required
Source code in src/lm_saes/runners/train.py 
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
def train_molt(settings: TrainMOLTSettings) -> None:
    """Train a Mixture of Linear Transforms (MOLT) model.

    Args:
        settings: Configuration settings for MOLT training
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.model_parallel_size, settings.data_parallel_size),  # TODO: check the order
            mesh_dim_names=("model", "data"),
        )
        if settings.model_parallel_size > 1 or settings.data_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    logger.info("Loading model and datasets")
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
            resume=settings.wandb.wandb_resume,
            id=settings.wandb.wandb_run_id,
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    logger.info("Initializing MOLT")

    assert settings.initializer is None or not isinstance(settings.initializer, str), (
        "Cannot use an initializer for a pretrained MOLT"
    )
    if isinstance(settings.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            settings.sae.pretrained_name_or_path,
            device_mesh=device_mesh,
            fold_activation_scale=settings.sae.fold_activation_scale,
            strict_loading=settings.sae.strict_loading,
            device=settings.sae.device,
            dtype=settings.sae.dtype,
        )
    elif settings.initializer is not None:
        initializer = Initializer(settings.initializer)
        sae = initializer.initialize_sae_from_config(
            settings.sae,
            activation_stream=activations_stream,
            device_mesh=device_mesh,
            wandb_logger=wandb_logger,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(settings.sae, device_mesh=device_mesh)

    logger.info(f"MOLT initialized: {type(sae).__name__}")

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting MOLT training")
    if settings.trainer.from_pretrained_path is not None:
        trainer = Trainer.from_checkpoint(
            sae,
            settings.trainer.from_pretrained_path,
        )
        trainer.wandb_logger = wandb_logger
    else:
        trainer = Trainer(settings.trainer)

    sae.cfg.save_hyperparameters(settings.trainer.exp_result_path)
    end_of_stream = trainer.fit(
        sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger
    )

    logger.info("Training completed, saving MOLT model")
    if end_of_stream:
        trainer.save_checkpoint(
            sae=sae,
            checkpoint_path=settings.trainer.exp_result_path,
        )
    else:
        sae.save_pretrained(
            save_path=settings.trainer.exp_result_path,
        )
        if is_primary_rank(device_mesh) and mongo_client is not None:
            assert settings.sae_name is not None and settings.sae_series is not None, (
                "sae_name and sae_series must be provided when saving to MongoDB"
            )
            mongo_client.create_sae(
                name=settings.sae_name,
                series=settings.sae_series,
                path=str(Path(settings.trainer.exp_result_path).absolute()),
                cfg=sae.cfg,
            )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed")

    logger.info("MOLT training completed successfully")

EvaluateSAESettings

Bases: BaseSettings

Settings for evaluating a Sparse Autoencoder.

sae instance-attribute 

sae: PretrainedSAE

Path to a pretrained SAE model

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

eval instance-attribute 

eval: EvalConfig

Configuration for evaluation

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

fold_activation_scale class-attribute instance-attribute 

fold_activation_scale: bool = False

Whether to fold the activation scale.

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

evaluate_sae 

evaluate_sae(settings: EvaluateSAESettings) -> None

Evaluate a SAE model.

Parameters:

Name Type Description Default
settings EvaluateSAESettings

Configuration settings for SAE evaluation

 required
Source code in src/lm_saes/runners/eval.py 
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
def evaluate_sae(settings: EvaluateSAESettings) -> None:
    """Evaluate a SAE model.

    Args:
        settings: Configuration settings for SAE evaluation
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.model_parallel_size,),
            mesh_dim_names=("model",),
        )
        if settings.model_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    activation_factory = ActivationFactory(settings.activation_factory)

    logger.info("Loading SAE model")

    sae = AbstractSparseAutoEncoder.from_pretrained(
        settings.sae.pretrained_name_or_path,
        device_mesh=device_mesh,
        fold_activation_scale=settings.fold_activation_scale,
        device=settings.sae.device,
        dtype=settings.sae.dtype,
        strict_loading=settings.sae.strict_loading,
    )

    logger.info(f"SAE model loaded: {type(sae).__name__}")

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    logger.info("Processing activations for evaluation")
    activations = activation_factory.process()
    evaluator = Evaluator(settings.eval)
    evaluator.evaluate(sae, activations, wandb_logger)
    logger.info("Evaluation completed")

EvaluateCrossCoderSettings

Bases: BaseSettings

Settings for evaluating a CrossCoder model.

sae instance-attribute 

sae: PretrainedSAE

Path to a pretrained CrossCoder model

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

activation_factories instance-attribute 

activation_factories: list[ActivationFactoryConfig]

Configuration for generating activations

eval instance-attribute 

eval: EvalConfig

Configuration for evaluation

wandb class-attribute instance-attribute 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

evaluate_crosscoder 

evaluate_crosscoder(
    settings: EvaluateCrossCoderSettings,
) -> None

Evaluate a CrossCoder model. The key difference to evaluate_sae is that the activation factories are a list of ActivationFactoryConfig, one for each head; and the evaluating contains a device mesh transformation from head parallelism to model (feature) parallelism.

Parameters:

Name Type Description Default
settings EvaluateCrossCoderSettings

Configuration settings for CrossCoder evaluation

 required
Source code in src/lm_saes/runners/eval.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
@torch.no_grad()
def evaluate_crosscoder(settings: EvaluateCrossCoderSettings) -> None:
    """Evaluate a CrossCoder model. The key difference to evaluate_sae is that the activation factories are a list of ActivationFactoryConfig, one for each head; and the evaluating contains a device mesh transformation from head parallelism to model (feature) parallelism.

    Args:
        settings: Configuration settings for CrossCoder evaluation
    """
    # Set up logging
    setup_logging(level="INFO")

    parallel_size = len(settings.activation_factories)

    logger.info(f"Analyzing CrossCoder with {parallel_size} parallel size")

    device_mesh = init_device_mesh(
        device_type=settings.device_type,
        mesh_shape=(parallel_size,),
        mesh_dim_names=("head",),
    )

    logger.info("Device meshes initialized for CrossCoder analysis")

    logger.info("Setting up activation factory for CrossCoder head")
    activation_factory = ActivationFactory(settings.activation_factories[device_mesh.get_local_rank("head")])

    logger.info("Loading CrossCoder model")
    sae = CrossCoder.from_pretrained(
        settings.sae.pretrained_name_or_path,
        device_mesh=device_mesh,
        device=settings.sae.device,
        dtype=settings.sae.dtype,
        fold_activation_scale=settings.sae.fold_activation_scale,
        strict_loading=settings.sae.strict_loading,
    )

    assert len(settings.activation_factories) * len(settings.activation_factories[0].hook_points) == sae.cfg.n_heads, (
        "Total number of hook points must match the number of heads in the CrossCoder"
    )

    wandb_logger = (
        wandb.init(
            project=settings.wandb.wandb_project,
            config=settings.model_dump(),
            name=settings.wandb.exp_name,
            entity=settings.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
        )
        if settings.wandb is not None and (device_mesh is None or mesh_rank(device_mesh) == 0)
        else None
    )

    if wandb_logger is not None:
        logger.info("WandB logger initialized")

    logger.info("Processing activations for CrossCoder evaluation")
    activations = activation_factory.process()
    evaluator = Evaluator(settings.eval)
    evaluator.evaluate(sae, activations, wandb_logger)

    logger.info("CrossCoder evaluation completed successfully")

AnalyzeSAESettings

Bases: BaseSettings

Settings for analyzing a Sparse Autoencoder.

sae instance-attribute 

sae: PretrainedSAE

Configuration for the SAE model architecture and parameters

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. LORSA may require a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

analyzer instance-attribute 

analyzer: FeatureAnalyzerConfig

Configuration for feature analysis

feature_analysis_name class-attribute instance-attribute 

feature_analysis_name: str = 'default'

Name of the feature analysis.

mongo class-attribute instance-attribute 

mongo: MongoDBConfig | None = None

Configuration for the MongoDB database.

output_dir class-attribute instance-attribute 

output_dir: Optional[Path] = None

Directory to save analysis results. Only used if MongoDB client is not provided.

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

analyze_sae 

analyze_sae(settings: AnalyzeSAESettings) -> None

Analyze a SAE model.

Parameters:

Name Type Description Default
settings AnalyzeSAESettings

Configuration settings for SAE analysis

 required
Source code in src/lm_saes/runners/analyze.py 
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
@torch.no_grad()
def analyze_sae(settings: AnalyzeSAESettings) -> None:
    """Analyze a SAE model.

    Args:
        settings: Configuration settings for SAE analysis
    """
    # Set up logging
    setup_logging(level="INFO")

    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(settings.model_parallel_size,),
            mesh_dim_names=("model",),
        )
        if settings.model_parallel_size > 1
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = None
    if settings.mongo is not None:
        mongo_client = MongoClient(settings.mongo)
        logger.info("MongoDB client initialized")
    else:
        assert settings.output_dir is not None, "Output directory must be provided if MongoDB client is not provided"
        logger.info(f"Analysis results will be saved to {settings.output_dir}")

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    sae = AbstractSparseAutoEncoder.from_pretrained(
        settings.sae.pretrained_name_or_path,
        device_mesh=device_mesh,
        device=settings.sae.device,
        dtype=settings.sae.dtype,
        fold_activation_scale=settings.sae.fold_activation_scale,
        strict_loading=settings.sae.strict_loading,
    )

    logger.info(f"SAE model loaded: {type(sae).__name__}")

    analyzer = FeatureAnalyzer(settings.analyzer)
    logger.info("Feature analyzer initialized")

    logger.info("Processing activations for analysis")

    with torch.amp.autocast(device_type=settings.device_type, dtype=settings.amp_dtype):
        result = analyzer.analyze_chunk(
            activation_factory,
            sae=sae,
            device_mesh=device_mesh,
            activation_factory_process_kwargs={
                "model": model,
                "model_name": settings.model_name,
                "datasets": datasets,
            },
        )

    logger.info("Analysis completed, saving results")
    start_idx = 0 if device_mesh is None else device_mesh.get_local_rank("model") * len(result)
    if mongo_client is not None:
        logger.info("Saving results to MongoDB")
        mongo_client.add_feature_analysis(
            name=settings.feature_analysis_name,
            sae_name=settings.sae_name,
            sae_series=settings.sae_series,
            analysis=result,
            start_idx=start_idx,
        )
        logger.info("Results saved to MongoDB")
    else:
        assert settings.output_dir is not None, "Output directory must be set when MongoDB is not used"
        logger.info(f"Saving results to output directory: {settings.output_dir}")
        pickle_path = save_analysis_to_file(
            output_dir=settings.output_dir,
            analysis_name=settings.feature_analysis_name,
            sae_name=settings.sae_name,
            sae_series=settings.sae_series,
            analysis=result,
            start_idx=start_idx,
            device_mesh=device_mesh,
        )
        logger.info(f"Results saved to: {pickle_path}")

    logger.info("SAE analysis completed successfully")

AnalyzeCrossCoderSettings

Bases: BaseSettings

Settings for analyzing a CrossCoder model.

sae instance-attribute 

sae: PretrainedSAE

Configuration for the CrossCoder model architecture and parameters

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

activation_factories instance-attribute 

activation_factories: list[ActivationFactoryConfig]

Configuration for generating activations

analyzer instance-attribute 

analyzer: FeatureAnalyzerConfig

Configuration for feature analysis

amp_dtype class-attribute instance-attribute 

amp_dtype: dtype = bfloat16

The dtype to use for outputting activations. If None, will not override the dtype.

feature_analysis_name class-attribute instance-attribute 

feature_analysis_name: str = 'default'

Name of the feature analysis.

mongo class-attribute instance-attribute 

mongo: MongoDBConfig | None = None

Configuration for the MongoDB database.

output_dir class-attribute instance-attribute 

output_dir: Optional[Path] = None

Directory to save analysis results. Only used if MongoDB client is not provided.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

analyze_crosscoder 

analyze_crosscoder(
    settings: AnalyzeCrossCoderSettings,
) -> None

Analyze a CrossCoder model. The key difference to analyze_sae is that the activation factories are a list of ActivationFactoryConfig, one for each head; and the analyzing contains a device mesh transformation from head parallelism to model (feature) parallelism.

Parameters:

Name Type Description Default
settings AnalyzeCrossCoderSettings

Configuration settings for CrossCoder analysis

 required
Source code in src/lm_saes/runners/analyze.py 
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
@torch.no_grad()
def analyze_crosscoder(settings: AnalyzeCrossCoderSettings) -> None:
    """Analyze a CrossCoder model. The key difference to analyze_sae is that the activation factories are a list of ActivationFactoryConfig, one for each head; and the analyzing contains a device mesh transformation from head parallelism to model (feature) parallelism.

    Args:
        settings: Configuration settings for CrossCoder analysis
    """
    # Set up logging
    setup_logging(level="INFO")

    parallel_size = len(settings.activation_factories)

    logger.info(f"Analyzing CrossCoder with {parallel_size} parallel size")

    crosscoder_device_mesh = init_device_mesh(
        device_type=settings.device_type,
        mesh_shape=(parallel_size,),
        mesh_dim_names=("head",),
    )

    device_mesh = init_device_mesh(
        device_type=settings.device_type,
        mesh_shape=(parallel_size,),
        mesh_dim_names=("model",),
    )

    logger.info("Device meshes initialized for CrossCoder analysis")

    mongo_client = None
    if settings.mongo is not None:
        mongo_client = MongoClient(settings.mongo)
        logger.info("MongoDB client initialized")
    else:
        assert settings.output_dir is not None, "Output directory must be provided if MongoDB client is not provided"
        logger.info(f"Analysis results will be saved to: {settings.output_dir}")

    logger.info("Setting up activation factory for CrossCoder head")
    activation_factory = ActivationFactory(settings.activation_factories[crosscoder_device_mesh.get_local_rank("head")])

    logger.info("Loading CrossCoder model")
    sae = CrossCoder.from_pretrained(
        settings.sae.pretrained_name_or_path,
        device_mesh=crosscoder_device_mesh,
        device=settings.sae.device,
        dtype=settings.sae.dtype,
        fold_activation_scale=settings.sae.fold_activation_scale,
        strict_loading=settings.sae.strict_loading,
    )

    assert len(settings.activation_factories) * len(settings.activation_factories[0].hook_points) == sae.cfg.n_heads, (
        "Total number of hook points must match the number of heads in the CrossCoder"
    )

    logger.info("Feature analyzer initialized")
    analyzer = FeatureAnalyzer(settings.analyzer)

    logger.info("Processing activations for CrossCoder analysis")

    with torch.amp.autocast(device_type=settings.device_type, dtype=settings.amp_dtype):
        result = analyzer.analyze_chunk(
            activation_factory,
            sae=sae,
            device_mesh=device_mesh,
        )

    logger.info("CrossCoder analysis completed, saving results to MongoDB")
    start_idx = 0 if device_mesh is None else device_mesh.get_local_rank("model") * len(result)
    if mongo_client is not None:
        mongo_client.add_feature_analysis(
            name=settings.feature_analysis_name,
            sae_name=settings.sae_name,
            sae_series=settings.sae_series,
            analysis=result,
            start_idx=start_idx,
        )
    else:
        assert settings.output_dir is not None, "Output directory must be set when MongoDB is not used"
        logger.info(f"Saving results to output directory: {settings.output_dir}")
        pickle_path = save_analysis_to_file(
            output_dir=settings.output_dir,
            analysis_name=settings.feature_analysis_name,
            sae_name=settings.sae_name,
            sae_series=settings.sae_series,
            analysis=result,
            start_idx=start_idx,
            device_mesh=device_mesh,
        )
        logger.info(f"Results saved to: {pickle_path}")

    logger.info("CrossCoder analysis completed successfully")

GenerateActivationsSettings

Bases: BaseSettings

Settings for activation generation.

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for loading the language model. If None, will read from the database.

model_name instance-attribute 

model_name: str

Name of the model to load. Use as identifier for the model in the database.

dataset class-attribute instance-attribute 

dataset: Optional[DatasetConfig] = None

Configuration for loading the dataset. If None, will read from the database.

dataset_name instance-attribute 

dataset_name: str

Name of the dataset. Use as identifier for the dataset in the database.

hook_points instance-attribute 

hook_points: list[str]

List of model hook points to capture activations from

output_dir instance-attribute 

output_dir: str

Directory to save activation files

target class-attribute instance-attribute 

target: ActivationFactoryTarget = ACTIVATIONS_2D

Target type for activation generation

model_batch_size class-attribute instance-attribute 

model_batch_size: int = 1

Batch size for model forward

batch_size instance-attribute 

batch_size: int

Size of the batch for activation generation

buffer_size class-attribute instance-attribute 

buffer_size: Optional[int] = None

Size of the buffer for activation generation

buffer_shuffle class-attribute instance-attribute 

buffer_shuffle: Optional[BufferShuffleConfig] = None

"Manual seed and device of generator for generating randomperm in buffer

total_tokens class-attribute instance-attribute 

total_tokens: Optional[int] = None

Optional total number of tokens to generate

context_size class-attribute instance-attribute 

context_size: int = 128

Context window size for tokenization

n_samples_per_chunk class-attribute instance-attribute 

n_samples_per_chunk: Optional[int] = None

Number of samples per saved chunk

num_workers class-attribute instance-attribute 

num_workers: Optional[int] = None

Number of workers for parallel writing

format class-attribute instance-attribute 

format: Literal['pt', 'safetensors'] = 'safetensors'

Format to save activations in ('pt' or 'safetensors')

n_shards class-attribute instance-attribute 

n_shards: Optional[int] = None

Number of shards to split the dataset into. If None, the dataset is split to the world size. Must be larger than the world size.

start_shard class-attribute instance-attribute 

start_shard: int = 0

The shard to start writing from

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for the MongoDB database. If None, will not use the database.

ignore_token_ids class-attribute instance-attribute 

ignore_token_ids: Optional[list[int]] = None

Tokens to ignore in the activations.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

override_dtype class-attribute instance-attribute 

override_dtype: dtype | None = None

Dtype to override the activations to. If None, will not override the dtype.

model_post_init 

model_post_init(__context: dict) -> None

Validate configuration after initialization.

Source code in src/lm_saes/runners/generate.py 
100
101
102
103
104
def model_post_init(self, __context: dict) -> None:
    """Validate configuration after initialization."""
    if self.mongo is not None:
        assert self.model is not None, "Database not provided. Must manually provide model config."
        assert self.dataset is not None, "Database not provided. Must manually provide dataset config."

generate_activations 

generate_activations(
    settings: GenerateActivationsSettings,
) -> None

Generate and save model activations from a dataset.

Parameters:

Name Type Description Default
settings GenerateActivationsSettings

Configuration settings for activation generation

 required
Source code in src/lm_saes/runners/generate.py 
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
def generate_activations(settings: GenerateActivationsSettings) -> None:
    """Generate and save model activations from a dataset.

    Args:
        settings: Configuration settings for activation generation
    """
    # Set up logging
    setup_logging(level="INFO")

    # Initialize device mesh
    device_mesh = (
        init_device_mesh(
            device_type=settings.device_type,
            mesh_shape=(int(os.environ.get("WORLD_SIZE", 1)), 1),
            mesh_dim_names=("data", "model"),
        )
        if os.environ.get("WORLD_SIZE") is not None
        else None
    )

    logger.info(f"Device mesh initialized: {device_mesh}")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None
    if mongo_client:
        logger.info("MongoDB client initialized")

    # Load configurations
    logger.info("Loading model and dataset configurations")
    model_cfg = load_config(
        config=settings.model, name=settings.model_name, mongo_client=mongo_client, config_type="model"
    )

    dataset_cfg = load_config(
        config=settings.dataset, name=settings.dataset_name, mongo_client=mongo_client, config_type="dataset"
    )

    # Load model and dataset
    logger.info("Loading model and dataset")
    model = load_model(model_cfg)
    dataset, metadata = load_dataset(
        dataset_cfg,
        device_mesh=device_mesh,
        n_shards=settings.n_shards,
        start_shard=settings.start_shard,
    )

    logger.info(f"Model loaded: {settings.model_name}")
    logger.info(f"Dataset loaded: {settings.dataset_name}")

    # Configure activation generation
    logger.info("Configuring activation factory")
    factory_cfg = ActivationFactoryConfig(
        sources=[ActivationFactoryDatasetSource(name=settings.dataset_name)],
        target=settings.target,
        hook_points=settings.hook_points,
        context_size=settings.context_size,
        model_batch_size=settings.model_batch_size,
        batch_size=settings.batch_size,
        buffer_size=settings.buffer_size,
        buffer_shuffle=settings.buffer_shuffle,
        ignore_token_ids=settings.ignore_token_ids,
        override_dtype=settings.override_dtype,
    )

    # Configure activation writer
    logger.info("Configuring activation writer")
    writer_cfg = ActivationWriterConfig(
        hook_points=settings.hook_points,
        total_generating_tokens=settings.total_tokens,
        n_samples_per_chunk=settings.n_samples_per_chunk,
        cache_dir=settings.output_dir,
        format=settings.format,
        num_workers=settings.num_workers,
    )

    # Create factory and writer
    factory = ActivationFactory(factory_cfg)
    writer = ActivationWriter(writer_cfg)

    logger.info("Starting activation generation and writing")
    # Generate and write activations
    activations = factory.process(
        model=model, model_name=settings.model_name, datasets={settings.dataset_name: (dataset, metadata)}
    )
    writer.process(activations, device_mesh=device_mesh, start_shard=settings.start_shard)

    logger.info("Activation generation completed successfully")

AutoInterpSettings

Bases: BaseSettings

Settings for automatic interpretation of SAE features.

sae_name instance-attribute 

sae_name: str

Name of the SAE model to interpret. Use as identifier for the SAE model in the database.

sae_series instance-attribute 

sae_series: str

Series of the SAE model to interpret. Use as identifier for the SAE model in the database.

model instance-attribute 

model: LanguageModelConfig

Configuration for the language model used to generate activations.

model_name instance-attribute 

model_name: str

Name of the model to load.

auto_interp instance-attribute 

auto_interp: AutoInterpConfig

Configuration for the auto-interpretation process.

mongo instance-attribute 

mongo: MongoDBConfig

Configuration for the MongoDB database.

features class-attribute instance-attribute 

features: Optional[list[int]] = None

List of specific feature indices to interpret. If None, will interpret all features.

analysis_name class-attribute instance-attribute 

analysis_name: str = 'default'

Name of the analysis to use for interpretation.

max_workers class-attribute instance-attribute 

max_workers: int = 10

Maximum number of workers to use for interpretation.

auto_interp 

auto_interp(settings: AutoInterpSettings)

Synchronous wrapper for interpret_feature.

Parameters:

Name Type Description Default
settings AutoInterpSettings

Configuration for feature interpretation

 required
Source code in src/lm_saes/runners/autointerp.py 
125
126
127
128
129
130
131
def auto_interp(settings: AutoInterpSettings):
    """Synchronous wrapper for interpret_feature.

    Args:
        settings: Configuration for feature interpretation
    """
    asyncio.run(interpret_feature(settings))

SweepSAESettings

Bases: BaseSettings

Settings for sweeping a Sparse Autoencoder (SAE).

items instance-attribute 

items: list[SweepingItem]

List of sweeping items

activation_factory instance-attribute 

activation_factory: ActivationFactoryConfig

Configuration for generating activations

eval class-attribute instance-attribute 

eval: bool = False

Whether to run in evaluation mode

data_parallel_size class-attribute instance-attribute 

data_parallel_size: int = 1

Size of data parallel mesh

model_parallel_size class-attribute instance-attribute 

model_parallel_size: int = 1

Size of model parallel (tensor parallel) mesh

mongo class-attribute instance-attribute 

mongo: Optional[MongoDBConfig] = None

Configuration for MongoDB

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model. Required if using dataset sources.

model_name class-attribute instance-attribute 

model_name: Optional[str] = None

Name of the tokenizer to load. Mixcoder requires a tokenizer to get the modality indices.

datasets class-attribute instance-attribute 

datasets: Optional[dict[str, Optional[DatasetConfig]]] = (
    None
)

Name to dataset config mapping. Required if using dataset sources.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

SweepingItem pydantic-model

Bases: BaseModel

A single item in a sweeping configuration.

Fields:

sae pydantic-field 

sae: BaseSAEConfig | PretrainedSAE

Configuration for the SAE model architecture and parameters, or the path to a pretrained SAE.

sae_name pydantic-field 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

sae_series pydantic-field 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

initializer pydantic-field 

initializer: InitializerConfig | None = None

Configuration for model initialization. Should be None for a pretrained SAE.

trainer pydantic-field 

trainer: TrainerConfig

Configuration for training process

wandb pydantic-field 

wandb: Optional[WandbConfig] = None

Configuration for Weights & Biases logging

sweep_sae 

sweep_sae(settings: SweepSAESettings) -> None

Sweep experiments for training SAE models.

Parameters:

Name Type Description Default
settings SweepSAESettings

Configuration settings for SAE sweeping

 required
Source code in src/lm_saes/runners/train.py 
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
def sweep_sae(settings: SweepSAESettings) -> None:
    """Sweep experiments for training SAE models.

    Args:
        settings: Configuration settings for SAE sweeping
    """
    # Set up logging
    setup_logging(level="INFO")

    n_sweeps = len(settings.items)

    device_mesh = init_device_mesh(
        device_type=settings.device_type,
        mesh_shape=(n_sweeps, settings.data_parallel_size, settings.model_parallel_size),
        mesh_dim_names=("sweep", "data", "model"),
    )

    logger.info(f"Device mesh initialized for sweep with {n_sweeps} configurations")

    mongo_client = MongoClient(settings.mongo) if settings.mongo is not None else None

    logger.info("Loading configurations on rank 0")
    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=False,
    )

    dataset_cfgs = (
        {
            dataset_name: load_config(
                config=dataset_cfg,
                name=dataset_name,
                mongo_client=mongo_client,
                config_type="dataset",
            )
            for dataset_name, dataset_cfg in settings.datasets.items()
        }
        if settings.datasets is not None
        else None
    )

    # Load model and datasets
    model = load_model(model_cfg) if model_cfg is not None else None
    datasets = (
        {
            dataset_name: load_dataset(dataset_cfg, device_mesh=device_mesh)
            for dataset_name, dataset_cfg in dataset_cfgs.items()
        }
        if dataset_cfgs is not None
        else None
    )

    activation_factory = ActivationFactory(settings.activation_factory, device_mesh=device_mesh)

    logger.info("Processing activations stream")
    activations_stream = activation_factory.process(
        model=model,
        model_name=settings.model_name,
        datasets=datasets,
    )

    sae_device_mesh = device_mesh["data", "model"]
    logger.info(f"Created 2D sub-mesh for SAE: {sae_device_mesh}")

    item = settings.items[device_mesh.get_local_rank("sweep")]
    logger.info(f"Processing sweep item: {item.sae_name}/{item.sae_series}")

    def convert_activations_to_2d_mesh(stream_3d, mesh_2d):
        from torch.distributed.tensor import DTensor

        for batch in stream_3d:
            converted_batch = {}
            for key, value in batch.items():
                if isinstance(value, torch.Tensor):
                    assert isinstance(value, DTensor), "value must be a DTensor"
                    local_tensor = value.to_local()
                    from lm_saes.utils.distributed import DimMap

                    converted_value = DTensor.from_local(
                        local_tensor,
                        device_mesh=mesh_2d,
                        placements=DimMap({"data": 0}).placements(mesh_2d),
                    )
                    converted_batch[key] = converted_value
                else:
                    converted_batch[key] = value
            yield converted_batch

    activations_stream = convert_activations_to_2d_mesh(activations_stream, sae_device_mesh)

    logger.info("Initializing SAE on 2D sub-mesh")

    assert item.initializer is None or not isinstance(item.initializer, str), (
        "Cannot use an initializer for a pretrained SAE"
    )
    if isinstance(item.sae, PretrainedSAE):
        sae = AbstractSparseAutoEncoder.from_pretrained(
            item.sae.pretrained_name_or_path,
            device_mesh=sae_device_mesh,
            fold_activation_scale=item.sae.fold_activation_scale,
            strict_loading=item.sae.strict_loading,
            device=item.sae.device,
            dtype=item.sae.dtype,
        )
    elif item.initializer is not None:
        initializer = Initializer(item.initializer)
        sae = initializer.initialize_sae_from_config(
            item.sae,
            activation_stream=activations_stream,
            device_mesh=sae_device_mesh,
            model=model,
        )
    else:
        sae = AbstractSparseAutoEncoder.from_config(item.sae, device_mesh=sae_device_mesh)

    wandb_logger = (
        wandb.init(
            project=item.wandb.wandb_project,
            config=item.model_dump(),
            name=item.wandb.exp_name,
            entity=item.wandb.wandb_entity,
            settings=wandb.Settings(x_disable_stats=True),
            mode=os.getenv("WANDB_MODE", "online"),  # type: ignore
        )
        if item.wandb is not None and is_primary_rank(device_mesh)
        else None
    )
    # TODO: implement eval_fn
    eval_fn = (lambda x: None) if settings.eval else None

    logger.info("Starting training for sweep item")
    trainer = Trainer(item.trainer)
    sae.cfg.save_hyperparameters(item.trainer.exp_result_path)
    trainer.fit(sae=sae, activation_stream=activations_stream, eval_fn=eval_fn, wandb_logger=wandb_logger)

    logger.info("Training completed, saving sweep item")
    sae.save_pretrained(
        save_path=item.trainer.exp_result_path,
    )
    if is_primary_rank(device_mesh) and mongo_client is not None:
        assert item.sae_name is not None and item.sae_series is not None, (
            "sae_name and sae_series must be provided when saving to MongoDB"
        )
        mongo_client.create_sae(
            name=item.sae_name,
            series=item.sae_series,
            path=str(Path(item.trainer.exp_result_path).absolute()),
            cfg=sae.cfg,
        )

    if wandb_logger is not None:
        wandb_logger.finish()
        logger.info("WandB session closed for sweep item")

    logger.info(f"Sweep item completed: {item.sae_name}/{item.sae_series}")

DirectLogitAttributeSettings

Bases: BaseSettings

Settings for analyzing a CrossCoder model.

sae instance-attribute 

sae: PretrainedSAE

Configuration for the SAE model architecture and parameters

sae_name instance-attribute 

sae_name: str

Name of the SAE model. Use as identifier for the SAE model in the database.

layer_idx class-attribute instance-attribute 

layer_idx: Optional[int | None] = None

The index of layer to DLA.

sae_series instance-attribute 

sae_series: str

Series of the SAE model. Use as identifier for the SAE model in the database.

model class-attribute instance-attribute 

model: Optional[LanguageModelConfig] = None

Configuration for the language model.

model_name instance-attribute 

model_name: str

Name of the language model.

direct_logit_attributor instance-attribute 

direct_logit_attributor: DirectLogitAttributorConfig

Configuration for the direct logit attributor.

mongo class-attribute instance-attribute 

mongo: MongoDBConfig | None = None

Configuration for the MongoDB database.

analysis_file class-attribute instance-attribute 

analysis_file: Path | None = None

The analysis results file to be updated. Only used if MongoDB client is not provided.

device_type class-attribute instance-attribute 

device_type: str = 'cuda'

Device type to use for distributed training ('cuda' or 'cpu')

direct_logit_attribute 

direct_logit_attribute(
    settings: DirectLogitAttributeSettings,
) -> None

Direct logit attribute a SAE model.

Parameters:

Name Type Description Default
settings DirectLogitAttributeSettings

Configuration settings for DirectLogitAttributor

 required
Source code in src/lm_saes/runners/analyze.py 
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
@torch.no_grad()
def direct_logit_attribute(settings: DirectLogitAttributeSettings) -> None:
    """Direct logit attribute a SAE model.

    Args:
        settings: Configuration settings for DirectLogitAttributor
    """
    # Set up logging
    setup_logging(level="INFO")

    # device_mesh = (
    #     init_device_mesh(
    #         device_type=settings.device_type,
    #         mesh_shape=(settings.head_parallel_size, settings.data_parallel_size, settings.model_parallel_size),
    #         mesh_dim_names=("head", "data", "model"),
    #     )
    #     if settings.head_parallel_size > 1 or settings.data_parallel_size > 1 or settings.model_parallel_size > 1
    #     else None
    # )

    mongo_client = None
    if settings.mongo is not None:
        mongo_client = MongoClient(settings.mongo)
        logger.info("MongoDB client initialized")
    else:
        assert settings.analysis_file is not None, (
            "Analysis directory must be provided if MongoDB client is not provided"
        )
        # the analysis directory should contain the analysis results to be updated
        logger.info(f"Analysis results to be updated: {settings.analysis_file}")

    logger.info("Loading SAE model")
    sae = AbstractSparseAutoEncoder.from_pretrained(
        settings.sae.pretrained_name_or_path,
        device=settings.sae.device,
        dtype=settings.sae.dtype,
        fold_activation_scale=settings.sae.fold_activation_scale,
        strict_loading=settings.sae.strict_loading,
    )

    # Load configurations
    model_cfg = load_config(
        config=settings.model,
        name=settings.model_name,
        mongo_client=mongo_client,
        config_type="model",
        required=True,
    )
    model_cfg.device = settings.device_type
    model_cfg.dtype = sae.cfg.dtype

    model = load_model(model_cfg)
    assert isinstance(model, TransformerLensLanguageModel), (
        "DirectLogitAttributor only supports TransformerLensLanguageModel as the model backend"
    )

    logger.info("Direct logit attribution")
    direct_logit_attributor = DirectLogitAttributor(settings.direct_logit_attributor)
    results = direct_logit_attributor.direct_logit_attribute(sae, model, settings.layer_idx)

    # if is_master():
    if mongo_client is not None:
        logger.info("Direct logit attribution completed, saving results to MongoDB")
        mongo_client.update_features(
            sae_name=settings.sae_name,
            sae_series=settings.sae_series,
            update_data=[{"logits": result} for result in results],
            start_idx=0,
        )
    else:
        assert settings.analysis_file is not None, "analysis_file must be set when MongoDB is not used"
        logger.info(f"Loading analysis results from: {settings.analysis_file}")

        # Load existing analysis results
        with open(settings.analysis_file, "rb") as f:
            analysis_data = pickle.load(f)

        # Update each feature with logits
        assert len(analysis_data) == len(results), (
            f"Number of features in analysis file ({len(analysis_data)}) does not match "
            f"number of results from direct logit attribution ({len(results)})"
        )

        for i, result in enumerate(results):
            assert analysis_data[i]["feature_idx"] == i, "Feature index mismatch"
            analysis_data[i]["logits"] = result

        # Save updated analysis back to file
        logger.info(f"Saving updated analysis results to: {settings.analysis_file}")
        with open(settings.analysis_file, "wb") as f:
            pickle.dump(analysis_data, f)

        logger.info(f"Updated {len(results)} features with logit attributions")

    logger.info("Direct logit attribution completed successfully")

CheckActivationConsistencySettings

Bases: BaseSettings

Settings for checking activation consistency. It will check if the activations are consistent across different hook points by comparing their token ids.

paths instance-attribute 

paths: dict[str, Path]

Paths to the activations to check.

device class-attribute instance-attribute 

device: str = 'cuda'

Device to use for checking activation consistency

num_workers class-attribute instance-attribute 

num_workers: int = 0

Number of workers to use for checking activation consistency

prefetch_factor class-attribute instance-attribute 

prefetch_factor: int | None = None

Number of samples loaded in advance by each worker

check_activation_consistency 

check_activation_consistency(
    settings: CheckActivationConsistencySettings,
) -> None

Check activation consistency.

Source code in src/lm_saes/runners/generate.py 
212
213
214
215
216
217
218
219
220
221
222
223
224
def check_activation_consistency(settings: CheckActivationConsistencySettings) -> None:
    """Check activation consistency."""

    loader = CachedActivationLoader(
        cache_dirs=settings.paths,
        device=settings.device,
        num_workers=settings.num_workers,
        prefetch_factor=settings.prefetch_factor,
    )

    activations = loader.process()
    for activation in activations:
        pass