126 lines
4.5 KiB
Python
126 lines
4.5 KiB
Python
import gc
|
|
import torch
|
|
|
|
from Context import Context
|
|
from Dataset import Dataset
|
|
from ModelLoader import ModelLoader
|
|
from transformers import AutoTokenizer
|
|
from peft import LoraConfig
|
|
from trl import SFTTrainer, SFTConfig
|
|
from Validator import Validator
|
|
|
|
class Trainer:
|
|
def train(self, context: Context):
|
|
if not context.train:
|
|
return
|
|
|
|
baseModel = ModelLoader().load(context.locBaseModel, False, True, context)
|
|
sftTrainer = self.__createTrainer(baseModel, context)
|
|
try:
|
|
vInplace = context.validate and context.vInplace
|
|
validator = None
|
|
if vInplace:
|
|
validator = Validator()
|
|
|
|
try:
|
|
cnt = 0
|
|
continueTraining = True
|
|
while continueTraining:
|
|
cnt += 1
|
|
print(f"Training run {cnt}")
|
|
|
|
self.__train(sftTrainer)
|
|
|
|
if vInplace:
|
|
continueTraining = not validator.validateInPlace(sftTrainer.tokenizer, sftTrainer.model, context)
|
|
else:
|
|
continueTraining = False
|
|
|
|
print(f"Total training runs: {cnt} epochs {cnt * (context.trEpochs or 1)}")
|
|
|
|
finally:
|
|
if vInplace:
|
|
print(f"Validation statistics:\n{validator.statistics}")
|
|
validator.unload(context)
|
|
|
|
if context.storeAdapter:
|
|
self.__storeAdapter(sftTrainer, context)
|
|
|
|
finally:
|
|
del sftTrainer
|
|
del baseModel
|
|
gc.collect()
|
|
if context.accel:
|
|
torch.cuda.empty_cache()
|
|
|
|
def __createTrainer(self, baseModel, context: Context):
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
context.locBaseModel,
|
|
trust_remote_code = True)
|
|
if tokenizer.pad_token == None:
|
|
tokenizer.pad_token = tokenizer.eos_token
|
|
|
|
dataset = Dataset().scan(context.locDataset, tokenizer, context)
|
|
|
|
peftConfig = LoraConfig(
|
|
r = context.loraR,
|
|
lora_alpha = context.loraAlpha,
|
|
lora_dropout = context.loraDropout,
|
|
bias = context.loraBias,
|
|
task_type = context.loraTaskType
|
|
#target_modules=["q_proj", "k_proj", "v_proj", "out_proj", "fc_in", "fc_out", "wte"],
|
|
)
|
|
|
|
sftArgs = {}
|
|
if context.accel:
|
|
sftArgs["fp16"] = True
|
|
else:
|
|
sftArgs["bf16"] = True
|
|
if context.trEpochs:
|
|
sftArgs["num_train_epochs"] = context.trEpochs
|
|
if context.trMaxSeqLength:
|
|
sftArgs["max_seq_length"] = context.trMaxSeqLength
|
|
if context.trPerDeviceTrainBatchSize:
|
|
sftArgs["per_device_train_batch_size"] = context.trPerDeviceTrainBatchSize
|
|
if context.trFindAutoBatchSize:
|
|
sftArgs["auto_find_batch_size"] = context.trFindAutoBatchSize
|
|
if context.trGradientAccSteps:
|
|
sftArgs["gradient_accumulation_steps"] = context.trGradientAccSteps
|
|
if context.trGradientCheckpointing:
|
|
sftArgs["gradient_checkpointing"] = context.trGradientCheckpointing
|
|
if context.trGroupByLength:
|
|
sftArgs["group_by_length"] = context.trGroupByLength
|
|
if context.trPacking:
|
|
sftArgs["packing"] = context.trPacking
|
|
if context.trOptim:
|
|
sftArgs["optim"] = context.trOptim
|
|
if context.trSchedulerType:
|
|
sftArgs["lr_scheduler_type"] = context.trSchedulerType
|
|
sftConfig = SFTConfig(
|
|
dataset_text_field="text",
|
|
output_dir=context.locWorkdir,
|
|
save_strategy="no",
|
|
#neftune_noise_alpha=5,
|
|
**sftArgs
|
|
)
|
|
|
|
sftTrainer = SFTTrainer(
|
|
model = baseModel,
|
|
train_dataset = dataset,
|
|
tokenizer = tokenizer,
|
|
peft_config = peftConfig,
|
|
args = sftConfig
|
|
)
|
|
|
|
return sftTrainer
|
|
|
|
def __train(self, sftTrainer):
|
|
sftTrainer.train()
|
|
print("Training finished")
|
|
|
|
def __storeAdapter(self, sftTrainer, context: Context):
|
|
print("Saving adapter...")
|
|
|
|
sftTrainer.model.save_pretrained(context.locAdapter)
|
|
sftTrainer.tokenizer.save_pretrained(context.locAdapter)
|