Spaces:
Running
Running
Commit
β’
6414c6a
1
Parent(s):
438fb10
Reorginize lenses (#32)
Browse files- reorginized lenses (02a3e822d0a0d80d92d6f4741359286d77f4597c)
- lens/{gpt-neox-20b β EleutherAI/gpt-neox-20b}/config.json +0 -0
- lens/{gpt-neox-20b β EleutherAI/gpt-neox-20b}/params.pt +0 -0
- lens/{pythia-1.4b-deduped-v0 β EleutherAI/pythia-1.4b-deduped-v0}/config.json +0 -0
- lens/{pythia-1.4b-deduped-v0 β EleutherAI/pythia-1.4b-deduped-v0}/params.pt +0 -0
- lens/{pythia-12b-deduped-v0 β EleutherAI/pythia-12b-deduped-v0}/config.json +0 -0
- lens/{pythia-12b-deduped-v0 β EleutherAI/pythia-12b-deduped-v0}/params.pt +0 -0
- lens/{pythia-160m-deduped-v0 β EleutherAI/pythia-160m-deduped-v0}/config.json +0 -0
- lens/{pythia-160m-deduped-v0 β EleutherAI/pythia-160m-deduped-v0}/params.pt +0 -0
- lens/{pythia-1b-deduped-v0 β EleutherAI/pythia-1b-deduped-v0}/config.json +0 -0
- lens/{pythia-1b-deduped-v0 β EleutherAI/pythia-1b-deduped-v0}/params.pt +0 -0
- lens/{pythia-410m-deduped-v0 β EleutherAI/pythia-410m-deduped-v0}/config.json +0 -0
- lens/{pythia-410m-deduped-v0 β EleutherAI/pythia-410m-deduped-v0}/params.pt +0 -0
- lens/{pythia-6.9b-deduped-v0 β EleutherAI/pythia-6.9b-deduped-v0}/config.json +0 -0
- lens/{pythia-6.9b-deduped-v0 β EleutherAI/pythia-6.9b-deduped-v0}/params.pt +0 -0
- lens/{pythia-70m-deduped-v0 β EleutherAI/pythia-70m-deduped-v0}/config.json +0 -0
- lens/{pythia-70m-deduped-v0 β EleutherAI/pythia-70m-deduped-v0}/params.pt +0 -0
- lens/{opt-1.3b β facebook/opt-1.3b}/config.json +0 -0
- lens/{opt-1.3b β facebook/opt-1.3b}/params.pt +0 -0
- lens/{opt-125m β facebook/opt-125m}/config.json +0 -0
- lens/{opt-125m β facebook/opt-125m}/params.pt +0 -0
- lens/{opt-6.7b β facebook/opt-6.7b}/config.json +0 -0
- lens/{opt-6.7b β facebook/opt-6.7b}/params.pt +0 -0
- lens_migration.py +0 -384
- migrate.sh +0 -12
lens/{gpt-neox-20b β EleutherAI/gpt-neox-20b}/config.json
RENAMED
File without changes
|
lens/{gpt-neox-20b β EleutherAI/gpt-neox-20b}/params.pt
RENAMED
File without changes
|
lens/{pythia-1.4b-deduped-v0 β EleutherAI/pythia-1.4b-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-1.4b-deduped-v0 β EleutherAI/pythia-1.4b-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-12b-deduped-v0 β EleutherAI/pythia-12b-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-12b-deduped-v0 β EleutherAI/pythia-12b-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-160m-deduped-v0 β EleutherAI/pythia-160m-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-160m-deduped-v0 β EleutherAI/pythia-160m-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-1b-deduped-v0 β EleutherAI/pythia-1b-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-1b-deduped-v0 β EleutherAI/pythia-1b-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-410m-deduped-v0 β EleutherAI/pythia-410m-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-410m-deduped-v0 β EleutherAI/pythia-410m-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-6.9b-deduped-v0 β EleutherAI/pythia-6.9b-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-6.9b-deduped-v0 β EleutherAI/pythia-6.9b-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{pythia-70m-deduped-v0 β EleutherAI/pythia-70m-deduped-v0}/config.json
RENAMED
File without changes
|
lens/{pythia-70m-deduped-v0 β EleutherAI/pythia-70m-deduped-v0}/params.pt
RENAMED
File without changes
|
lens/{opt-1.3b β facebook/opt-1.3b}/config.json
RENAMED
File without changes
|
lens/{opt-1.3b β facebook/opt-1.3b}/params.pt
RENAMED
File without changes
|
lens/{opt-125m β facebook/opt-125m}/config.json
RENAMED
File without changes
|
lens/{opt-125m β facebook/opt-125m}/params.pt
RENAMED
File without changes
|
lens/{opt-6.7b β facebook/opt-6.7b}/config.json
RENAMED
File without changes
|
lens/{opt-6.7b β facebook/opt-6.7b}/params.pt
RENAMED
File without changes
|
lens_migration.py
DELETED
@@ -1,384 +0,0 @@
|
|
1 |
-
#!/usr/bin/env python3
|
2 |
-
from huggingface_hub import model_info
|
3 |
-
import argparse
|
4 |
-
from copy import deepcopy
|
5 |
-
import inspect
|
6 |
-
from logging import warn
|
7 |
-
from pathlib import Path
|
8 |
-
from tqdm import tqdm
|
9 |
-
import json
|
10 |
-
|
11 |
-
from tuned_lens.model_surgery import get_final_norm, get_transformer_layers
|
12 |
-
from tuned_lens.load_artifacts import load_lens_artifacts
|
13 |
-
from tuned_lens.nn import TunedLens
|
14 |
-
from transformers.models.bloom.modeling_bloom import BloomBlock
|
15 |
-
from transformers import PreTrainedModel, AutoModelForCausalLM
|
16 |
-
from typing import Optional, Generator, Union
|
17 |
-
import torch as th
|
18 |
-
|
19 |
-
from tuned_lens.stats.distance import js_divergence
|
20 |
-
|
21 |
-
|
22 |
-
def instantiate_layer(model_config, layer_idx: int, model_type: str) -> th.nn.Module:
|
23 |
-
if model_type == "bloom":
|
24 |
-
from transformers.models.bloom.modeling_bloom import BloomBlock
|
25 |
-
|
26 |
-
return _BloomBlockWrapper(BloomBlock(model_config)) # type: ignore[arg-type]
|
27 |
-
if model_type == "gpt_neo":
|
28 |
-
from transformers.models.gpt_neo.modeling_gpt_neo import GPTNeoBlock
|
29 |
-
|
30 |
-
return GPTNeoBlock(model_config, layer_idx)
|
31 |
-
if model_type == "gpt_neox":
|
32 |
-
from transformers.models.gpt_neox.modeling_gpt_neox import (
|
33 |
-
GPTNeoXLayer,
|
34 |
-
)
|
35 |
-
|
36 |
-
return GPTNeoXLayer(model_config) # type: ignore[arg-type]
|
37 |
-
if model_type == "gpt2":
|
38 |
-
from transformers.models.gpt2.modeling_gpt2 import GPT2Block
|
39 |
-
|
40 |
-
return GPT2Block(model_config, layer_idx) # type: ignore[arg-type]
|
41 |
-
if model_type == "opt":
|
42 |
-
from transformers.models.opt.modeling_opt import OPTDecoderLayer
|
43 |
-
|
44 |
-
return OPTDecoderLayer(model_config) # type: ignore[arg-type]
|
45 |
-
else:
|
46 |
-
raise ValueError(f"Unknown model type '{model_type}'")
|
47 |
-
|
48 |
-
|
49 |
-
def maybe_wrap(layer: th.nn.Module) -> th.nn.Module:
|
50 |
-
return _BloomBlockWrapper(layer) if isinstance(layer, BloomBlock) else layer
|
51 |
-
|
52 |
-
|
53 |
-
# Very annoying that we have to do this. See https://bit.ly/3XSQ7W6 for context on
|
54 |
-
# what we're doing here.
|
55 |
-
class _BloomBlockWrapper(th.nn.Module):
|
56 |
-
def __init__(self, block: BloomBlock):
|
57 |
-
super().__init__()
|
58 |
-
self.block = block
|
59 |
-
|
60 |
-
def forward(self, x: th.Tensor) -> th.Tensor:
|
61 |
-
from transformers.models.bloom.modeling_bloom import (
|
62 |
-
BloomModel,
|
63 |
-
build_alibi_tensor,
|
64 |
-
)
|
65 |
-
|
66 |
-
batch_size, seq_len, _ = x.shape
|
67 |
-
dummy_mask = x.new_ones([batch_size, seq_len])
|
68 |
-
|
69 |
-
# Causal mask isn't created inside the block itself, so we have to do it here.
|
70 |
-
# Weirdly _prepare_attn_mask doesn't depend on `self` at all but is still an
|
71 |
-
# instance method for some reason, so we pass `None` as the first argument.
|
72 |
-
causal_mask = BloomModel._prepare_attn_mask(
|
73 |
-
None, dummy_mask, (batch_size, seq_len), 0 # type: ignore[arg-type]
|
74 |
-
)
|
75 |
-
alibi = build_alibi_tensor(dummy_mask, self.block.num_heads, x.dtype)
|
76 |
-
h, *_ = self.block(x, alibi, causal_mask)
|
77 |
-
return h
|
78 |
-
|
79 |
-
|
80 |
-
class TunedLensOld(th.nn.Module):
|
81 |
-
"""A tuned lens for decoding hidden states into logits."""
|
82 |
-
|
83 |
-
layer_norm: th.nn.LayerNorm
|
84 |
-
unembedding: th.nn.Linear
|
85 |
-
extra_layers: th.nn.Sequential
|
86 |
-
layer_translators: th.nn.ModuleList
|
87 |
-
|
88 |
-
def __init__(
|
89 |
-
self,
|
90 |
-
model: Optional[PreTrainedModel] = None,
|
91 |
-
*,
|
92 |
-
bias: bool = True,
|
93 |
-
extra_layers: int = 0,
|
94 |
-
include_input: bool = True,
|
95 |
-
reuse_unembedding: bool = True,
|
96 |
-
# Used when saving and loading the lens
|
97 |
-
model_config: Optional[dict] = None,
|
98 |
-
d_model: Optional[int] = None,
|
99 |
-
num_layers: Optional[int] = None,
|
100 |
-
vocab_size: Optional[int] = None,
|
101 |
-
):
|
102 |
-
"""Create a TunedLensOld.
|
103 |
-
|
104 |
-
Args:
|
105 |
-
model : A pertained model from the transformers library you wish to inspect.
|
106 |
-
bias : Whether to include a bias term in the translator layers.
|
107 |
-
extra_layers : The number of extra layers to apply to the hidden states
|
108 |
-
before decoding into logits.
|
109 |
-
|
110 |
-
include_input : Whether to include a lens that decodes the word embeddings.
|
111 |
-
reuse_unembedding : Weather to reuse the unembedding matrix from the model.
|
112 |
-
model_config : The config of the model. Used for saving and loading.
|
113 |
-
d_model : The models hidden size. Used for saving and loading.
|
114 |
-
num_layers : The number of layers in the model. Used for saving and loading.
|
115 |
-
vocab_size : The size of the vocabulary. Used for saving and loading.
|
116 |
-
|
117 |
-
Raises:
|
118 |
-
ValueError: if neither a model or d_model, num_layers, and vocab_size,
|
119 |
-
are provided.
|
120 |
-
"""
|
121 |
-
super().__init__()
|
122 |
-
|
123 |
-
self.extra_layers = th.nn.Sequential()
|
124 |
-
|
125 |
-
if (
|
126 |
-
model
|
127 |
-
is None
|
128 |
-
== (d_model is None or num_layers is None or vocab_size is None)
|
129 |
-
):
|
130 |
-
raise ValueError(
|
131 |
-
"Must provide either a model or d_model, num_layers, and vocab_size"
|
132 |
-
)
|
133 |
-
|
134 |
-
# Initializing from scratch without a model
|
135 |
-
if not model:
|
136 |
-
assert d_model and num_layers and vocab_size
|
137 |
-
self.layer_norm = th.nn.LayerNorm(d_model)
|
138 |
-
self.unembedding = th.nn.Linear(d_model, vocab_size, bias=False)
|
139 |
-
|
140 |
-
# Use HuggingFace methods to get decoder layers
|
141 |
-
else:
|
142 |
-
assert not (d_model or num_layers or vocab_size)
|
143 |
-
d_model = model.config.hidden_size
|
144 |
-
num_layers = model.config.num_hidden_layers
|
145 |
-
vocab_size = model.config.vocab_size
|
146 |
-
assert isinstance(d_model, int) and isinstance(vocab_size, int)
|
147 |
-
|
148 |
-
model_config = model.config.to_dict() # type: ignore[F841]
|
149 |
-
|
150 |
-
# Currently we convert the decoder to full precision
|
151 |
-
self.unembedding = deepcopy(model.get_output_embeddings()).float()
|
152 |
-
if ln := get_final_norm(model):
|
153 |
-
self.layer_norm = deepcopy(ln).float()
|
154 |
-
else:
|
155 |
-
self.layer_norm = th.nn.Identity()
|
156 |
-
|
157 |
-
if extra_layers:
|
158 |
-
_, layers = get_transformer_layers(model)
|
159 |
-
self.extra_layers.extend(
|
160 |
-
[maybe_wrap(layer) for layer in layers[-extra_layers:]]
|
161 |
-
)
|
162 |
-
|
163 |
-
# Save config for later
|
164 |
-
config_keys = set(inspect.getfullargspec(TunedLensOld).kwonlyargs)
|
165 |
-
self.config = {k: v for k, v in locals().items() if k in config_keys}
|
166 |
-
del model_config
|
167 |
-
|
168 |
-
# Try to prevent finetuning the decoder
|
169 |
-
assert d_model and num_layers
|
170 |
-
self.layer_norm.requires_grad_(False)
|
171 |
-
self.unembedding.requires_grad_(False)
|
172 |
-
|
173 |
-
out_features = d_model if reuse_unembedding else vocab_size
|
174 |
-
translator = th.nn.Linear(d_model, out_features, bias=bias)
|
175 |
-
if not reuse_unembedding:
|
176 |
-
translator.weight.data = self.unembedding.weight.data.clone()
|
177 |
-
translator.bias.data.zero_()
|
178 |
-
else:
|
179 |
-
translator.weight.data.zero_()
|
180 |
-
translator.bias.data.zero_()
|
181 |
-
|
182 |
-
self.add_module("input_translator", translator if include_input else None)
|
183 |
-
# Don't include the final layer
|
184 |
-
num_layers -= 1
|
185 |
-
|
186 |
-
self.layer_translators = th.nn.ModuleList(
|
187 |
-
[deepcopy(translator) for _ in range(num_layers)]
|
188 |
-
)
|
189 |
-
|
190 |
-
def __getitem__(self, item: int) -> th.nn.Module:
|
191 |
-
"""Get the probe module at the given index."""
|
192 |
-
if isinstance(self.input_translator, th.nn.Module):
|
193 |
-
if item == 0:
|
194 |
-
return self.input_translator
|
195 |
-
else:
|
196 |
-
item -= 1
|
197 |
-
|
198 |
-
return self.layer_translators[item]
|
199 |
-
|
200 |
-
def __iter__(self) -> Generator[th.nn.Module, None, None]:
|
201 |
-
"""Get iterator over the translators within the lens."""
|
202 |
-
if isinstance(self.input_translator, th.nn.Module):
|
203 |
-
yield self.input_translator
|
204 |
-
|
205 |
-
yield from self.layer_translators
|
206 |
-
|
207 |
-
@classmethod
|
208 |
-
def load(cls, resource_id: str, **kwargs) -> "TunedLensOld":
|
209 |
-
"""Load a tuned lens from a or hugging face hub.
|
210 |
-
|
211 |
-
Args:
|
212 |
-
resource_id : The path to the directory containing the config and checkpoint
|
213 |
-
or the name of the model on the hugging face hub.
|
214 |
-
**kwargs : Additional arguments to pass to torch.load.
|
215 |
-
|
216 |
-
Returns:
|
217 |
-
A TunedLensOld instance.
|
218 |
-
"""
|
219 |
-
config_path, ckpt_path = load_lens_artifacts(resource_id)
|
220 |
-
# Load config
|
221 |
-
with open(config_path, "r") as f:
|
222 |
-
config = json.load(f)
|
223 |
-
|
224 |
-
# Load parameters
|
225 |
-
state = th.load(ckpt_path, **kwargs)
|
226 |
-
|
227 |
-
# Backwards compatibility we really need to stop renaming things
|
228 |
-
keys = list(state.keys())
|
229 |
-
for key in keys:
|
230 |
-
for old_key in ["probe", "adapter"]:
|
231 |
-
if old_key in key:
|
232 |
-
warn(
|
233 |
-
f"Loading a checkpoint with a '{old_key}' key. "
|
234 |
-
"This is deprecated and may be removed in a future version. "
|
235 |
-
)
|
236 |
-
new_key = key.replace(old_key, "translator")
|
237 |
-
state[new_key] = state.pop(key)
|
238 |
-
|
239 |
-
# Drop unrecognized config keys
|
240 |
-
unrecognized = set(config) - set(inspect.getfullargspec(cls).kwonlyargs)
|
241 |
-
for key in unrecognized:
|
242 |
-
warn(f"Ignoring config key '{key}'")
|
243 |
-
del config[key]
|
244 |
-
|
245 |
-
lens = cls(**config)
|
246 |
-
|
247 |
-
if num_extras := config.get("extra_layers"):
|
248 |
-
# This is sort of a hack but AutoConfig doesn't appear to have a from_dict
|
249 |
-
# for some reason.
|
250 |
-
from transformers.models.auto import CONFIG_MAPPING
|
251 |
-
|
252 |
-
model_conf_dict = config.get("model_config")
|
253 |
-
del model_conf_dict["torch_dtype"]
|
254 |
-
assert model_conf_dict, "Need a 'model_config' entry to load extra layers"
|
255 |
-
|
256 |
-
model_type = model_conf_dict["model_type"]
|
257 |
-
config_cls = CONFIG_MAPPING[model_type]
|
258 |
-
model_config = config_cls.from_dict(model_conf_dict)
|
259 |
-
|
260 |
-
lens.extra_layers = th.nn.Sequential(
|
261 |
-
*[
|
262 |
-
instantiate_layer(
|
263 |
-
model_config, model_config.num_hidden_layers - i - 1, model_type
|
264 |
-
)
|
265 |
-
for i in range(num_extras)
|
266 |
-
]
|
267 |
-
)
|
268 |
-
|
269 |
-
lens.load_state_dict(state)
|
270 |
-
return lens
|
271 |
-
|
272 |
-
def save(
|
273 |
-
self,
|
274 |
-
path: Union[Path, str],
|
275 |
-
ckpt: str = "params.pt",
|
276 |
-
config: str = "config.json",
|
277 |
-
) -> None:
|
278 |
-
"""Save the lens to a directory.
|
279 |
-
|
280 |
-
Args:
|
281 |
-
path : The path to the directory to save the lens to.
|
282 |
-
ckpt : The name of the checkpoint file to save the parameters to.
|
283 |
-
config : The name of the config file to save the config to.
|
284 |
-
"""
|
285 |
-
path = Path(path)
|
286 |
-
path.mkdir(exist_ok=True, parents=True)
|
287 |
-
th.save(self.state_dict(), path / ckpt)
|
288 |
-
|
289 |
-
with open(path / config, "w") as f:
|
290 |
-
json.dump(self.config, f)
|
291 |
-
|
292 |
-
def normalize_(self):
|
293 |
-
"""Canonicalize the transforms by centering their weights and biases."""
|
294 |
-
for linear in self:
|
295 |
-
assert isinstance(linear, th.nn.Linear)
|
296 |
-
|
297 |
-
A, b = linear.weight.data, linear.bias.data
|
298 |
-
A -= A.mean(dim=0, keepdim=True)
|
299 |
-
b -= b.mean()
|
300 |
-
|
301 |
-
def transform_hidden(self, h: th.Tensor, idx: int) -> th.Tensor:
|
302 |
-
"""Transform hidden state from layer `idx`."""
|
303 |
-
if not self.config["reuse_unembedding"]:
|
304 |
-
raise RuntimeError("TunedLensOld.transform_hidden requires reuse_unembedding")
|
305 |
-
|
306 |
-
# Note that we add the translator output residually, in contrast to the formula
|
307 |
-
# in the paper. By parametrizing it this way we ensure that weight decay
|
308 |
-
# regularizes the transform toward the identity, not the zero transformation.
|
309 |
-
return h + self[idx](h)
|
310 |
-
|
311 |
-
def to_logits(self, h: th.Tensor) -> th.Tensor:
|
312 |
-
"""Decode a hidden state into logits."""
|
313 |
-
h = self.extra_layers(h)
|
314 |
-
while isinstance(h, tuple):
|
315 |
-
h, *_ = h
|
316 |
-
|
317 |
-
return self.unembedding(self.layer_norm(h))
|
318 |
-
|
319 |
-
def forward(self, h: th.Tensor, idx: int) -> th.Tensor:
|
320 |
-
"""Transform and then decode the hidden states into logits."""
|
321 |
-
# Sanity check to make sure we don't finetune the decoder
|
322 |
-
# if any(p.requires_grad for p in self.parameters(recurse=False)):
|
323 |
-
# raise RuntimeError("Make sure to freeze the decoder")
|
324 |
-
|
325 |
-
# We're learning a separate unembedding for each layer
|
326 |
-
if not self.config["reuse_unembedding"]:
|
327 |
-
h_ = self.layer_norm(h)
|
328 |
-
return self[idx](h_)
|
329 |
-
|
330 |
-
h = self.transform_hidden(h, idx)
|
331 |
-
return self.to_logits(h)
|
332 |
-
|
333 |
-
def __len__(self) -> int:
|
334 |
-
"""Return the number of layer translators in the lens."""
|
335 |
-
N = len(self.layer_translators)
|
336 |
-
if self.input_translator:
|
337 |
-
N += 1
|
338 |
-
|
339 |
-
return N
|
340 |
-
|
341 |
-
|
342 |
-
if __name__ == "__main__":
|
343 |
-
parser = argparse.ArgumentParser()
|
344 |
-
parser.add_argument("--model", type=str, default="gpt2")
|
345 |
-
parser.add_argument("--resource-id", type=str, default="gpt2")
|
346 |
-
parser.add_argument("--output-dir", type=str, default="lens/gpt2")
|
347 |
-
args = parser.parse_args()
|
348 |
-
|
349 |
-
model = AutoModelForCausalLM.from_pretrained(args.model)
|
350 |
-
revision = model_info(args.model).sha
|
351 |
-
model.eval()
|
352 |
-
model.requires_grad_(False)
|
353 |
-
|
354 |
-
device = th.device("cuda:0" if th.cuda.is_available() else "cpu")
|
355 |
-
|
356 |
-
print("Loading old lens")
|
357 |
-
tuned_lens_old = TunedLensOld.load(args.resource_id, map_location=device)
|
358 |
-
|
359 |
-
print("Initializing new lens")
|
360 |
-
tuned_lens = TunedLens.from_model(
|
361 |
-
model, bias=tuned_lens_old.config['bias'], revision=revision
|
362 |
-
)
|
363 |
-
|
364 |
-
for i in tqdm(range(len(tuned_lens_old)), desc="Copying parameters"):
|
365 |
-
tuned_lens[i].load_state_dict(tuned_lens_old[i].state_dict())
|
366 |
-
|
367 |
-
|
368 |
-
tuned_lens = tuned_lens.to(device)
|
369 |
-
tuned_lens_old = tuned_lens_old.to(device)
|
370 |
-
model = model.to(device)
|
371 |
-
|
372 |
-
# Fuzz the new lens against the old one's
|
373 |
-
with th.no_grad():
|
374 |
-
for i in tqdm(range(len(tuned_lens)), desc="Fuzzing layers"):
|
375 |
-
for _ in range(10):
|
376 |
-
a = th.randn(1, 1, tuned_lens.config.d_model, device=device)
|
377 |
-
logits_new = tuned_lens(a, i)
|
378 |
-
logits_old = tuned_lens_old(a, i)
|
379 |
-
log_ps_new = logits_new.log_softmax(-1)
|
380 |
-
log_ps_old = logits_old.log_softmax(-1)
|
381 |
-
print("js div", js_divergence(log_ps_new, log_ps_old))
|
382 |
-
assert (th.allclose(log_ps_new, log_ps_old, atol=1e-4)), (log_ps_new - log_ps_old).abs().max()
|
383 |
-
print("Saving new lens to", args.output_dir)
|
384 |
-
tuned_lens.to(th.device("cpu")).save(args.output_dir)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
migrate.sh
DELETED
@@ -1,12 +0,0 @@
|
|
1 |
-
#!/bin/bash
|
2 |
-
|
3 |
-
set -e
|
4 |
-
|
5 |
-
for i in pythia-70m-deduped-v0,EleutherAI/pythia-70m-deduped-v0
|
6 |
-
do
|
7 |
-
IFS=","
|
8 |
-
set -- $i
|
9 |
-
echo "migrating $2"
|
10 |
-
CUDA_VISIBLE_DEVICES=-1 python3 lens_migration.py --model $2 --resource-id $1 --output lens/$1
|
11 |
-
git commit -am "$1 migrated"
|
12 |
-
done
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|