Coverage for transformer_lens/HookedEncoderDecoder.py: 72%
284 statements
« prev ^ index » next coverage.py v7.6.1, created at 2025-07-09 19:34 +0000
« prev ^ index » next coverage.py v7.6.1, created at 2025-07-09 19:34 +0000
1"""Hooked EncoderDecoder
3Contains a T5 style model. This is separate from :class:`transformer_lens.HookedTransformer`
4because it has a significantly different architecture to e.g. GPT style transformers.
5"""
7from __future__ import annotations
9import logging
10import os
11from itertools import chain
12from pathlib import Path
13from typing import (
14 Any,
15 Dict,
16 List,
17 Optional,
18 Tuple,
19 Type,
20 TypeVar,
21 Union,
22 cast,
23 overload,
24)
26import torch
27import tqdm
28from einops import repeat
29from jaxtyping import Float, Int
30from torch import nn
31from transformers import AutoTokenizer, PreTrainedTokenizerBase
32from typing_extensions import Literal
34import transformer_lens.loading_from_pretrained as loading
35from transformer_lens.ActivationCache import ActivationCache
36from transformer_lens.components import MLP, Embed, GatedMLP, RMSNorm, T5Block, Unembed
37from transformer_lens.FactoredMatrix import FactoredMatrix
38from transformer_lens.hook_points import HookedRootModule, HookPoint
39from transformer_lens.HookedTransformerConfig import HookedTransformerConfig
40from transformer_lens.utilities import devices
41from transformer_lens.utils import sample_logits
43T = TypeVar("T", bound="HookedEncoderDecoder")
46class HookedEncoderDecoder(HookedRootModule):
47 """
48 This class implements a T5 encoder-decoder using the components in ./components.py, with HookPoints on every interesting activation. It inherits from HookedRootModule.
50 Limitations:
51 - Also note that model does not include dropouts, which may lead to inconsistent results from training or fine-tuning.
53 Like HookedTransformer, it can have a pretrained Transformer's weights loaded via `.from_pretrained`. There are a few features you might know from HookedTransformer which are not yet supported:
54 - There is no preprocessing (e.g. LayerNorm folding) when loading a pretrained model
55 - The model only accepts tokens as inputs, and not strings, or lists of strings
56 """
58 tokenizer: Optional[PreTrainedTokenizerBase]
60 def __init__(
61 self,
62 cfg: Union[HookedTransformerConfig, Dict],
63 tokenizer: Optional[PreTrainedTokenizerBase] = None,
64 move_to_device: bool = True,
65 **kwargs: Any,
66 ):
67 super().__init__()
68 if isinstance(cfg, Dict): 68 ↛ 69line 68 didn't jump to line 69 because the condition on line 68 was never true
69 cfg = HookedTransformerConfig(**cfg)
70 elif isinstance(cfg, str): 70 ↛ 71line 70 didn't jump to line 71 because the condition on line 70 was never true
71 raise ValueError(
72 "Please pass in a config dictionary or HookedTransformerConfig object. If you want to load a pretrained model, use HookedEncoderDecoder.from_pretrained() instead."
73 )
74 self.cfg: HookedTransformerConfig = cfg
76 if self.cfg.n_devices != 1: 76 ↛ 77line 76 didn't jump to line 77 because the condition on line 76 was never true
77 raise ValueError("Multiple devices not supported for HookedEncoderDecoder")
78 if tokenizer is not None: 78 ↛ 79line 78 didn't jump to line 79 because the condition on line 78 was never true
79 self.tokenizer = tokenizer
80 elif self.cfg.tokenizer_name is not None: 80 ↛ 87line 80 didn't jump to line 87 because the condition on line 80 was always true
81 huggingface_token = os.environ.get("HF_TOKEN", "")
82 self.tokenizer = AutoTokenizer.from_pretrained(
83 self.cfg.tokenizer_name,
84 token=huggingface_token if len(huggingface_token) > 0 else None,
85 )
86 else:
87 self.tokenizer = None
89 if self.cfg.d_vocab == -1: 89 ↛ 91line 89 didn't jump to line 91 because the condition on line 89 was never true
90 # If we have a tokenizer, vocab size can be inferred from it.
91 if self.tokenizer is None:
92 raise ValueError("Must provide a tokenizer if d_vocab is not provided")
94 self.cfg.d_vocab = len(self.tokenizer)
95 if self.cfg.d_vocab_out == -1: 95 ↛ 96line 95 didn't jump to line 96 because the condition on line 95 was never true
96 self.cfg.d_vocab_out = self.cfg.d_vocab
98 self.embed = Embed(self.cfg)
99 self.encoder = nn.ModuleList(
100 [
101 T5Block(self.cfg, num_layer, is_decoder=False)
102 for num_layer in range(self.cfg.n_layers)
103 ]
104 )
105 self.encoder_final_ln = RMSNorm(self.cfg)
106 self.decoder = nn.ModuleList(
107 [
108 T5Block(self.cfg, num_layer, is_decoder=True)
109 for num_layer in range(self.cfg.n_layers)
110 ]
111 )
112 self.decoder_final_ln = RMSNorm(self.cfg)
113 # self.lm_head = nn.Linear(self.cfg.d_model, self.cfg.d_vocab_out)
114 self.unembed = Unembed(self.cfg)
116 self.hook_embed = HookPoint()
118 if move_to_device: 118 ↛ 119line 118 didn't jump to line 119 because the condition on line 118 was never true
119 self.to(self.cfg.device)
121 self.setup()
123 def to_tokens(
124 self,
125 input: Union[str, List[str]],
126 move_to_device: bool = True,
127 truncate: bool = True,
128 ) -> Tuple[Int[torch.Tensor, "batch pos"], Int[torch.Tensor, "batch pos"]]:
129 """Converts a string to a tensor of tokens.
130 Taken mostly from the HookedTransformer implementation, but does not support default padding
131 sides or prepend_bos.
133 Args:
134 input (Union[str, List[str]]): The input to tokenize.
135 move_to_device (bool): Whether to move the output tensor of tokens to the device the
136 model lives on. Defaults to True
137 truncate (bool): If the output tokens are too long, whether to truncate the output
138 tokens to the model's max context window. Does nothing for shorter inputs.
139 Defaults to True.
140 """
142 assert self.tokenizer is not None, "Cannot use to_tokens without a tokenizer"
144 encodings = self.tokenizer(
145 input,
146 return_tensors="pt",
147 padding=True,
148 truncation=truncate,
149 max_length=self.cfg.n_ctx if truncate else None,
150 )
152 tokens = encodings.input_ids
153 attention_mask = encodings.attention_mask
155 if move_to_device: 155 ↛ 158line 155 didn't jump to line 158 because the condition on line 155 was always true
156 tokens = tokens.to(self.cfg.device)
157 attention_mask = attention_mask.to(self.cfg.device)
158 return tokens, attention_mask
160 @overload
161 def forward(
162 self,
163 input: Union[str, List[str], Int[torch.Tensor, "batch pos"]],
164 decoder_input: Optional[Int[torch.Tensor, "batch decoder_pos"]] = None,
165 return_type: Literal["logits"] = "logits",
166 one_zero_attention_mask: Optional[Int[torch.Tensor, "batch pos"]] = None,
167 ) -> Float[torch.Tensor, "batch pos d_vocab"]:
168 ...
170 @overload
171 def forward(
172 self,
173 input: Union[str, List[str], Int[torch.Tensor, "batch pos"]],
174 decoder_input: Optional[Int[torch.Tensor, "batch decoder_pos"]] = None,
175 return_type: Optional[Literal[None]] = None,
176 one_zero_attention_mask: Optional[Int[torch.Tensor, "batch pos"]] = None,
177 ) -> Optional[Float[torch.Tensor, "batch pos d_vocab"]]:
178 ...
180 def forward(
181 self,
182 input: Union[str, List[str], Int[torch.Tensor, "batch pos"]],
183 decoder_input: Optional[Int[torch.Tensor, "batch decoder_pos"]] = None,
184 return_type: Optional[str] = "logits",
185 one_zero_attention_mask: Optional[Int[torch.Tensor, "batch pos"]] = None,
186 ) -> Optional[Float[torch.Tensor, "batch decoder_pos d_vocab"]]:
187 """Forward pass of the T5 model.
189 Args:
190 input: Input to be processed. Can be one of:
191 - str: A single string input
192 - List[str]: A batch of string inputs
193 - Int[torch.Tensor, "batch pos"]: A batch of token IDs
194 decoder_input: Tensor of shape (batch, decoder_pos) containing the decoder input sequence.
195 If None and input is of type str or List[str], starts with batch of beginning-of-sequence (BOS) tokens.
196 return_type: Specifies the model output type:
197 - "logits": Return logits tensor
198 - None: Returns nothing
199 one_zero_attention_mask: A binary mask which indicates
200 which tokens should be attended to (1) and which should be ignored (0).
201 Primarily used for padding variable-length sentences in a batch.
202 For instance, in a batch with sentences of differing lengths, shorter
203 sentences are padded with 0s on the right. If not provided, the model
204 assumes all tokens should be attended to.
205 This parameter gets inferred from the tokenizer if input is a string or list of strings.
206 Shape is (batch_size, sequence_length).
208 Returns:
209 Optional[Float[torch.Tensor, "batch decoder_pos d_vocab"]]:
210 If return_type="logits": Returns logits tensor of shape (batch, decoder_pos, vocab_size)
211 If return_type=None: Returns None
212 """
214 if isinstance(input, (str, list)):
215 tokens, attention_mask = self.to_tokens(input)
217 # If attention mask is not provided, use the ones from the tokenizer
218 one_zero_attention_mask = (
219 attention_mask if one_zero_attention_mask is None else one_zero_attention_mask
220 )
222 # If decoder_input is not provided, start with tensor of PAD tokens of shape (batch, 1)
223 if decoder_input is None: 223 ↛ 243line 223 didn't jump to line 243 because the condition on line 223 was always true
224 assert self.tokenizer is not None
225 decoder_input = torch.full(
226 (tokens.shape[0], 1),
227 self.tokenizer.pad_token_id,
228 device=self.cfg.device,
229 )
230 else:
231 tokens = input
233 if one_zero_attention_mask is None:
234 logging.warning(
235 "No attention mask provided. Assuming all tokens should be attended to."
236 )
238 if decoder_input is None: 238 ↛ 239line 238 didn't jump to line 239 because the condition on line 238 was never true
239 raise ValueError(
240 "Must provide decoder_input if input is not a string or list of strings"
241 )
243 if tokens.device.type != self.cfg.device: 243 ↛ 244line 243 didn't jump to line 244 because the condition on line 243 was never true
244 tokens = tokens.to(self.cfg.device)
246 if one_zero_attention_mask is not None:
247 one_zero_attention_mask = one_zero_attention_mask.to(self.cfg.device)
249 resid = self.hook_embed(self.embed(tokens))
251 if one_zero_attention_mask is not None:
252 additive_attention_mask = (
253 repeat(1 - one_zero_attention_mask, "batch pos -> batch 1 1 pos")
254 ) * torch.finfo(self.cfg.dtype).min
255 else:
256 additive_attention_mask = None
258 query_len = key_len = tokens.shape[1]
260 encoder_positional_bias = cast(
261 T5Block, self.encoder[0]
262 ).attn.compute_relative_attention_bias(query_len, key_len, device=self.cfg.device)
264 for encoder_block in self.encoder:
265 resid = encoder_block(
266 resid_pre=resid,
267 additive_attention_mask=additive_attention_mask,
268 position_bias=encoder_positional_bias,
269 )
271 encoder_resid = self.encoder_final_ln(resid)
273 if decoder_input is None: 273 ↛ 274line 273 didn't jump to line 274 because the condition on line 273 was never true
274 raise ValueError("decoder_input cannot be None when input is not a string")
275 decoder_resid = self.embed(decoder_input)
276 decoder_query_len = decoder_key_len = decoder_input.shape[1]
277 decoder_positional_bias = cast(
278 T5Block, self.decoder[0]
279 ).attn.compute_relative_attention_bias(
280 decoder_query_len, decoder_key_len, device=self.cfg.device
281 )
283 for decoder_block in self.decoder:
284 decoder_resid = decoder_block(
285 resid_pre=decoder_resid,
286 position_bias=decoder_positional_bias,
287 encoder_hidden_states=encoder_resid,
288 encoder_additive_attention_mask=additive_attention_mask,
289 )
291 decoder_resid = self.decoder_final_ln(decoder_resid)
293 if self.cfg.tie_word_embeddings: 293 ↛ 298line 293 didn't jump to line 298 because the condition on line 293 was always true
294 # Rescale output before projecting on vocab
295 # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
296 decoder_resid *= self.cfg.d_model**-0.5
298 logits = self.unembed(decoder_resid)
299 if return_type is None: 299 ↛ 300line 299 didn't jump to line 300 because the condition on line 299 was never true
300 return None
301 return logits
303 @torch.inference_mode()
304 def generate(
305 self,
306 input: Union[str, Int[torch.Tensor, "batch pos"]] = "",
307 one_zero_attention_mask: Optional[Int[torch.Tensor, "batch pos"]] = None,
308 max_new_tokens: int = 10,
309 stop_at_eos: bool = True,
310 eos_token_id: Optional[Union[int, List[int]]] = None,
311 do_sample: bool = True,
312 top_k: Optional[int] = None,
313 top_p: Optional[float] = None,
314 temperature: float = 1.0,
315 freq_penalty: float = 0.0,
316 return_type: Optional[str] = "input",
317 verbose: bool = True,
318 ) -> Union[Int[torch.Tensor, "batch new_tokens"], str]:
319 """Sample tokens from the T5 encoder-decoder model.
321 Sample tokens from the model until the model outputs eos_token or max_new_tokens is reached.
322 This function is primarily taken from HookedTransformer but adjusted for the HookedEncoderDecoder
323 architecture.
324 This function does not support key value caching and no default padding sides or prepend_bos.
326 To avoid fiddling with ragged tensors, if we input a batch of text and some sequences finish
327 (by producing an EOT token), we keep running the model on the entire batch, but throw away
328 the output for a finished sequence and just keep adding EOTs to pad.
330 This supports entering a single string, but not a list of strings - if the strings don't
331 tokenize to exactly the same length, this gets messy. If that functionality is needed,
332 convert them to a batch of tokens and input that instead.
334 Args:
335 input (Union[str, Int[torch.Tensor, "batch pos"])]): Either a batch of tokens ([batch,
336 pos]) or a text string (this will be converted to a batch of tokens with batch size
337 1).
338 max_new_tokens (int): Maximum number of tokens to generate.
339 stop_at_eos (bool): If True, stop generating tokens when the model outputs eos_token.
340 eos_token_id (Optional[Union[int, Sequence]]): The token ID to use for end
341 of sentence. If None, use the tokenizer's eos_token_id - required if using
342 stop_at_eos. It's also possible to provide a list of token IDs (not just the
343 eos_token_id), in which case the generation will stop when any of them are output
344 (useful e.g. for stable_lm).
345 do_sample (bool): If True, sample from the model's output distribution. Otherwise, use
346 greedy search (take the max logit each time).
347 top_k (int): Number of tokens to sample from. If None, sample from all tokens.
348 top_p (float): Probability mass to sample from. If 1.0, sample from all tokens. If <1.0,
349 we take the top tokens with cumulative probability >= top_p.
350 temperature (float): Temperature for sampling. Higher values will make the model more
351 random (limit of temp -> 0 is just taking the top token, limit of temp -> inf is
352 sampling from a uniform distribution).
353 freq_penalty (float): Frequency penalty for sampling - how much to penalise previous
354 tokens. Higher values will make the model more random.
355 return_type (Optional[str]): The type of the output to return - either a string (str),
356 a tensor of tokens (tensor) or whatever the format of the input was (input).
357 verbose (bool): If True, show tqdm progress bars for generation.
359 Returns:
360 outputs (torch.Tensor): [batch, new_tokens], generated sequence of new tokens
361 (by default returns same type as input).
362 """
364 if isinstance(input, str): 364 ↛ 376line 364 didn't jump to line 376 because the condition on line 364 was always true
365 # If text, convert to tokens (batch_size=1)
366 assert (
367 self.tokenizer is not None
368 ), "Must provide a tokenizer if passing a string to the model"
369 encoder_input, attention_mask = self.to_tokens(input)
371 # If attention mask is not provided, use the one from the tokenizer
372 one_zero_attention_mask = (
373 attention_mask if one_zero_attention_mask is None else one_zero_attention_mask
374 )
375 else:
376 assert isinstance(input, torch.Tensor) # keep mypy happy
377 encoder_input = input
379 # If tokens are provided, user should be aware that attention mask will not be inferred
380 if one_zero_attention_mask is None:
381 logging.warning(
382 "No attention mask provided. Assuming all tokens should be attended to."
383 )
385 if return_type == "input": 385 ↛ 391line 385 didn't jump to line 391 because the condition on line 385 was always true
386 if isinstance(input, str): 386 ↛ 389line 386 didn't jump to line 389 because the condition on line 386 was always true
387 return_type = "str"
388 else:
389 return_type = "tensor"
391 assert isinstance(encoder_input, torch.Tensor)
392 batch_size = encoder_input.shape[0]
393 device = devices.get_device_for_block_index(0, self.cfg)
395 # For the decoder input, we start with a tensor of PAD tokens of shape (batch, 1)
396 assert self.tokenizer is not None
397 decoder_input = torch.full((batch_size, 1), self.tokenizer.pad_token_id).to(device)
399 stop_tokens: List[int] = []
400 eos_token_for_padding = 0
401 if stop_at_eos: 401 ↛ 427line 401 didn't jump to line 427 because the condition on line 401 was always true
402 tokenizer_has_eos_token = self.tokenizer.eos_token_id is not None
404 local_eos_token_id: Optional[Union[int, List[int]]] = eos_token_id
405 if local_eos_token_id is None: 405 ↛ 412line 405 didn't jump to line 412 because the condition on line 405 was always true
406 assert (
407 tokenizer_has_eos_token
408 ), "Must pass a eos_token_id if stop_at_eos is True and tokenizer is None or has no eos_token_id"
410 local_eos_token_id = self.tokenizer.eos_token_id
412 if isinstance(local_eos_token_id, int): 412 ↛ 417line 412 didn't jump to line 417 because the condition on line 412 was always true
413 stop_tokens = [local_eos_token_id]
414 eos_token_for_padding = local_eos_token_id
415 else:
416 # eos_token_id is a Sequence (e.g. list or tuple)
417 if local_eos_token_id is None:
418 raise ValueError("eos_token_id cannot be None here")
419 stop_tokens = local_eos_token_id
420 eos_token_for_padding = (
421 self.tokenizer.eos_token_id
422 if tokenizer_has_eos_token
423 else local_eos_token_id[0]
424 )
426 # An array to track which sequences in the batch have finished.
427 finished_sequences = torch.zeros(batch_size, dtype=torch.bool, device=self.cfg.device)
429 # Currently nothing in HookedTransformer changes with eval, but this is here in case
430 # that changes in the future.
431 self.eval()
432 for _ in tqdm.tqdm(range(max_new_tokens), disable=not verbose): 432 ↛ 481line 432 didn't jump to line 481 because the loop on line 432 didn't complete
433 # While generating, we keep generating logits, throw away all but the final logits,
434 # and then use those logits to sample from the distribution We keep adding the
435 # sampled tokens to the end of tokens.
436 # We input the entire sequence, as a [batch, pos] tensor, since we aren't using
437 # the cache.
439 # Encoder input will be the same for all iterations
440 # Decoder input will be appended with the new token each iteration
441 logits = self.forward(
442 encoder_input,
443 decoder_input=decoder_input,
444 one_zero_attention_mask=one_zero_attention_mask,
445 )
446 assert logits is not None
447 final_logits = logits[:, -1, :]
449 if do_sample: 449 ↛ 450line 449 didn't jump to line 450 because the condition on line 449 was never true
450 sampled_tokens = sample_logits(
451 final_logits,
452 top_k=top_k,
453 top_p=top_p,
454 temperature=temperature,
455 freq_penalty=freq_penalty,
456 tokens=decoder_input,
457 ).to(devices.get_device_for_block_index(0, self.cfg))
458 else:
459 sampled_tokens = final_logits.argmax(-1).to(
460 devices.get_device_for_block_index(0, self.cfg)
461 )
463 if stop_at_eos: 463 ↛ 476line 463 didn't jump to line 476 because the condition on line 463 was always true
464 # For all unfinished sequences, add on the next token. If a sequence was
465 # finished, throw away the generated token and add eos_token_for_padding
466 # instead.
467 sampled_tokens[finished_sequences] = eos_token_for_padding
468 finished_sequences.logical_or_(
469 torch.isin(
470 sampled_tokens.to(self.cfg.device),
471 torch.tensor(stop_tokens).to(self.cfg.device),
472 )
473 )
475 # Append new token to the decoder input
476 decoder_input = torch.cat([decoder_input, sampled_tokens.unsqueeze(-1)], dim=-1)
478 if stop_at_eos and finished_sequences.all():
479 break
481 if return_type == "str": 481 ↛ 487line 481 didn't jump to line 487 because the condition on line 481 was always true
482 assert self.tokenizer is not None
483 # Convert tokens to string
484 return self.tokenizer.decode(decoder_input[0], skip_special_tokens=True)
486 else:
487 return decoder_input
489 @overload
490 def run_with_cache(
491 self, *model_args: Any, return_cache_object: Literal[True] = True, **kwargs: Any
492 ) -> Tuple[Float[torch.Tensor, "batch pos d_vocab"], ActivationCache]:
493 ...
495 @overload
496 def run_with_cache(
497 self, *model_args: Any, return_cache_object: Literal[False] = False, **kwargs: Any
498 ) -> Tuple[Float[torch.Tensor, "batch pos d_vocab"], Dict[str, torch.Tensor]]:
499 ...
501 def run_with_cache(
502 self,
503 *model_args: Any,
504 return_cache_object: bool = True,
505 remove_batch_dim: bool = False,
506 **kwargs: Any,
507 ) -> Tuple[
508 Float[torch.Tensor, "batch pos d_vocab"],
509 Union[ActivationCache, Dict[str, torch.Tensor]],
510 ]:
511 """
512 Wrapper around run_with_cache in HookedRootModule. If return_cache_object is True, this will return an ActivationCache object, with a bunch of useful HookedTransformer specific methods, otherwise it will return a dictionary of activations as in HookedRootModule. This function was copied directly from HookedTransformer.
513 """
514 out, cache_dict = super().run_with_cache(
515 *model_args, remove_batch_dim=remove_batch_dim, **kwargs
516 )
517 if return_cache_object: 517 ↛ 521line 517 didn't jump to line 521 because the condition on line 517 was always true
518 cache = ActivationCache(cache_dict, self, has_batch_dim=not remove_batch_dim)
519 return out, cache
520 else:
521 return out, cache_dict
523 def to(self: T, *args: Any, **kwargs: Any) -> T:
524 return super().to(*args, **kwargs)
526 def cuda(self: T, device: Optional[Union[int, torch.device]] = None) -> T:
527 if isinstance(device, int):
528 return self.to(f"cuda:{device}")
529 elif device is None:
530 return self.to("cuda")
531 else:
532 return self.to(device)
534 def cpu(self: T) -> T:
535 return self.to("cpu")
537 def mps(self: T) -> T:
538 return self.to(torch.device("mps"))
540 @classmethod
541 def from_pretrained(
542 cls: Type[T],
543 model_name: str,
544 checkpoint_index: Optional[int] = None,
545 checkpoint_value: Optional[int] = None,
546 hf_model: Optional[Any] = None,
547 device: Optional[str] = None,
548 tokenizer: Optional[Any] = None,
549 move_to_device: bool = True,
550 dtype: Optional[torch.dtype] = torch.float32,
551 **from_pretrained_kwargs: Any,
552 ) -> T:
553 """Loads in the pretrained weights from huggingface. Currently supports loading weight from HuggingFace BertForMaskedLM. Unlike HookedTransformer, this does not yet do any preprocessing on the model."""
554 logging.warning(
555 "Support for T5 in TransformerLens is currently experimental, until such a time when it has feature "
556 "parity with HookedTransformer and has been tested on real research tasks. Until then, backward "
557 "compatibility is not guaranteed. Please see the docs for information on the limitations of the current "
558 "implementation."
559 "\n"
560 "If using T5 for interpretability research, keep in mind that T5 has some significant architectural "
561 "differences to GPT. The major one is that T5 is an Encoder-Decoder model"
562 "Also, it uses relative positional embeddings, different types of Attention (without bias) and LayerNorm"
563 )
565 if from_pretrained_kwargs.get("load_in_8bit", False) or from_pretrained_kwargs.get( 565 ↛ 568line 565 didn't jump to line 568 because the condition on line 565 was never true
566 "load_in_4bit", False
567 ):
568 raise ValueError("Quantization not supported")
570 if "torch_dtype" in from_pretrained_kwargs: 570 ↛ 571line 570 didn't jump to line 571 because the condition on line 570 was never true
571 dtype = from_pretrained_kwargs["torch_dtype"]
573 if dtype is None: 573 ↛ 574line 573 didn't jump to line 574 because the condition on line 573 was never true
574 dtype = torch.float32
576 name_or_path = (
577 model_name if Path(model_name).exists() else loading.get_official_model_name(model_name)
578 )
580 cfg = loading.get_pretrained_model_config(
581 name_or_path,
582 checkpoint_index=checkpoint_index,
583 checkpoint_value=checkpoint_value,
584 fold_ln=False,
585 device=device,
586 n_devices=1,
587 dtype=dtype,
588 **from_pretrained_kwargs,
589 )
591 state_dict = loading.get_pretrained_state_dict(
592 name_or_path, cfg, hf_model, dtype=dtype, **from_pretrained_kwargs
593 )
595 model = cls(cfg, tokenizer, move_to_device=False)
597 model.load_state_dict(state_dict, strict=False)
599 if move_to_device: 599 ↛ 602line 599 didn't jump to line 602 because the condition on line 599 was always true
600 model.to(cfg.device)
602 print(f"Loaded pretrained model {model_name} into HookedTransformer")
604 return model
606 @property
607 def W_U(self) -> Float[torch.Tensor, "d_model d_vocab"]:
608 """
609 Convenience to get the unembedding matrix (ie the linear map from the final residual stream to the output logits)
610 """
611 return self.unembed.W_U
613 @property
614 def b_U(self) -> Float[torch.Tensor, "d_vocab"]:
615 """
616 Convenience to get the unembedding bias
617 """
618 return self.unembed.b_U
620 @property
621 def W_E(self) -> Float[torch.Tensor, "d_vocab d_model"]:
622 """
623 Convenience to get the embedding matrix
624 """
625 return self.embed.W_E
627 @property
628 def W_pos(self) -> None:
629 """
630 Convenience function to get the positional embedding. Only works on models with absolute positional embeddings!
631 """
632 raise NotImplementedError(
633 "T5 does not have absolute positional embeddings. Uses relative positional embeddings instead."
634 )
636 @property
637 def W_K(self) -> Float[torch.Tensor, "n_layers n_heads d_model d_head"]:
638 """Stacks the key weights across all layers"""
639 return torch.stack( 639 ↛ exit, 639 ↛ exit2 missed branches: 1) line 639 didn't jump to the function exit, 2) line 639 didn't return from function 'W_K' because the return on line 639 wasn't executed
640 [cast(T5Block, block).attn.W_K for block in chain(self.encoder, self.decoder)],
641 dim=0,
642 )
644 @property
645 def W_Q(self) -> Float[torch.Tensor, "n_layers n_heads d_model d_head"]:
646 """Stacks the query weights across all layers"""
647 return torch.stack( 647 ↛ exit, 647 ↛ exit2 missed branches: 1) line 647 didn't jump to the function exit, 2) line 647 didn't return from function 'W_Q' because the return on line 647 wasn't executed
648 [cast(T5Block, block).attn.W_Q for block in chain(self.encoder, self.decoder)],
649 dim=0,
650 )
652 @property
653 def W_V(self) -> Float[torch.Tensor, "n_layers n_heads d_model d_head"]:
654 """Stacks the value weights across all layers"""
655 return torch.stack( 655 ↛ exit, 655 ↛ exit2 missed branches: 1) line 655 didn't jump to the function exit, 2) line 655 didn't return from function 'W_V' because the return on line 655 wasn't executed
656 [cast(T5Block, block).attn.W_V for block in chain(self.encoder, self.decoder)],
657 dim=0,
658 )
660 @property
661 def W_O(self) -> Float[torch.Tensor, "n_layers n_heads d_head d_model"]:
662 """Stacks the attn output weights across all layers"""
663 return torch.stack( 663 ↛ exit, 663 ↛ exit2 missed branches: 1) line 663 didn't jump to the function exit, 2) line 663 didn't return from function 'W_O' because the return on line 663 wasn't executed
664 [cast(T5Block, block).attn.W_O for block in chain(self.encoder, self.decoder)],
665 dim=0,
666 )
668 @property
669 def W_in(self) -> Float[torch.Tensor, "n_layers d_model d_mlp"]:
670 """Stacks the MLP input weights across all layers"""
671 weights: List[torch.Tensor] = []
672 for block in chain(self.encoder, self.decoder):
673 mlp = cast(T5Block, block).mlp
674 if isinstance(mlp, (MLP, GatedMLP)):
675 weights.append(mlp.W_in)
676 else:
677 raise NotImplementedError(
678 f"W_in property is not supported for MLP of type {type(mlp).__name__}"
679 )
680 return torch.stack(weights, dim=0)
682 @property
683 def W_out(self) -> Float[torch.Tensor, "n_layers d_mlp d_model"]:
684 """Stacks the MLP output weights across all layers"""
685 weights: List[torch.Tensor] = []
686 for block in chain(self.encoder, self.decoder):
687 mlp = cast(T5Block, block).mlp
688 if isinstance(mlp, (MLP, GatedMLP)):
689 weights.append(mlp.W_out)
690 else:
691 raise NotImplementedError(
692 f"W_out property is not supported for MLP of type {type(mlp).__name__}"
693 )
694 return torch.stack(weights, dim=0)
696 @property
697 def b_K(self) -> Float[torch.Tensor, "n_layers n_heads d_head"]:
698 """Stacks the key biases across all layers"""
699 return torch.stack( 699 ↛ exit, 699 ↛ exit2 missed branches: 1) line 699 didn't jump to the function exit, 2) line 699 didn't return from function 'b_K' because the return on line 699 wasn't executed
700 [cast(T5Block, block).attn.b_K for block in chain(self.encoder, self.decoder)],
701 dim=0,
702 )
704 @property
705 def b_Q(self) -> Float[torch.Tensor, "n_layers n_heads d_head"]:
706 """Stacks the query biases across all layers"""
707 return torch.stack( 707 ↛ exit, 707 ↛ exit2 missed branches: 1) line 707 didn't jump to the function exit, 2) line 707 didn't return from function 'b_Q' because the return on line 707 wasn't executed
708 [cast(T5Block, block).attn.b_Q for block in chain(self.encoder, self.decoder)],
709 dim=0,
710 )
712 @property
713 def b_V(self) -> Float[torch.Tensor, "n_layers n_heads d_head"]:
714 """Stacks the value biases across all layers"""
715 return torch.stack( 715 ↛ exit, 715 ↛ exit2 missed branches: 1) line 715 didn't jump to the function exit, 2) line 715 didn't return from function 'b_V' because the return on line 715 wasn't executed
716 [cast(T5Block, block).attn.b_V for block in chain(self.encoder, self.decoder)],
717 dim=0,
718 )
720 @property
721 def b_O(self) -> Float[torch.Tensor, "n_layers d_model"]:
722 """Stacks the attn output biases across all layers"""
723 return torch.stack( 723 ↛ exit, 723 ↛ exit2 missed branches: 1) line 723 didn't jump to the function exit, 2) line 723 didn't return from function 'b_O' because the return on line 723 wasn't executed
724 [cast(T5Block, block).attn.b_O for block in chain(self.encoder, self.decoder)],
725 dim=0,
726 )
728 @property
729 def b_in(self) -> Float[torch.Tensor, "n_layers d_mlp"]:
730 """Stacks the MLP input biases across all layers"""
731 biases: List[torch.Tensor] = []
732 for block in chain(self.encoder, self.decoder):
733 mlp = cast(T5Block, block).mlp
734 if isinstance(mlp, (MLP, GatedMLP)):
735 biases.append(mlp.b_in)
736 else:
737 raise NotImplementedError(
738 f"b_in property is not supported for MLP of type {type(mlp).__name__}"
739 )
740 return torch.stack(biases, dim=0)
742 @property
743 def b_out(self) -> Float[torch.Tensor, "n_layers d_model"]:
744 """Stacks the MLP output biases across all layers"""
745 biases: List[torch.Tensor] = []
746 for block in chain(self.encoder, self.decoder):
747 mlp = cast(T5Block, block).mlp
748 if isinstance(mlp, (MLP, GatedMLP)):
749 biases.append(mlp.b_out)
750 else:
751 raise NotImplementedError(
752 f"b_out property is not supported for MLP of type {type(mlp).__name__}"
753 )
754 return torch.stack(biases, dim=0)
756 @property
757 def QK(self) -> FactoredMatrix: # [n_layers, n_heads, d_model, d_model]
758 """Returns a FactoredMatrix object with the product of the Q and K matrices for each layer and head.
759 Useful for visualizing attention patterns."""
760 return FactoredMatrix(self.W_Q, self.W_K.transpose(-2, -1))
762 @property
763 def OV(self) -> FactoredMatrix: # [n_layers, n_heads, d_model, d_model]
764 """Returns a FactoredMatrix object with the product of the O and V matrices for each layer and head."""
765 return FactoredMatrix(self.W_V, self.W_O)
767 def all_head_labels(self) -> List[str]:
768 """Returns a list of strings with the format "L{l}H{h}", where l is the layer index and h is the head index."""
769 return [f"EL{l}H{h}" for l in range(self.cfg.n_layers) for h in range(self.cfg.n_heads)] + [ 769 ↛ exit, 769 ↛ exit2 missed branches: 1) line 769 didn't run the list comprehension on line 769 or line 769 didn't run the list comprehension on line 769, 2) line 769 didn't return from function 'all_head_labels' because the return on line 769 wasn't executed
770 f"DL{l}H{h}" for l in range(self.cfg.n_layers) for h in range(self.cfg.n_heads)
771 ]