FastChat
FastChat copied to clipboard
Published
20 hours ago •
lm-sys
lm-sys
Add scaled_dot_product_attention to replace flash attention
I have set scaled_dot_product_attention as default when the torch 2.0 was installed. It should be as efficient as original.
I have set scaled_dot_product_attention as default when the torch 2.0 was installed. It should be as efficient as original.
I used this code. It worked very well and also avoided installing flash-attn (I couldn't get this library installed). Thank you.
I tested this PR with Torch 2.0 on my 4x40GB A100, but found that it is 2x slower than the original flash attention implementation. I haven't dug into the details. I will keep this PR open as a solution for those who have difficulty installing flash attention.
I tested this PR with Torch 2.0 on my 4x40GB A100, but found that it is 2x slower than the original flash attention implementation. I haven't dug into the details. I will keep this PR open as a solution for those who have difficulty installing flash attention.
I have test it with 8x40GB A100, i don't found this gap.
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
runtime if needed. Op compatibility means that your system
meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-devel package with yum
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0
[WARNING] using untested triton version (2.0.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
utils .................. [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/opt/miniconda/lib/python3.8/site-packages/torch']
torch version .................... 2.0.0
deepspeed install path ........... ['/opt/miniconda/lib/python3.8/site-packages/deepspeed']
deepspeed info ................... 0.9.1, unknown, unknown
torch cuda version ............... 11.7
torch hip version ................ None
nvcc version ..................... 11.7
deepspeed wheel compiled w. ...... torch 2.0, cuda 11.7
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
#
# 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.
"""
Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...)
on a text file or a dataset without using HuggingFace Trainer.
Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=text-generation
"""
# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
import argparse
import json
import logging
import math
import os
import sys
from llama_attention_monkey_patch import replace_llama_attn_with_flash_attn
replace_llama_attn_with_flash_attn()
import random
from itertools import chain
from pathlib import Path
import datasets
import torch
from accelerate import Accelerator, DistributedType
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from datasets import load_dataset
from huggingface_hub import Repository, create_repo
from torch.utils.data import DataLoader
from tqdm.auto import tqdm
import transformers
from transformers import (
CONFIG_MAPPING,
MODEL_MAPPING,
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
SchedulerType,
default_data_collator,
get_scheduler,
)
from transformers.utils import get_full_repo_name, send_example_telemetry
from peft import LoraConfig, TaskType, get_peft_model
from accelerate.utils import DummyOptim, DummyScheduler
from deepspeed.accelerator import get_accelerator
datasets.disable_caching()
logger = get_logger(__name__)
MODEL_CONFIG_CLASSES = list(MODEL_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)
# Converting Bytes to Megabytes
def b2mb(x):
return int(x / 2 ** 20)
# This context manager is used to track the peak memory usage of the process
def parse_args():
parser = argparse.ArgumentParser(description="Finetune a transformers model on a causal language modeling task")
parser.add_argument(
"--dataset_name",
type=str,
default=None,
help="The name of the dataset to use (via the datasets library).",
)
parser.add_argument(
"--dataset_config_name",
type=str,
default=None,
help="The configuration name of the dataset to use (via the datasets library).",
)
parser.add_argument(
"--train_file", type=str, default=None, help="A csv or a json file containing the training data."
)
parser.add_argument(
"--validation_file", type=str, default=None, help="A csv or a json file containing the validation data."
)
parser.add_argument(
"--validation_split_percentage",
default=5,
help="The percentage of the train set used as validation set in case there's no validation split",
)
parser.add_argument(
"--model_name_or_path",
type=str,
help="Path to pretrained model or model identifier from huggingface.co/models.",
required=False,
)
parser.add_argument(
"--config_name",
type=str,
default=None,
help="Pretrained config name or path if not the same as model_name",
)
parser.add_argument(
"--tokenizer_name",
type=str,
default=None,
help="Pretrained tokenizer name or path if not the same as model_name",
)
parser.add_argument(
"--use_slow_tokenizer",
action="store_true",
help="If passed, will use a slow tokenizer (not backed by the 🤗 Tokenizers library).",
)
parser.add_argument(
"--per_device_train_batch_size",
type=int,
default=8,
help="Batch size (per device) for the training dataloader.",
)
parser.add_argument(
"--per_device_eval_batch_size",
type=int,
default=8,
help="Batch size (per device) for the evaluation dataloader.",
)
parser.add_argument(
"--learning_rate",
type=float,
default=5e-5,
help="Initial learning rate (after the potential warmup period) to use.",
)
parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.")
parser.add_argument("--num_train_epochs", type=int, default=3, help="Total number of training epochs to perform.")
parser.add_argument(
"--max_train_steps",
type=int,
default=None,
help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
)
parser.add_argument(
"--gradient_accumulation_steps",
type=int,
default=1,
help="Number of updates steps to accumulate before performing a backward/update pass.",
)
parser.add_argument(
"--lr_scheduler_type",
type=SchedulerType,
default="linear",
help="The scheduler type to use.",
choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
)
parser.add_argument(
"--lora_r",
type=int,
default=8,
)
parser.add_argument(
"--lora_alpa",
type=int,
default=32,
)
parser.add_argument("--gradient_checkpointing", action="store_true")
parser.add_argument(
"--num_warmup_steps", type=int, default=0, help="Number of steps for the warmup in the lr scheduler."
)
parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.")
parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
parser.add_argument(
"--model_type",
type=str,
default=None,
help="Model type to use if training from scratch.",
choices=MODEL_TYPES,
)
parser.add_argument(
"--block_size",
type=int,
default=None,
help=(
"Optional input sequence length after tokenization. The training dataset will be truncated in block of"
" this size for training. Default to the model max input length for single sentence inputs (take into"
" account special tokens)."
),
)
parser.add_argument("--use_group_texts",action="store_true")
parser.add_argument(
"--preprocessing_num_workers",
type=int,
default=None,
help="The number of processes to use for the preprocessing.",
)
parser.add_argument(
"--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
)
parser.add_argument(
"--no_keep_linebreaks", action="store_true", help="Do not keep line breaks when using TXT files."
)
parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
parser.add_argument(
"--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
)
parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
parser.add_argument(
"--checkpointing_steps",
type=str,
default="epoch",
help="Whether the various states should be saved at the end of every n steps, or 'epoch' for each epoch.",
)
parser.add_argument(
"--resume_from_checkpoint",
type=str,
default=None,
help="If the training should continue from a checkpoint folder.",
)
parser.add_argument(
"--with_tracking",
action="store_true",
help="Whether to enable experiment trackers for logging.",
)
parser.add_argument(
"--report_to",
type=str,
default="all",
help=(
'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
' `"wandb"`, `"comet_ml"` and `"clearml"`. Use `"all"` (default) to report to all integrations.'
"Only applicable when `--with_tracking` is passed."
),
)
args = parser.parse_args()
# Sanity checks
if args.dataset_name is None and args.train_file is None and args.validation_file is None:
raise ValueError("Need either a dataset name or a training/validation file.")
else:
if args.train_file is not None:
extension = args.train_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, json or txt file."
if args.validation_file is not None:
extension = args.validation_file.split(".")[-1]
assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, json or txt file."
if args.push_to_hub:
assert args.output_dir is not None, "Need an `output_dir` to create a repo when `--push_to_hub` is passed."
return args
def main():
args = parse_args()
# Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
# information sent is the one passed as arguments along with your Python/PyTorch versions.
send_example_telemetry("run_clm_no_trainer", args)
# Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
# If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
# in the environment
accelerator_log_kwargs = {}
if args.with_tracking:
accelerator_log_kwargs["log_with"] = args.report_to
accelerator_log_kwargs["project_dir"] = args.output_dir
accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps, **accelerator_log_kwargs)
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
# If passed along, set the training seed now.
if args.seed is not None:
set_seed(args.seed)
# Handle the repository creation
if accelerator.is_main_process:
if args.push_to_hub:
if args.hub_model_id is None:
repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
else:
repo_name = args.hub_model_id
create_repo(repo_name, exist_ok=True, token=args.hub_token)
repo = Repository(args.output_dir, clone_from=repo_name, token=args.hub_token)
with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
if "step_*" not in gitignore:
gitignore.write("step_*\n")
if "epoch_*" not in gitignore:
gitignore.write("epoch_*\n")
elif args.output_dir is not None:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[:{args.validation_split_percentage}%]",
)
raw_datasets["train"] = load_dataset(
args.dataset_name,
args.dataset_config_name,
split=f"train[{args.validation_split_percentage}%:]",
)
else:
data_files = {}
dataset_args = {}
if args.train_file is not None:
data_files["train"] = args.train_file
if args.validation_file is not None:
data_files["validation"] = args.validation_file
extension = args.train_file.split(".")[-1]
if extension == "txt":
extension = "text"
dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
raw_datasets = load_dataset(extension, data_files=data_files, **dataset_args)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
extension,
data_files=data_files,
split=f"train[:{args.validation_split_percentage}%]",
**dataset_args,
)
raw_datasets["train"] = load_dataset(
extension,
data_files=data_files,
split=f"train[{args.validation_split_percentage}%:]",
**dataset_args,
)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
# Load pretrained model and tokenizer
#
# In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
if args.config_name:
config = AutoConfig.from_pretrained(args.config_name)
elif args.model_name_or_path:
config = AutoConfig.from_pretrained(args.model_name_or_path)
else:
config = CONFIG_MAPPING[args.model_type]()
logger.warning("You are instantiating a new config instance from scratch.")
if args.tokenizer_name:
tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
elif args.model_name_or_path:
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
else:
raise ValueError(
"You are instantiating a new tokenizer from scratch. This is not supported by this script."
"You can do it from another script, save it, and load it from here, using --tokenizer_name."
)
# tokenizer.pad_token = tokenizer.eos_token
if args.model_name_or_path:
model = AutoModelForCausalLM.from_pretrained(
args.model_name_or_path,
from_tf=bool(".ckpt" in args.model_name_or_path),
config=config,
).half()
else:
logger.info("Training new model from scratch")
model = AutoModelForCausalLM.from_config(config)
# Preprocessing the datasets.
# First we tokenize all the texts.
column_names = raw_datasets["train"].column_names
text_column_name = "text" if "text" in column_names else column_names[0]
def tokenize_function(examples):
result = tokenizer(examples[text_column_name],
padding='max_length' if not args.use_group_texts else False,
truncation=True if not args.use_group_texts else False,
max_length = args.block_size if not args.use_group_texts else None,
return_tensors='pt')
result["labels"] = result["input_ids"].clone()
result["labels"][ result["labels"]==tokenizer.pad_token_id] = -100
return result
tokenizer.pad_token_id = tokenizer.eos_token_id
with accelerator.main_process_first():
tokenized_datasets = raw_datasets.map(
tokenize_function,
batched=True,
num_proc=args.preprocessing_num_workers,
remove_columns=column_names,
load_from_cache_file=not args.overwrite_cache,
desc="Running tokenizer on dataset",
)
if args.block_size is None:
block_size = tokenizer.model_max_length
if block_size > 1024:
logger.warning(
"The chosen tokenizer supports a `model_max_length` that is longer than the default `block_size` value"
" of 1024. If you would like to use a longer `block_size` up to `tokenizer.model_max_length` you can"
" override this default with `--block_size xxx`."
)
block_size = 1024
else:
if args.block_size > tokenizer.model_max_length:
logger.warning(
f"The block_size passed ({args.block_size}) is larger than the maximum length for the model"
f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
)
block_size = min(args.block_size, tokenizer.model_max_length)
# Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
def group_texts(examples):
# Concatenate all texts.
concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
total_length = len(concatenated_examples[list(examples.keys())[0]])
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
# customize this part to your needs.
if total_length >= block_size:
total_length = (total_length // block_size) * block_size
# Split by chunks of max_len.
result = {
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
for k, t in concatenated_examples.items()
}
result["labels"] = [i[1:] for i in result["input_ids"].copy()]
# result.pop("token_type_ids",None)
return result
# Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
# for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
# to preprocess.
#
# To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
# https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
with accelerator.main_process_first():
if args.use_group_texts:
lm_datasets = tokenized_datasets.map(
group_texts,
batched=True,
num_proc=args.preprocessing_num_workers,
load_from_cache_file=not args.overwrite_cache,
desc=f"Grouping texts in chunks of {block_size}",
)
else:
lm_datasets = tokenized_datasets
train_dataset = lm_datasets["train"]
eval_dataset = lm_datasets["validation"]
# Log a few random samples from the training set:
for index in random.sample(range(len(train_dataset)), 3):
logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
# DataLoaders creation:
train_dataloader = DataLoader(
train_dataset, shuffle=True, collate_fn=default_data_collator, batch_size=args.per_device_train_batch_size
)
eval_dataloader = DataLoader(
eval_dataset, collate_fn=default_data_collator, batch_size=args.per_device_eval_batch_size
)
peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM, inference_mode=False, r=args.lora_r, lora_alpha=args.lora_alpa, lora_dropout=0.1,#lora_match=True,
#target_modules=["query_key_value"], # ["query_key_value", "dense", "dense_h_to_4h", "dense_4h_to_h"],
)
for param in model.parameters():
param.requires_grad = False # freeze the model - train adapters later
if param.ndim == 1:
# cast the small parameters (e.g. layernorm) to fp32 for stability
param.data = param.data.to(torch.float32)
if args.gradient_checkpointing:
model.gradient_checkpointing_enable()
model.enable_input_require_grads()
model = get_peft_model(model, peft_config)
# Creates Dummy Optimizer if `optimizer` was spcified in the config file else creates Adam Optimizer
optimizer_cls = (
torch.optim.AdamW
if accelerator.state.deepspeed_plugin is None
or "optimizer" not in accelerator.state.deepspeed_plugin.deepspeed_config
else DummyOptim
)
# print(optimizer_cls,(optimizer_cls))
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters() if p.requires_grad],
"weight_decay": args.weight_decay,
},
]
optimizer = optimizer_cls(optimizer_grouped_parameters, lr=args.learning_rate, weight_decay=args.weight_decay)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
# Creates Dummy Scheduler if `scheduler` was spcified in the config file else creates `args.lr_scheduler_type` Scheduler
if (
accelerator.state.deepspeed_plugin is None
or "scheduler" not in accelerator.state.deepspeed_plugin.deepspeed_config
):
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps * args.gradient_accumulation_steps,
num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
)
else:
lr_scheduler = DummyScheduler(
optimizer, total_num_steps=args.max_train_steps, warmup_num_steps=args.num_warmup_steps
)
is_ds_zero_3 = False
if getattr(accelerator.state, "deepspeed_plugin", None):
is_ds_zero_3 = accelerator.state.deepspeed_plugin.zero_stage == 3
# Prepare everything with our `accelerator`.
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
)
#accelerator.print(model.print_trainable_parameters())
# On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
if accelerator.distributed_type == DistributedType.TPU:
model.tie_weights()
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
checkpointing_steps = args.checkpointing_steps
if checkpointing_steps is not None and checkpointing_steps.isdigit():
checkpointing_steps = int(checkpointing_steps)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initializes automatically on the main process.
if args.with_tracking:
experiment_config = vars(args)
# TensorBoard cannot log Enums, need the raw value
experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
accelerator.init_trackers("clm_no_trainer", experiment_config)
# Train!
total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
logger.info("***** Running training *****")
logger.info(f" Num examples = {len(train_dataset)}")
logger.info(f" Num Epochs = {args.num_train_epochs}")
logger.info(f" Instantaneous batch size per device = {args.per_device_train_batch_size}")
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
logger.info(f" Gradient Accumulation steps = {args.gradient_accumulation_steps}")
logger.info(f" Total optimization steps = {args.max_train_steps}")
# Only show the progress bar once on each machine.
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
completed_steps = 0
starting_epoch = 0
# Potentially load in the weights and states from a previous save
if args.resume_from_checkpoint:
if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
accelerator.print(f"Resumed from checkpoint: {args.resume_from_checkpoint}")
accelerator.load_state(args.resume_from_checkpoint)
path = os.path.basename(args.resume_from_checkpoint)
else:
# Get the most recent checkpoint
dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
dirs.sort(key=os.path.getctime)
path = dirs[-1] # Sorts folders by date modified, most recent checkpoint is the last
# Extract `epoch_{i}` or `step_{i}`
training_difference = os.path.splitext(path)[0]
if "epoch" in training_difference:
starting_epoch = int(training_difference.replace("epoch_", "")) + 1
resume_step = None
else:
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
# update the progress_bar if load from checkpoint
progress_bar.update(starting_epoch * num_update_steps_per_epoch)
completed_steps = starting_epoch * num_update_steps_per_epoch
with accelerator.main_process_first():
torch.cuda.empty_cache()
for epoch in range(starting_epoch, args.num_train_epochs):
model.train()
if args.with_tracking:
total_loss = 0
for step, batch in enumerate(train_dataloader):
# We need to skip steps until we reach the resumed step
if args.resume_from_checkpoint and epoch == starting_epoch:
if resume_step is not None and step < resume_step:
if step % args.gradient_accumulation_steps == 0:
progress_bar.update(1)
completed_steps += 1
continue
with accelerator.accumulate(model):
optimizer.zero_grad()
# see_memory_usage(f'before forward {model.global_steps}', force=True)
outputs = model(**batch,use_cache=False)
# see_memory_usage(f'after forward {model.global_steps}', force=True)
loss = outputs.loss
# We keep track of the loss at each epoch
if args.with_tracking:
total_loss += loss.detach().float()
# see_memory_usage(f'before backward {model.global_steps}', force=True)
accelerator.backward(loss)
# see_memory_usage(f'before optimizer {model.global_steps}', force=True)
optimizer.step()
lr_scheduler.step()
# see_memory_usage(f'after optimizer {model.global_steps}', force=True)
if is_ds_zero_3:
get_accelerator().empty_cache()
if args.with_tracking:
accelerator.log(
{
"train_loss_step": loss.item(),
},
step=completed_steps,
)
if accelerator.sync_gradients:
progress_bar.update(1)
completed_steps += 1
if isinstance(checkpointing_steps, int):
if completed_steps % checkpointing_steps == 0 and completed_steps > 0:
output_dir = f"step_{completed_steps}"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if completed_steps >= args.max_train_steps:
break
model.eval()
losses = []
for step, batch in enumerate(eval_dataloader):
with torch.no_grad():
outputs = model(**batch,use_cache=False)
loss = outputs.loss
losses.append(accelerator.gather_for_metrics(loss.repeat(args.per_device_eval_batch_size)))
losses = torch.cat(losses)
try:
eval_loss = torch.mean(losses)
perplexity = math.exp(eval_loss)
except OverflowError:
perplexity = float("inf")
logger.info(f"epoch {epoch}: perplexity: {perplexity} eval_loss: {eval_loss}")
if args.with_tracking:
accelerator.log(
{
"perplexity": perplexity,
"eval_loss": eval_loss,
"train_loss": total_loss.item() / len(train_dataloader),
"epoch": epoch,
"step": completed_steps,
},
step=completed_steps,
)
if args.checkpointing_steps == "epoch":
output_dir = f"epoch_{epoch}"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
accelerator.save_state(output_dir)
if args.push_to_hub and epoch < args.num_train_epochs - 1:
accelerator.wait_for_everyone()
output_dir = f"epoch_{epoch}"
if args.output_dir is not None:
output_dir = os.path.join(args.output_dir, output_dir)
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if args.with_tracking:
accelerator.end_training()
if args.output_dir is not None:
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(
args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
)
if accelerator.is_main_process:
with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
json.dump({"perplexity": perplexity}, f)
if __name__ == "__main__":
main()
LoRA Gradient Accumulation steps = 16 36/17340 [02:31<20:18:17, 4.22s/it] deepspeed zero 2 torch 2.0 42/17340 [03:05<21:08:16, 4.40s/it] deepspeed zero 2 flash_att
Gradient Accumulation steps = 1 373/277425 [01:45<21:42:30, 3.55it/s] deepspeed zero 2 flash_att 238/277425 [01:05<20:36:17, 3.74it/s] deepspeed zero 2 torch 2.0
@dumpmemory Hi would you mind sharing your deepspeed command and config? I am also interested in trying this out. Thanks!
haotian-liu
Hi, my deepspeed config is just accelerate config with zero 2 setting and cpu offload and bf16 enabled. I will upload later.
deepspeed.json
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"optimizer": {
"type": "Adam",
"params": {
"lr": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupLR",
"params": {
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu"
},
"offload_param": {
"device": "cpu"
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1e8,
"stage3_max_reuse_distance": 1e8,
"stage3_gather_16bit_weights_on_model_save": true
},
"gradient_clipping": 0.7,
"steps_per_print": 1,
"bf16": {
"enabled": true
},
"fp16": {
"loss_scale": 0,
"auto_cast":true,
"loss_scale_window": 1000,
"initial_scale_power": 32,
"hysteresis": 2,
"min_loss_scale": 0.5,
"enabled": false
}
}
As @zhisbug said, there's a fix for using xformers instead. Which wouldn't force a full upgrade to Pytorch2 , which is VERY undesirable in many big systems.
Hi @dumpmemory , I noticed that you use
tokenizer.pad_token_id = tokenizer.eos_token_idto set pad_token_id. May I ask why? Was the original Vicuna trained with this kind of operation?
Hi @dumpmemory , I noticed that you use
tokenizer.pad_token_id = tokenizer.eos_token_idto set pad_token_id. May I ask why? Was the original Vicuna trained with this kind of operation?
Sorry to bother you, I figured it out and found in FastChat/fastchat/train/train.py:269 that the original script use tokenizer.pad_token = tokenizer.unk_token