File size: 10,439 Bytes
302920f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 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 |
# Copyright 2023-present the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import tempfile
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import (
AutoPeftModel,
AutoPeftModelForCausalLM,
AutoPeftModelForFeatureExtraction,
AutoPeftModelForQuestionAnswering,
AutoPeftModelForSeq2SeqLM,
AutoPeftModelForSequenceClassification,
AutoPeftModelForTokenClassification,
LoraConfig,
PeftModel,
PeftModelForCausalLM,
PeftModelForFeatureExtraction,
PeftModelForQuestionAnswering,
PeftModelForSeq2SeqLM,
PeftModelForSequenceClassification,
PeftModelForTokenClassification,
get_peft_model,
)
from peft.utils import infer_device
class TestPeftAutoModel:
dtype = torch.float16 if infer_device() == "mps" else torch.bfloat16
def test_peft_causal_lm(self):
model_id = "peft-internal-testing/tiny-OPTForCausalLM-lora"
model = AutoPeftModelForCausalLM.from_pretrained(model_id)
assert isinstance(model, PeftModelForCausalLM)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForCausalLM.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForCausalLM)
# check if kwargs are passed correctly
model = AutoPeftModelForCausalLM.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForCausalLM)
assert model.base_model.lm_head.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForCausalLM.from_pretrained(model_id, adapter_name, is_trainable, torch_dtype=self.dtype)
def test_peft_causal_lm_extended_vocab(self):
model_id = "peft-internal-testing/tiny-random-OPTForCausalLM-extended-vocab"
model = AutoPeftModelForCausalLM.from_pretrained(model_id)
assert isinstance(model, PeftModelForCausalLM)
# check if kwargs are passed correctly
model = AutoPeftModelForCausalLM.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForCausalLM)
assert model.base_model.lm_head.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForCausalLM.from_pretrained(model_id, adapter_name, is_trainable, torch_dtype=self.dtype)
def test_peft_seq2seq_lm(self):
model_id = "peft-internal-testing/tiny_T5ForSeq2SeqLM-lora"
model = AutoPeftModelForSeq2SeqLM.from_pretrained(model_id)
assert isinstance(model, PeftModelForSeq2SeqLM)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForSeq2SeqLM.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForSeq2SeqLM)
# check if kwargs are passed correctly
model = AutoPeftModelForSeq2SeqLM.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForSeq2SeqLM)
assert model.base_model.lm_head.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForSeq2SeqLM.from_pretrained(model_id, adapter_name, is_trainable, torch_dtype=self.dtype)
def test_peft_sequence_cls(self):
model_id = "peft-internal-testing/tiny_OPTForSequenceClassification-lora"
model = AutoPeftModelForSequenceClassification.from_pretrained(model_id)
assert isinstance(model, PeftModelForSequenceClassification)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForSequenceClassification.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForSequenceClassification)
# check if kwargs are passed correctly
model = AutoPeftModelForSequenceClassification.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForSequenceClassification)
assert model.score.original_module.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForSequenceClassification.from_pretrained(
model_id, adapter_name, is_trainable, torch_dtype=self.dtype
)
def test_peft_token_classification(self):
model_id = "peft-internal-testing/tiny_GPT2ForTokenClassification-lora"
model = AutoPeftModelForTokenClassification.from_pretrained(model_id)
assert isinstance(model, PeftModelForTokenClassification)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForTokenClassification.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForTokenClassification)
# check if kwargs are passed correctly
model = AutoPeftModelForTokenClassification.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForTokenClassification)
assert model.base_model.classifier.original_module.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForTokenClassification.from_pretrained(
model_id, adapter_name, is_trainable, torch_dtype=self.dtype
)
def test_peft_question_answering(self):
model_id = "peft-internal-testing/tiny_OPTForQuestionAnswering-lora"
model = AutoPeftModelForQuestionAnswering.from_pretrained(model_id)
assert isinstance(model, PeftModelForQuestionAnswering)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForQuestionAnswering.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForQuestionAnswering)
# check if kwargs are passed correctly
model = AutoPeftModelForQuestionAnswering.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForQuestionAnswering)
assert model.base_model.qa_outputs.original_module.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForQuestionAnswering.from_pretrained(
model_id, adapter_name, is_trainable, torch_dtype=self.dtype
)
def test_peft_feature_extraction(self):
model_id = "peft-internal-testing/tiny_OPTForFeatureExtraction-lora"
model = AutoPeftModelForFeatureExtraction.from_pretrained(model_id)
assert isinstance(model, PeftModelForFeatureExtraction)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModelForFeatureExtraction.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModelForFeatureExtraction)
# check if kwargs are passed correctly
model = AutoPeftModelForFeatureExtraction.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModelForFeatureExtraction)
assert model.base_model.model.decoder.embed_tokens.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModelForFeatureExtraction.from_pretrained(
model_id, adapter_name, is_trainable, torch_dtype=self.dtype
)
def test_peft_whisper(self):
model_id = "peft-internal-testing/tiny_WhisperForConditionalGeneration-lora"
model = AutoPeftModel.from_pretrained(model_id)
assert isinstance(model, PeftModel)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model = AutoPeftModel.from_pretrained(tmp_dirname)
assert isinstance(model, PeftModel)
# check if kwargs are passed correctly
model = AutoPeftModel.from_pretrained(model_id, torch_dtype=self.dtype)
assert isinstance(model, PeftModel)
assert model.base_model.model.model.encoder.embed_positions.weight.dtype == self.dtype
adapter_name = "default"
is_trainable = False
# This should work
_ = AutoPeftModel.from_pretrained(model_id, adapter_name, is_trainable, torch_dtype=self.dtype)
def test_embedding_size_not_reduced_if_greater_vocab_size(self, tmp_path):
# See 2415
# There was a bug in AutoPeftModels where the embedding was always resized to the vocab size of the tokenizer
# when the tokenizer was found. This makes sense if the vocabulary was extended, but some models like Qwen
# already start out with "spare" embeddings, i.e. the embedding size is larger than the vocab size. This could
# result in the embedding being shrunk, which in turn resulted in an error when loading the weights.
# first create a checkpoint; it is important that the tokenizer is also saved in the same location
model_id = "Qwen/Qwen2-0.5B"
model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = get_peft_model(model, LoraConfig(modules_to_save=["lm_head", "embed_token"]))
model.save_pretrained(tmp_path)
tokenizer.save_pretrained(tmp_path)
# does not raise; without the fix, it raises:
# > size mismatch for base_model.model.lm_head.modules_to_save.default.weight: copying a param with shape
# torch.Size([151936, 896]) from checkpoint, the shape in current model is torch.Size([151646, 896]).
AutoPeftModelForCausalLM.from_pretrained(tmp_path)
|