zetavg
update
e9c5abc unverified
raw
history blame
7.53 kB
import os
import sys
from typing import Any, List
import fire
import torch
import transformers
from datasets import Dataset, load_dataset
from peft import (
LoraConfig,
get_peft_model,
get_peft_model_state_dict,
prepare_model_for_int8_training,
set_peft_model_state_dict,
)
from transformers import LlamaForCausalLM, LlamaTokenizer
def train(
# model/data params
base_model: Any,
tokenizer: Any,
output_dir: str,
train_dataset_data: List[Any],
# training hyperparams
micro_batch_size: int = 4,
gradient_accumulation_steps: int = 32,
num_epochs: int = 3,
learning_rate: float = 3e-4,
cutoff_len: int = 256,
val_set_size: int = 2000,
# lora hyperparams
lora_r: int = 8,
lora_alpha: int = 16,
lora_dropout: float = 0.05,
lora_target_modules: List[str] = [
"q_proj",
"v_proj",
],
# llm hyperparams
train_on_inputs: bool = True, # if False, masks out inputs in loss
group_by_length: bool = False, # faster, but produces an odd training loss curve
# either training checkpoint or final adapter
resume_from_checkpoint: str = None,
# logging
callbacks: List[Any] = []
):
device_map = "auto"
world_size = int(os.environ.get("WORLD_SIZE", 1))
ddp = world_size != 1
if ddp:
device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
model = base_model
if isinstance(model, str):
model = LlamaForCausalLM.from_pretrained(
base_model,
load_in_8bit=True,
torch_dtype=torch.float16,
device_map=device_map,
)
if isinstance(tokenizer, str):
tokenizer = LlamaTokenizer.from_pretrained(tokenizer)
tokenizer.pad_token_id = (
0 # unk. we want this to be different from the eos token
)
tokenizer.padding_side = "left" # Allow batched inference
def tokenize(prompt, add_eos_token=True):
# there's probably a way to do this with the tokenizer settings
# but again, gotta move fast
result = tokenizer(
prompt,
truncation=True,
max_length=cutoff_len,
padding=False,
return_tensors=None,
)
if (
result["input_ids"][-1] != tokenizer.eos_token_id
and len(result["input_ids"]) < cutoff_len
and add_eos_token
):
result["input_ids"].append(tokenizer.eos_token_id)
result["attention_mask"].append(1)
result["labels"] = result["input_ids"].copy()
return result
def generate_and_tokenize_prompt(data_point):
full_prompt = data_point["prompt"] + data_point["completion"]
tokenized_full_prompt = tokenize(full_prompt)
if not train_on_inputs:
user_prompt = data_point["prompt"]
tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
user_prompt_len = len(tokenized_user_prompt["input_ids"])
tokenized_full_prompt["labels"] = [
-100
] * user_prompt_len + tokenized_full_prompt["labels"][
user_prompt_len:
] # could be sped up, probably
return tokenized_full_prompt
# will fail anyway.
try:
model = prepare_model_for_int8_training(model)
except Exception as e:
print(
f"Got error while running prepare_model_for_int8_training(model), maybe the model has already be prepared. Original error: {e}.")
# model = prepare_model_for_int8_training(model)
config = LoraConfig(
r=lora_r,
lora_alpha=lora_alpha,
target_modules=lora_target_modules,
lora_dropout=lora_dropout,
bias="none",
task_type="CAUSAL_LM",
)
model = get_peft_model(model, config)
# If train_dataset_data is a list, convert it to datasets.Dataset
if isinstance(train_dataset_data, list):
train_dataset_data = Dataset.from_list(train_dataset_data)
if resume_from_checkpoint:
# Check the available weights and load them
checkpoint_name = os.path.join(
resume_from_checkpoint, "pytorch_model.bin"
) # Full checkpoint
if not os.path.exists(checkpoint_name):
checkpoint_name = os.path.join(
resume_from_checkpoint, "adapter_model.bin"
) # only LoRA model - LoRA config above has to fit
resume_from_checkpoint = (
False # So the trainer won't try loading its state
)
# The two files above have a different name depending on how they were saved, but are actually the same.
if os.path.exists(checkpoint_name):
print(f"Restarting from {checkpoint_name}")
adapters_weights = torch.load(checkpoint_name)
model = set_peft_model_state_dict(model, adapters_weights)
else:
print(f"Checkpoint {checkpoint_name} not found")
# Be more transparent about the % of trainable params.
model.print_trainable_parameters()
if val_set_size > 0:
train_val = train_dataset_data.train_test_split(
test_size=val_set_size, shuffle=True, seed=42
)
train_data = (
train_val["train"].shuffle().map(generate_and_tokenize_prompt)
)
val_data = (
train_val["test"].shuffle().map(generate_and_tokenize_prompt)
)
else:
train_data = train_dataset_data.shuffle().map(generate_and_tokenize_prompt)
val_data = None
if not ddp and torch.cuda.device_count() > 1:
# keeps Trainer from trying its own DataParallelism when more than 1 gpu is available
model.is_parallelizable = True
model.model_parallel = True
trainer = transformers.Trainer(
model=model,
train_dataset=train_data,
eval_dataset=val_data,
args=transformers.TrainingArguments(
per_device_train_batch_size=micro_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
warmup_steps=100,
num_train_epochs=num_epochs,
learning_rate=learning_rate,
fp16=True,
logging_steps=10,
optim="adamw_torch",
evaluation_strategy="steps" if val_set_size > 0 else "no",
save_strategy="steps",
eval_steps=200 if val_set_size > 0 else None,
save_steps=200,
output_dir=output_dir,
save_total_limit=3,
load_best_model_at_end=True if val_set_size > 0 else False,
ddp_find_unused_parameters=False if ddp else None,
group_by_length=group_by_length,
# report_to="wandb" if use_wandb else None,
# run_name=wandb_run_name if use_wandb else None,
),
data_collator=transformers.DataCollatorForSeq2Seq(
tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
),
callbacks=callbacks,
)
model.config.use_cache = False
old_state_dict = model.state_dict
model.state_dict = (
lambda self, *_, **__: get_peft_model_state_dict(
self, old_state_dict()
)
).__get__(model, type(model))
if torch.__version__ >= "2" and sys.platform != "win32":
model = torch.compile(model)
result = trainer.train(resume_from_checkpoint=resume_from_checkpoint)
model.save_pretrained(output_dir)
print(f"Model saved to {output_dir}.")
return result