Coverage for transformer_lens/HookedTransformerConfig.py: 87%
122 statements
« prev ^ index » next coverage.py v7.4.4, created at 2024-06-11 01:46 +0000
« prev ^ index » next coverage.py v7.4.4, created at 2024-06-11 01:46 +0000
1"""Hooked Transformer Config.
3Module with a dataclass for storing the configuration of a
4:class:`transformer_lens.HookedTransformer` model.
5"""
7from __future__ import annotations
9import logging
10import pprint
11import random
12from dataclasses import dataclass
13from typing import Any, Dict, List, Optional, Union
15import numpy as np
16import torch
18from transformer_lens import utils
20SUPPORTED_ACTIVATIONS = ["relu", "gelu", "silu", "gelu_new", "solu_ln", "gelu_fast"]
23@dataclass 23 ↛ 25line 23 didn't jump to line 25, because
24class HookedTransformerConfig:
25 """
26 Configuration class to store the configuration of a HookedTransformer model.
28 See further_comments.md for more details on the more complex arguments.
30 Args:
31 d_model (int): The dimensionality of the embeddings.
32 d_head (int): The dimensionality of each attention head.
33 n_layers (int): The number of transformer blocks (one block = one attn layer AND one MLP layer).
34 n_ctx (int): The maximum sequence length.
35 n_heads (int): The number of attention heads. If not
36 specified, will be set to d_model // d_head. (This is represented by a default value of -1)
37 d_mlp (int, *optional*): The dimensionality of the feedforward mlp
38 network. Defaults to 4 * d_model, and in an attn-only model is None.
39 d_vocab (int): The size of the vocabulary. Defaults to -1, which means not set. If not set, will be
40 automatically set from the tokenizer's vocab size.
41 act_fn (str, *optional*): The activation function to use. Always
42 lowercase. Supports ['relu', 'gelu', 'silu', 'gelu_new', 'solu_ln',
43 'gelu_fast']. Must be set unless using an attn-only model.
44 eps (float): The epsilon value to use for layer normalization. Defaults
45 to 1e-5
46 use_attn_result (bool): whether to explicitly calculate the amount
47 each head adds to the residual stream (with a hook) and THEN add it
48 up, vs just calculating the sum. This can be very memory intensive
49 for large models, so defaults to False
50 use_split_qkv_input (bool): whether to explicitly calculate the input of
51 each head separately, with a hook. Defaults to false to save memory.
52 use_hook_mlp_in (bool): whether to use a hook to get the input to the
53 MLP layer. Defaults to false to save memory.
54 use_attn_in (bool): whether to explicitly calculate the input of each
55 attention head separately, with a hook. Defaults to false to save memory
56 use_attn_scale (bool): whether to scale the attention weights by
57 1/sqrt(d_head)
58 model_name (str): the name of the model, used to load
59 weights from HuggingFace or initialized to "custom" if not passed
60 original_architecture (str, *optional*): the family of the model, used
61 to help load
62 weights from HuggingFace or initialized to "custom" if not passed
63 from_checkpoint (bool): Whether the model weights were
64 loaded from a checkpoint (only applies to pretrained models)
65 checkpoint_index (int, *optional*): The index of the
66 checkpoint loaded (only applies to pretrained models).
67 checkpoint_label_type (str, *optional*): Whether
68 checkpoints are labelled by the number of steps or number of tokens.
69 checkpoint_value (int, *optional*): The value of the
70 checkpoint label (whether of steps or tokens).
71 tokenizer_name (str, *optional*): the full name of the model, passed into
72 HuggingFace to access the tokenizer. Only used when passing in
73 custom config, if loading from pretrained then this is not needed.
74 use_local_attn (bool): whether to use local attention - ie each
75 destination token can only attend to source tokens a certain distance back.
76 window_size (int, *optional*): the size of the window for local
77 attention
78 attn_types (List[str], *optional*): the types of attention to use for
79 local attention
80 weight_init_mode (str): the initialization mode to use for the
81 weights. Only relevant for custom models, ignored for pre-trained.
82 We now support 'gpt2', 'xavier_uniform', 'xavier_normal', 'kaiming_uniform',
83 'kaiming_normal'. MuP support to come. Defaults to 'gpt2'.
84 normalization_type (str, *optional*): the type of normalization to use.
85 Options are None (no normalization), 'LN' (use LayerNorm, including weights
86 & biases) and 'LNPre' (use LayerNorm, but no weights & biases).
87 Defaults to LN
88 device(str): The device to use for the model. Defaults to 'cuda' if
89 available, else 'cpu'. Must be 'cuda' if `n_devices` > 1.
90 n_devices (int): The number of devices to use for the model. Defaults to 1. Layers are loaded
91 to support "pipeline parallelism", where each device is responsible for a subset of the layers.
92 attention_dir (str): Whether to use causal (aka unidirectional aka GPT-2
93 style) or bidirectional attention. Options are 'causal' and
94 'bidirectional'. Defaults to 'causal'
95 attn_only (bool): Whether to only use attention layers, no feedforward
96 layers. Defaults to False
97 seed (int, *optional*): The seed to use for the model.
98 Used to set sources of randomness (Python, PyTorch and NumPy) and to initialize weights.
99 Defaults to None. We recommend setting a seed, so your experiments are reproducible.
100 initializer_range (float): The standard deviation of the normal used to
101 initialise the weights, initialized to 0.8 / sqrt(d_model). If weight_init_mode is
102 'xavier_uniform' or 'xavier_normal', this value is instead treated as the `gain` parameter for the weight
103 initialisation (a constant factor to scale the weights by). Defaults to -1.0, which means not set.
104 init_weights (bool): Whether to initialize the weights. Defaults to
105 True. If False, does not initialize weights.
106 scale_attn_by_inverse_layer_idx (bool): Whether to scale the attention
107 weights by 1/(layer_id+1), used by Mistral (Stanford) models for numerical stability when
108 training in FP16. Defaults to False.
109 positional_embedding_type (str): The positional embedding used. Options
110 are 'standard' (ie GPT-2 style, absolute, randomly initialized learned positional
111 embeddings, directly added to the residual stream), 'rotary'
112 (described here: https://blog.eleuther.ai/rotary-embeddings/ ) and
113 'shortformer' (GPT-2 style absolute & learned, but rather than being
114 added to the residual stream they're only added to the inputs to the
115 keys and the queries (ie key = W_K(res_stream + pos_embed), but
116 values and MLPs don't get any positional info)). Sinusoidal are not
117 currently supported. Defaults to 'standard'.
118 final_rms (bool): Whether to replace the final normalization (just
119 before the unembed) with RMSNorm (ie no centering or bias, just
120 scaling + weights). Only included because of a dumb bug in my
121 original SoLU code. Defaults to False.
122 d_vocab_out (int, *optional*): The size of the output vocabulary. Defaults to -1, which means not set. If not
123 set, will be equal to d_vocab. Mainly useful for algorithmic tasks
124 where the input and output vocabularies may be different.
125 parallel_attn_mlp (bool): Whether to parallelize the attention and MLP
126 layers - a weird cursed thing done by GPT-J. Means that
127 mlp_out=MLP(ln1(resid_pre)) and resid_post=resid_pre+attn_out+mlp_out. Defaults to False.
128 rotary_dim (int, *optional*): The dimensionality of the rotary
129 embeddings, may be d_head in which case only the first rotary_dim
130 dimensions of each head are rotated. Defaults to None, if
131 positional_embedding_type=="rotary" it defaults to d_head.
132 n_params (int, *optional*): The number of (hidden weight)
133 parameters in the model. This is automatically calculated and not
134 intended to be set by the user. (Non embedding parameters, because
135 the [scaling laws paper](https://arxiv.org/pdf/2001.08361.pdf) found
136 that that was a more meaningful number. Ignoring biases and layer
137 norms, for convenience)
138 use_hook_tokens (bool): Will add a hook point on the token input to
139 HookedTransformer.forward, which lets you cache or intervene on the tokens.
140 Defaults to False.
141 default_prepend_bos (bool, optional): Default behavior of whether to prepend the BOS token when the
142 methods of HookedTransformer process input text to tokenize (only when input is a string).
143 Defaults to True - even for models not explicitly trained with this, heads often use the
144 first position as a resting position and accordingly lose information from the first token,
145 so this empirically seems to give better results. To change the default behavior to False, pass in
146 default_prepend_bos=False. Note that you can also locally override the default behavior by passing
147 in prepend_bos=True/False when you call a method that processes the input string.
148 dtype (torch.dtype, *optional*): The model's dtype. Defaults to torch.float32.
149 tokenizer_prepends_bos (bool, *optional*): This flag is set by set_tokenizer. It is set to True only
150 when the tokenizer automatically prepends the BOS token if initialized with add_bos_token=True.
151 We need this information to dynamically control bos prepending.
152 load_in_4bit(bool): If this flag is set, then it's assumed that parameters are 4-bit quantized
153 with bitsandbytes. Currently only supported for Llama.
154 n_key_value_heads (int, *optional*): The number of groups of heads that use the same key and value matrix.
155 Only for models that use Grouped Query Attention.
156 post_embedding_ln (bool): Whether to apply layer normalization after embedding the tokens. Defaults
157 to False.
158 num_experts (int, *optional*): The number of experts to use in the MoE layer. If set, experts_per_token
159 must also be set. Set to None if not using MoE.
160 experts_per_token (int, *optional*): The number of experts to use for each pass in the MoE layer. If set,
161 num_experts must also be set. Set to None if not using MoE.
162 relative_attention_max_distance (int, *optional*): The maximum distance between tokens for relative
163 attention. If set, relative_attention_num_buckets must also be set.Only used in EncoderDecoder models, like T5.
164 relative_attention_num_buckets (int, *optional*): The number of buckets to use for relative attention.
165 If set, relative_attention_max_distance must also be set.Only used in EncoderDecoder models, like T5.
166 decoder_start_token_id (int, *optional*): The start token id for the decoder. Only used in EncoderDecoder models, like T5.
167 tie_word_embeddings (bool): Whether to tie the word embeddings and the output layer weights. Defaults to False. Only used in EncoderDecoder (T5) by now.
168 """
170 n_layers: int
171 d_model: int
172 n_ctx: int
173 d_head: int
174 model_name: str = "custom"
175 n_heads: int = -1
176 d_mlp: Optional[int] = None
177 act_fn: Optional[str] = None
178 d_vocab: int = -1
179 eps: float = 1e-5
180 use_attn_result: bool = False
181 use_attn_scale: bool = True
182 use_split_qkv_input: bool = False
183 use_hook_mlp_in: bool = False
184 use_attn_in: bool = False
185 use_local_attn: bool = False
186 original_architecture: Optional[str] = None
187 from_checkpoint: bool = False
188 checkpoint_index: Optional[int] = None
189 checkpoint_label_type: Optional[str] = None
190 checkpoint_value: Optional[int] = None
191 tokenizer_name: Optional[str] = None
192 window_size: Optional[int] = None
193 attn_types: Optional[List] = None
194 init_mode: str = "gpt2"
195 normalization_type: Optional[str] = "LN"
196 device: Optional[str] = None
197 n_devices: int = 1
198 attention_dir: str = "causal"
199 attn_only: bool = False
200 seed: Optional[int] = None
201 initializer_range: float = -1.0
202 init_weights: bool = True
203 scale_attn_by_inverse_layer_idx: bool = False
204 positional_embedding_type: str = "standard"
205 final_rms: bool = False
206 d_vocab_out: int = -1
207 parallel_attn_mlp: bool = False
208 rotary_dim: Optional[int] = None
209 n_params: Optional[int] = None
210 use_hook_tokens: bool = False
211 gated_mlp: bool = False
212 default_prepend_bos: bool = True
213 dtype: torch.dtype = torch.float32
214 tokenizer_prepends_bos: Optional[bool] = None
215 n_key_value_heads: Optional[int] = None
216 post_embedding_ln: bool = False
217 rotary_base: int = 10000
218 trust_remote_code: bool = False
219 rotary_adjacent_pairs: bool = False
220 load_in_4bit: bool = False
221 num_experts: Optional[int] = None
222 experts_per_token: Optional[int] = None
223 relative_attention_max_distance: Optional[int] = None
224 relative_attention_num_buckets: Optional[int] = None
225 decoder_start_token_id: Optional[int] = None
226 tie_word_embeddings: bool = False
228 def __post_init__(self):
229 if self.n_heads == -1:
230 self.n_heads = self.d_model // self.d_head
232 if not self.d_model % (self.d_head) == 0: 232 ↛ 233line 232 didn't jump to line 233, because the condition on line 232 was never true
233 logging.warning(
234 "d_model %d is not divisible by d_head %d."
235 "n_heads was inferred to be %d, rounding down the ratio.",
236 self.d_model,
237 self.d_head,
238 self.n_heads,
239 )
241 if self.seed is not None: 241 ↛ 242line 241 didn't jump to line 242, because the condition on line 241 was never true
242 self.set_seed_everywhere(self.seed)
243 if self.use_local_attn:
244 assert self.window_size is not None, "window_size must be specified for local attention"
245 assert self.attn_types is not None, "attn_types must be specified for local attention"
246 if not self.attn_only:
247 if self.d_mlp is None:
248 # For some reason everyone hard codes in this hyper-parameter!
249 self.d_mlp: int = self.d_model * 4
250 assert self.act_fn is not None, "act_fn must be specified for non-attn-only models"
251 assert (
252 self.act_fn in SUPPORTED_ACTIVATIONS
253 ), f"act_fn={self.act_fn} must be one of {SUPPORTED_ACTIVATIONS}"
254 if self.initializer_range < 0 and self.init_mode == "gpt2": 254 ↛ 257line 254 didn't jump to line 257, because the condition on line 254 was never false
255 # Roughly copy the GPT-2 value, but proportional to sqrt(1/d_model)
256 self.initializer_range = 0.8 / np.sqrt(self.d_model)
257 if self.initializer_range < 0 and self.init_mode != "gpt2": 257 ↛ 259line 257 didn't jump to line 259, because the condition on line 257 was never true
258 # This is the gain parameter for the weight initialisation
259 self.initializer_range = 1.0
261 if self.d_vocab_out == -1:
262 # d_vocab_out defaults to d_vocab, unless there's an algorithmic task
263 # If d_vocab is not set, it'll be inferred from tokenizer_name or from a tokenizer
264 # explicitly passed to HookedTransformer initialisation.
265 self.d_vocab_out = self.d_vocab
267 if self.positional_embedding_type == "rotary" and self.rotary_dim is None: 267 ↛ 268line 267 didn't jump to line 268, because the condition on line 267 was never true
268 self.rotary_dim = self.d_head
270 if self.num_experts is not None: 270 ↛ 271line 270 didn't jump to line 271, because the condition on line 270 was never true
271 assert (
272 self.experts_per_token is not None
273 ), "experts_per_token must be set if num_experts is set"
274 if self.experts_per_token is not None: 274 ↛ 275line 274 didn't jump to line 275, because the condition on line 274 was never true
275 assert (
276 self.num_experts is not None
277 ), "num_experts must be set if experts_per_token is set"
279 # The number of parameters in attention layers (ignoring biases and layer norm). 4 because W_Q, W_K, W_V and W_O
280 self.n_params = self.n_layers * ((self.d_model * self.d_head * self.n_heads * 4))
281 if not self.attn_only:
282 assert self.d_mlp is not None # mypy
283 # Number of parameters in MLP layers (ignoring biases and layer norm). 2 because W_in and W_out
284 mlp_params_per_layer = self.d_model * self.d_mlp * (2 + self.gated_mlp)
286 if self.num_experts: 286 ↛ 288line 286 didn't jump to line 288, because the condition on line 286 was never true
287 # If we are using MoE, we multiply by num_experts, and add the expert gate parameters (d_model * num_experts)
288 mlp_params_per_layer = (mlp_params_per_layer + self.d_model) * self.num_experts
289 self.n_params += self.n_layers * mlp_params_per_layer
291 if self.device is None:
292 self.device = utils.get_device()
294 if self.n_devices > 1: 294 ↛ 295line 294 didn't jump to line 295, because the condition on line 294 was never true
295 assert (
296 torch.cuda.device_count() >= self.n_devices
297 ), f"Not enough CUDA devices to support n_devices {self.n_devices}"
299 assert self.default_prepend_bos in [
300 True,
301 False,
302 ], f"padding_side must be either True or False, but {self.default_prepend_bos} is given"
304 @classmethod
305 def unwrap(cls, config: Union[Dict, "HookedTransformerConfig"]) -> HookedTransformerConfig:
306 """
307 Convenience function to avoid duplicate code from a common way config is passed to various components
308 """
309 return HookedTransformerConfig.from_dict(config) if isinstance(config, Dict) else config
311 @classmethod
312 def from_dict(cls, config_dict: Dict[str, Any]) -> HookedTransformerConfig:
313 """
314 Instantiates a `HookedTransformerConfig` from a Python dictionary of
315 parameters.
316 """
317 return cls(**config_dict)
319 def to_dict(self):
320 return self.__dict__
322 def __repr__(self):
323 return "HookedTransformerConfig:\n" + pprint.pformat(self.to_dict())
325 def set_seed_everywhere(self, seed: int):
326 torch.manual_seed(seed)
327 random.seed(seed)
328 np.random.seed(seed)