Spaces:
Runtime error
Runtime error
zetavg
commited on
Commit
•
300b660
1
Parent(s):
85fb243
update
Browse files
llama_lora/lib/finetune.py
CHANGED
@@ -70,7 +70,13 @@ def train(
|
|
70 |
wandb_tags: List[str] = [],
|
71 |
wandb_watch: str = "false", # options: false | gradients | all
|
72 |
wandb_log_model: str = "true", # options: false | true
|
|
|
73 |
):
|
|
|
|
|
|
|
|
|
|
|
74 |
if lora_modules_to_save is not None and len(lora_modules_to_save) <= 0:
|
75 |
lora_modules_to_save = None
|
76 |
|
@@ -171,6 +177,16 @@ def train(
|
|
171 |
if ddp:
|
172 |
device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
|
173 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
model = base_model
|
175 |
if isinstance(model, str):
|
176 |
model_name = model
|
@@ -216,51 +232,16 @@ def train(
|
|
216 |
# )
|
217 |
tokenizer.padding_side = "left" # Allow batched inference
|
218 |
|
219 |
-
def tokenize(prompt, add_eos_token=True):
|
220 |
-
# there's probably a way to do this with the tokenizer settings
|
221 |
-
# but again, gotta move fast
|
222 |
-
result = tokenizer(
|
223 |
-
prompt,
|
224 |
-
truncation=True,
|
225 |
-
max_length=cutoff_len,
|
226 |
-
padding=False,
|
227 |
-
return_tensors=None,
|
228 |
-
)
|
229 |
-
if (
|
230 |
-
result["input_ids"][-1] != tokenizer.eos_token_id
|
231 |
-
and len(result["input_ids"]) < cutoff_len
|
232 |
-
and add_eos_token
|
233 |
-
):
|
234 |
-
result["input_ids"].append(tokenizer.eos_token_id)
|
235 |
-
result["attention_mask"].append(1)
|
236 |
-
|
237 |
-
result["labels"] = result["input_ids"].copy()
|
238 |
-
|
239 |
-
return result
|
240 |
-
|
241 |
-
def generate_and_tokenize_prompt(data_point):
|
242 |
-
full_prompt = data_point["prompt"] + data_point["completion"]
|
243 |
-
tokenized_full_prompt = tokenize(full_prompt)
|
244 |
-
if not train_on_inputs:
|
245 |
-
user_prompt = data_point["prompt"]
|
246 |
-
tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
|
247 |
-
user_prompt_len = len(tokenized_user_prompt["input_ids"])
|
248 |
-
|
249 |
-
tokenized_full_prompt["labels"] = [
|
250 |
-
-100
|
251 |
-
] * user_prompt_len + tokenized_full_prompt["labels"][
|
252 |
-
user_prompt_len:
|
253 |
-
] # could be sped up, probably
|
254 |
-
return tokenized_full_prompt
|
255 |
-
|
256 |
-
# will fail anyway.
|
257 |
try:
|
258 |
model = prepare_model_for_int8_training(model)
|
259 |
except Exception as e:
|
260 |
print(
|
261 |
f"Got error while running prepare_model_for_int8_training(model), maybe the model has already be prepared. Original error: {e}.")
|
262 |
|
263 |
-
|
|
|
|
|
|
|
264 |
|
265 |
lora_config_args = {
|
266 |
'r': lora_r,
|
@@ -279,12 +260,6 @@ def train(
|
|
279 |
if bf16:
|
280 |
model = model.to(torch.bfloat16)
|
281 |
|
282 |
-
# If train_data is a list, convert it to datasets.Dataset
|
283 |
-
if isinstance(train_data, list):
|
284 |
-
with open(os.path.join(output_dir, "train_data_samples.json"), 'w') as file:
|
285 |
-
json.dump(list(train_data[:100]), file, indent=2)
|
286 |
-
train_data = Dataset.from_list(train_data)
|
287 |
-
|
288 |
if resume_from_checkpoint:
|
289 |
# Check the available weights and load them
|
290 |
checkpoint_name = os.path.join(
|
@@ -320,6 +295,54 @@ def train(
|
|
320 |
wandb.config.update({"model": {"all_params": all_params, "trainable_params": trainable_params,
|
321 |
"trainable%": 100 * trainable_params / all_params}})
|
322 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
323 |
if val_set_size > 0:
|
324 |
train_val = train_data.train_test_split(
|
325 |
test_size=val_set_size, shuffle=True, seed=42
|
@@ -339,6 +362,11 @@ def train(
|
|
339 |
model.is_parallelizable = True
|
340 |
model.model_parallel = True
|
341 |
|
|
|
|
|
|
|
|
|
|
|
342 |
# https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments
|
343 |
training_args = {
|
344 |
'output_dir': output_dir,
|
|
|
70 |
wandb_tags: List[str] = [],
|
71 |
wandb_watch: str = "false", # options: false | gradients | all
|
72 |
wandb_log_model: str = "true", # options: false | true
|
73 |
+
status_message_callback: Any = None,
|
74 |
):
|
75 |
+
if status_message_callback:
|
76 |
+
cb_result = status_message_callback("Preparing training...")
|
77 |
+
if cb_result:
|
78 |
+
return
|
79 |
+
|
80 |
if lora_modules_to_save is not None and len(lora_modules_to_save) <= 0:
|
81 |
lora_modules_to_save = None
|
82 |
|
|
|
177 |
if ddp:
|
178 |
device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
|
179 |
|
180 |
+
if status_message_callback:
|
181 |
+
if isinstance(base_model, str):
|
182 |
+
cb_result = status_message_callback(f"Preparing model '{base_model}' for training...")
|
183 |
+
if cb_result:
|
184 |
+
return
|
185 |
+
else:
|
186 |
+
cb_result = status_message_callback("Preparing model for training...")
|
187 |
+
if cb_result:
|
188 |
+
return
|
189 |
+
|
190 |
model = base_model
|
191 |
if isinstance(model, str):
|
192 |
model_name = model
|
|
|
232 |
# )
|
233 |
tokenizer.padding_side = "left" # Allow batched inference
|
234 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
try:
|
236 |
model = prepare_model_for_int8_training(model)
|
237 |
except Exception as e:
|
238 |
print(
|
239 |
f"Got error while running prepare_model_for_int8_training(model), maybe the model has already be prepared. Original error: {e}.")
|
240 |
|
241 |
+
if status_message_callback:
|
242 |
+
cb_result = status_message_callback("Preparing PEFT model for training...")
|
243 |
+
if cb_result:
|
244 |
+
return
|
245 |
|
246 |
lora_config_args = {
|
247 |
'r': lora_r,
|
|
|
260 |
if bf16:
|
261 |
model = model.to(torch.bfloat16)
|
262 |
|
|
|
|
|
|
|
|
|
|
|
|
|
263 |
if resume_from_checkpoint:
|
264 |
# Check the available weights and load them
|
265 |
checkpoint_name = os.path.join(
|
|
|
295 |
wandb.config.update({"model": {"all_params": all_params, "trainable_params": trainable_params,
|
296 |
"trainable%": 100 * trainable_params / all_params}})
|
297 |
|
298 |
+
if status_message_callback:
|
299 |
+
cb_result = status_message_callback("Preparing train data...")
|
300 |
+
if cb_result:
|
301 |
+
return
|
302 |
+
|
303 |
+
def tokenize(prompt, add_eos_token=True):
|
304 |
+
# there's probably a way to do this with the tokenizer settings
|
305 |
+
# but again, gotta move fast
|
306 |
+
result = tokenizer(
|
307 |
+
prompt,
|
308 |
+
truncation=True,
|
309 |
+
max_length=cutoff_len,
|
310 |
+
padding=False,
|
311 |
+
return_tensors=None,
|
312 |
+
)
|
313 |
+
if (
|
314 |
+
result["input_ids"][-1] != tokenizer.eos_token_id
|
315 |
+
and len(result["input_ids"]) < cutoff_len
|
316 |
+
and add_eos_token
|
317 |
+
):
|
318 |
+
result["input_ids"].append(tokenizer.eos_token_id)
|
319 |
+
result["attention_mask"].append(1)
|
320 |
+
|
321 |
+
result["labels"] = result["input_ids"].copy()
|
322 |
+
|
323 |
+
return result
|
324 |
+
|
325 |
+
def generate_and_tokenize_prompt(data_point):
|
326 |
+
full_prompt = data_point["prompt"] + data_point["completion"]
|
327 |
+
tokenized_full_prompt = tokenize(full_prompt)
|
328 |
+
if not train_on_inputs:
|
329 |
+
user_prompt = data_point["prompt"]
|
330 |
+
tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
|
331 |
+
user_prompt_len = len(tokenized_user_prompt["input_ids"])
|
332 |
+
|
333 |
+
tokenized_full_prompt["labels"] = [
|
334 |
+
-100
|
335 |
+
] * user_prompt_len + tokenized_full_prompt["labels"][
|
336 |
+
user_prompt_len:
|
337 |
+
] # could be sped up, probably
|
338 |
+
return tokenized_full_prompt
|
339 |
+
|
340 |
+
# If train_data is a list, convert it to datasets.Dataset
|
341 |
+
if isinstance(train_data, list):
|
342 |
+
with open(os.path.join(output_dir, "train_data_samples.json"), 'w') as file:
|
343 |
+
json.dump(list(train_data[:100]), file, indent=2)
|
344 |
+
train_data = Dataset.from_list(train_data)
|
345 |
+
|
346 |
if val_set_size > 0:
|
347 |
train_val = train_data.train_test_split(
|
348 |
test_size=val_set_size, shuffle=True, seed=42
|
|
|
362 |
model.is_parallelizable = True
|
363 |
model.model_parallel = True
|
364 |
|
365 |
+
if status_message_callback:
|
366 |
+
cb_result = status_message_callback("Train starting...")
|
367 |
+
if cb_result:
|
368 |
+
return
|
369 |
+
|
370 |
# https://huggingface.co/docs/transformers/main/en/main_classes/trainer#transformers.TrainingArguments
|
371 |
training_args = {
|
372 |
'output_dir': output_dir,
|
llama_lora/ui/finetune/finetune_ui.py
CHANGED
@@ -309,6 +309,7 @@ def handle_lora_modules_to_save_add(choices, new_module, selected_modules):
|
|
309 |
|
310 |
def do_abort_training():
|
311 |
Global.should_stop_training = True
|
|
|
312 |
|
313 |
|
314 |
def finetune_ui():
|
|
|
309 |
|
310 |
def do_abort_training():
|
311 |
Global.should_stop_training = True
|
312 |
+
Global.training_status_text = "Aborting..."
|
313 |
|
314 |
|
315 |
def finetune_ui():
|
llama_lora/ui/finetune/training.py
CHANGED
@@ -22,6 +22,13 @@ from ..trainer_callback import (
|
|
22 |
from .data_processing import get_data_from_input
|
23 |
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
def do_train(
|
26 |
# Dataset
|
27 |
template,
|
@@ -254,6 +261,7 @@ def do_train(
|
|
254 |
train_output = Global.finetune_train_fn(
|
255 |
train_data=train_data,
|
256 |
callbacks=training_callbacks,
|
|
|
257 |
**finetune_args,
|
258 |
)
|
259 |
|
|
|
22 |
from .data_processing import get_data_from_input
|
23 |
|
24 |
|
25 |
+
def status_message_callback(message):
|
26 |
+
if Global.should_stop_training:
|
27 |
+
return True
|
28 |
+
|
29 |
+
Global.training_status_text = message
|
30 |
+
|
31 |
+
|
32 |
def do_train(
|
33 |
# Dataset
|
34 |
template,
|
|
|
261 |
train_output = Global.finetune_train_fn(
|
262 |
train_data=train_data,
|
263 |
callbacks=training_callbacks,
|
264 |
+
status_message_callback=status_message_callback,
|
265 |
**finetune_args,
|
266 |
)
|
267 |
|
llama_lora/ui/trainer_callback.py
CHANGED
@@ -57,6 +57,9 @@ def update_training_states(
|
|
57 |
Global.training_log_history = log_history
|
58 |
Global.training_eta = Global.training_eta_predictor.predict_eta(current_step, total_steps)
|
59 |
|
|
|
|
|
|
|
60 |
last_history = None
|
61 |
last_loss = None
|
62 |
if len(Global.training_log_history) > 0:
|
|
|
57 |
Global.training_log_history = log_history
|
58 |
Global.training_eta = Global.training_eta_predictor.predict_eta(current_step, total_steps)
|
59 |
|
60 |
+
if Global.should_stop_training:
|
61 |
+
return
|
62 |
+
|
63 |
last_history = None
|
64 |
last_loss = None
|
65 |
if len(Global.training_log_history) > 0:
|