load the tokenizer seperately from the model
Browse files- scripts/finetune.py +21 -12
- src/axolotl/utils/models.py +41 -42
scripts/finetune.py
CHANGED
@@ -21,7 +21,7 @@ src_dir = os.path.join(project_root, "src")
|
|
21 |
sys.path.insert(0, src_dir)
|
22 |
|
23 |
from axolotl.utils.data import load_prepare_datasets
|
24 |
-
from axolotl.utils.models import load_model
|
25 |
from axolotl.utils.trainer import setup_trainer
|
26 |
from axolotl.utils.wandb import setup_wandb_env_vars
|
27 |
|
@@ -161,13 +161,30 @@ def train(
|
|
161 |
|
162 |
validate_config(cfg)
|
163 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
# Load the model and tokenizer
|
165 |
-
logging.info("loading model
|
166 |
-
model,
|
167 |
cfg.base_model,
|
168 |
cfg.base_model_config,
|
169 |
cfg.model_type,
|
170 |
-
|
171 |
cfg,
|
172 |
adapter=cfg.adapter,
|
173 |
inference=("inference" in kwargs),
|
@@ -192,10 +209,6 @@ def train(
|
|
192 |
model.save_pretrained(cfg.output_dir)
|
193 |
return
|
194 |
|
195 |
-
train_dataset, eval_dataset = load_prepare_datasets(
|
196 |
-
tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
|
197 |
-
)
|
198 |
-
|
199 |
if cfg.debug:
|
200 |
logging.info("check_dataset_labels...")
|
201 |
check_dataset_labels(
|
@@ -205,10 +218,6 @@ def train(
|
|
205 |
tokenizer,
|
206 |
)
|
207 |
|
208 |
-
if prepare_ds_only:
|
209 |
-
logging.info("Finished preparing dataset. Exiting...")
|
210 |
-
return
|
211 |
-
|
212 |
trainer = setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer)
|
213 |
|
214 |
model.config.use_cache = False
|
|
|
21 |
sys.path.insert(0, src_dir)
|
22 |
|
23 |
from axolotl.utils.data import load_prepare_datasets
|
24 |
+
from axolotl.utils.models import load_model, load_tokenizer
|
25 |
from axolotl.utils.trainer import setup_trainer
|
26 |
from axolotl.utils.wandb import setup_wandb_env_vars
|
27 |
|
|
|
161 |
|
162 |
validate_config(cfg)
|
163 |
|
164 |
+
# load the tokenizer first
|
165 |
+
logging.info("loading tokenizer...")
|
166 |
+
tokenizer = load_tokenizer(
|
167 |
+
cfg.base_model_config,
|
168 |
+
cfg.tokenizer_type,
|
169 |
+
cfg
|
170 |
+
)
|
171 |
+
|
172 |
+
if "inference" not in kwargs and "shard" not in kwargs: # don't need to load dataset for these
|
173 |
+
train_dataset, eval_dataset = load_prepare_datasets(
|
174 |
+
tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
|
175 |
+
)
|
176 |
+
|
177 |
+
if prepare_ds_only:
|
178 |
+
logging.info("Finished preparing dataset. Exiting...")
|
179 |
+
return
|
180 |
+
|
181 |
# Load the model and tokenizer
|
182 |
+
logging.info("loading model and peft_config...")
|
183 |
+
model, peft_config = load_model(
|
184 |
cfg.base_model,
|
185 |
cfg.base_model_config,
|
186 |
cfg.model_type,
|
187 |
+
tokenizer,
|
188 |
cfg,
|
189 |
adapter=cfg.adapter,
|
190 |
inference=("inference" in kwargs),
|
|
|
209 |
model.save_pretrained(cfg.output_dir)
|
210 |
return
|
211 |
|
|
|
|
|
|
|
|
|
212 |
if cfg.debug:
|
213 |
logging.info("check_dataset_labels...")
|
214 |
check_dataset_labels(
|
|
|
218 |
tokenizer,
|
219 |
)
|
220 |
|
|
|
|
|
|
|
|
|
221 |
trainer = setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer)
|
222 |
|
223 |
model.config.use_cache = False
|
src/axolotl/utils/models.py
CHANGED
@@ -7,7 +7,6 @@ from typing import Optional, Tuple, TYPE_CHECKING
|
|
7 |
import bitsandbytes as bnb
|
8 |
import torch
|
9 |
import transformers
|
10 |
-
from torch import nn
|
11 |
from transformers import (
|
12 |
AutoModelForCausalLM,
|
13 |
AutoTokenizer,
|
@@ -34,20 +33,56 @@ if TYPE_CHECKING:
|
|
34 |
from transformers import PreTrainedTokenizer
|
35 |
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
def load_model(
|
38 |
base_model,
|
39 |
base_model_config,
|
40 |
model_type,
|
41 |
-
|
42 |
cfg,
|
43 |
adapter="lora",
|
44 |
inference=False,
|
45 |
):
|
46 |
-
# type: (str, str, str, str, AttrDefault, Optional[str], bool) -> Tuple[PreTrainedModel,
|
47 |
|
48 |
# TODO refactor as a kwarg
|
49 |
load_in_8bit = cfg.load_in_8bit
|
50 |
-
tokenizer = None
|
51 |
is_llama_derived_model = "llama" in base_model or (
|
52 |
cfg.model_type and "llama" in cfg.model_type.lower()
|
53 |
)
|
@@ -122,7 +157,7 @@ def load_model(
|
|
122 |
model_path = str(cache_model_path)
|
123 |
except:
|
124 |
model_path = cfg.base_model
|
125 |
-
model,
|
126 |
base_model_config if base_model_config else base_model,
|
127 |
model_path,
|
128 |
device_map=cfg.device_map,
|
@@ -207,42 +242,6 @@ def load_model(
|
|
207 |
**model_kwargs,
|
208 |
)
|
209 |
|
210 |
-
if not tokenizer:
|
211 |
-
try:
|
212 |
-
if is_llama_derived_model and "LlamaTokenizer" in globals():
|
213 |
-
tokenizer = LlamaTokenizer.from_pretrained(
|
214 |
-
base_model_config,
|
215 |
-
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
216 |
-
)
|
217 |
-
else:
|
218 |
-
tokenizer = getattr(transformers, tokenizer_type).from_pretrained(
|
219 |
-
base_model_config,
|
220 |
-
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
221 |
-
)
|
222 |
-
except:
|
223 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
224 |
-
base_model_config,
|
225 |
-
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
226 |
-
)
|
227 |
-
|
228 |
-
logging.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
|
229 |
-
logging.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
|
230 |
-
logging.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
|
231 |
-
logging.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
|
232 |
-
|
233 |
-
if tokenizer.__class__.__name__ in ["LlamaTokenizer", "LlamaTokenizerFast"]:
|
234 |
-
tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
|
235 |
-
|
236 |
-
if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
|
237 |
-
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
|
238 |
-
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
239 |
-
|
240 |
-
if cfg.special_tokens:
|
241 |
-
for k, v in cfg.special_tokens.items():
|
242 |
-
tokenizer.add_special_tokens({k: v})
|
243 |
-
if cfg.tokens:
|
244 |
-
tokenizer.add_tokens(list(cfg.tokens))
|
245 |
-
|
246 |
embeddings_len = math.ceil(len(tokenizer) / 32) * 32
|
247 |
model.resize_token_embeddings(embeddings_len)
|
248 |
|
@@ -291,7 +290,7 @@ def load_model(
|
|
291 |
model.config.use_cache = False
|
292 |
|
293 |
# TODO resume_from_checkpoint handling
|
294 |
-
return model,
|
295 |
|
296 |
|
297 |
def load_adapter(model, cfg, adapter):
|
|
|
7 |
import bitsandbytes as bnb
|
8 |
import torch
|
9 |
import transformers
|
|
|
10 |
from transformers import (
|
11 |
AutoModelForCausalLM,
|
12 |
AutoTokenizer,
|
|
|
33 |
from transformers import PreTrainedTokenizer
|
34 |
|
35 |
|
36 |
+
def load_tokenizer(
|
37 |
+
base_model_config,
|
38 |
+
tokenizer_type,
|
39 |
+
cfg,
|
40 |
+
):
|
41 |
+
if tokenizer_type:
|
42 |
+
tokenizer = getattr(transformers, tokenizer_type).from_pretrained(
|
43 |
+
base_model_config,
|
44 |
+
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
45 |
+
)
|
46 |
+
else:
|
47 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
48 |
+
base_model_config,
|
49 |
+
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
50 |
+
)
|
51 |
+
|
52 |
+
logging.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
|
53 |
+
logging.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
|
54 |
+
logging.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
|
55 |
+
logging.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
|
56 |
+
|
57 |
+
if tokenizer.__class__.__name__ in ["LlamaTokenizer", "LlamaTokenizerFast"]:
|
58 |
+
tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
|
59 |
+
|
60 |
+
if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
|
61 |
+
tokenizer.add_special_tokens({"pad_token": "[PAD]"})
|
62 |
+
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
63 |
+
|
64 |
+
if cfg.special_tokens:
|
65 |
+
for k, v in cfg.special_tokens.items():
|
66 |
+
tokenizer.add_special_tokens({k: v})
|
67 |
+
if cfg.tokens:
|
68 |
+
tokenizer.add_tokens(list(cfg.tokens))
|
69 |
+
|
70 |
+
return tokenizer
|
71 |
+
|
72 |
+
|
73 |
def load_model(
|
74 |
base_model,
|
75 |
base_model_config,
|
76 |
model_type,
|
77 |
+
tokenizer,
|
78 |
cfg,
|
79 |
adapter="lora",
|
80 |
inference=False,
|
81 |
):
|
82 |
+
# type: (str, str, str, str, AttrDefault, Optional[str], bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
|
83 |
|
84 |
# TODO refactor as a kwarg
|
85 |
load_in_8bit = cfg.load_in_8bit
|
|
|
86 |
is_llama_derived_model = "llama" in base_model or (
|
87 |
cfg.model_type and "llama" in cfg.model_type.lower()
|
88 |
)
|
|
|
157 |
model_path = str(cache_model_path)
|
158 |
except:
|
159 |
model_path = cfg.base_model
|
160 |
+
model, _ = load_llama_model_4bit_low_ram(
|
161 |
base_model_config if base_model_config else base_model,
|
162 |
model_path,
|
163 |
device_map=cfg.device_map,
|
|
|
242 |
**model_kwargs,
|
243 |
)
|
244 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
embeddings_len = math.ceil(len(tokenizer) / 32) * 32
|
246 |
model.resize_token_embeddings(embeddings_len)
|
247 |
|
|
|
290 |
model.config.use_cache = False
|
291 |
|
292 |
# TODO resume_from_checkpoint handling
|
293 |
+
return model, lora_config
|
294 |
|
295 |
|
296 |
def load_adapter(model, cfg, adapter):
|