Benchmarking SFT trainer with 8bit models

finetune_sft_trl.py

# coding=utf-8
# Copyright 2023 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.
import os
import tempfile
from dataclasses import dataclass, field
from typing import Optional
import torch

from datasets import load_dataset
from tqdm import tqdm
from accelerate import Accelerator
from transformers import (
    AutoModelForCausalLM,
    AutoModelForSeq2SeqLM,
    LlamaTokenizer,
    HfArgumentParser,
    AutoTokenizer,
    TrainingArguments,
    BitsAndBytesConfig,
)
from peft import LoraConfig

from trl import SFTTrainer


tqdm.pandas()

########################################################################
# This is a fully working simple example to use trl's SFTTrainer.
#
# This example fine-tunes any causal language model (GPT-2, GPT-Neo, etc.)
# by using the SFTTrainer from trl, we will leverage PEFT library to finetune
# adapters on the model.
#
########################################################################

@dataclass
class ScriptArguments:
    """
    Define the arguments used in this script.
    """

    model_name: Optional[str] = field(default="decapoda-research/llama-7b-hf", metadata={"help": "the model name"})
    dataset_name: Optional[str] = field(default="ybelkada/oasst1-tiny-subset", metadata={"help": "the dataset name"})
    use_8_bit: Optional[bool] = field(default=False, metadata={"help": "use 8 bit precision"})
    use_seq2seq_lm: Optional[bool] = field(default=False, metadata={"help": "use seq2seq LM"})
    use_4_bit: Optional[bool] = field(default=True, metadata={"help": "use 4 bit precision"})
    bnb_4bit_quant_type: Optional[str] = field(default="nf4", metadata={"help": "precise the quantization type (fp4 or nf4)"})
    use_bnb_nested_quant: Optional[bool] = field(default=False, metadata={"help": "use nested quantization"})
    use_multi_gpu: Optional[bool] = field(default=False, metadata={"help": "use multi GPU"})
    use_adapters: Optional[bool] = field(default=True, metadata={"help": "use adapters"})
    batch_size: Optional[int] = field(default=1, metadata={"help": "input batch size"})
    max_seq_length: Optional[int] = field(default=512, metadata={"help": "max sequence length"})
    optimizer_name: Optional[str] = field(default="adamw_hf", metadata={"help": "Optimizer name"})

def get_current_device():
    return Accelerator().process_index

parser = HfArgumentParser(ScriptArguments)
script_args = parser.parse_args_into_dataclasses()[0]

dataset = load_dataset(script_args.dataset_name, split="train[:1%]")

# We load the model
if script_args.use_multi_gpu:
    device_map = "auto"
else:
    device_map = {"":get_current_device()}

if script_args.use_8_bit and script_args.use_4_bit:
    raise ValueError(
        "You can't use 8 bit and 4 bit precision at the same time"
    )

if script_args.use_4_bit:
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float16,
        bnb_4bit_quant_type=script_args.bnb_4bit_quant_type,
        bnb_4bit_use_double_quant=script_args.use_bnb_nested_quant,
    )   
else:
    bnb_config = None

transformers_class = AutoModelForSeq2SeqLM if script_args.use_seq2seq_lm else AutoModelForCausalLM

model = transformers_class.from_pretrained(
    script_args.model_name, 
    load_in_8bit=script_args.use_8_bit, 
    load_in_4bit=script_args.use_4_bit,
    device_map=device_map if (script_args.use_8_bit or script_args.use_4_bit) else None,
    quantization_config=bnb_config,
    torch_dtype=torch.float16,
)

if script_args.use_adapters:
    peft_config = LoraConfig(
        lora_alpha=32,
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM" if not script_args.use_seq2seq_lm else "SEQ_2_SEQ_LM",
    )
else:
    peft_config = None
    if script_args.use_8_bit:
        raise ValueError(
            "You need to use adapters to use 8 bit precision"
        )

if "llama" in script_args.model_name:
    tokenizer = LlamaTokenizer.from_pretrained(script_args.model_name)
    tokenizer.add_special_tokens({'pad_token': '[PAD]'})
else:
    tokenizer = AutoTokenizer.from_pretrained(script_args.model_name)

with tempfile.TemporaryDirectory() as tmp_dir:
    training_arguments = TrainingArguments(
        per_device_train_batch_size=script_args.batch_size,
        max_steps=10,
        gradient_accumulation_steps=4,
        per_device_eval_batch_size=script_args.batch_size,
        output_dir=tmp_dir,
        report_to=["none"],
        optim=script_args.optimizer_name,
        fp16=True,
    )

    trainer = SFTTrainer(
        model=model,
        tokenizer=tokenizer,
        train_dataset=dataset,
        dataset_text_field="messages",
        peft_config=peft_config,
        max_seq_length=script_args.max_seq_length,
        args=training_arguments,
    )

    trainer.train()

    assert "adapter_model.bin" in os.listdir(tmp_dir)

can I use my csv file?

Out of interest, would this work for AutoModelForSeq2SeqLM. For starters, wouldn't you need an input_text_field and a output_text_field, and right now the SFTTrainer only takes in one dataset_text_field.

@sachinruk you are right, SFTTrainer is not supposed to work for SeqToSeq (encoder-decoder) models.. The premise of SFTTrainer is to easily train decoder only/generative models.

@younesbelkada , do we need to pass fp16=True in the TrainingArguments if we are using bnb_config while loading the model?

@dhruvmullick those are 2 separate things. fp16=True refers to the PEFT adapter, and bnb_config refers to the base model. You can set both however you like independent of each other.

Makes sense, thanks @tdolega