sdeed8/collie

A Light Toolkit to Finetune Large Models.

CoLLiE

Full Parameter Fine-tuning of Large Language Models on Consumer Hardware.

Beta release (0.0.1)

To install:

python setup.py install

Speed Benchmark

Throughput when training LLaMA with CoLLiE

Model Size	Parallel	Optimizer	Hardware	Seq. Length	Batch Size	Sec. per Step	TGS
7B	Pipeline	SGD	8 * RTX 3090	1024	128	31.0	528.5
7B	ZeRO-3	Inplace SGD	1 * RTX 3090	1024	3	4.0	832.5
13B	Pipeline	SGD	8 * RTX 3090	1024	128	47.0	348.5
13B	ZeRO-3	Inplace SGD	2 * RTX 3090	1024	4	11.7	350.1
30B	ZeRO-3	Inplace SGD	4 * RTX 3090	960	2	34.1	56.4
65B	ZeRO-3	Inplace SGD	8 * RTX 3090	512	2	67.1	15.3

* TGS = Tokens / GPU / Second

How to use CoLLiE

Here's a simple example to run pipeline parallel:

# Command: $torchrun --nproc_per_node=8 train.py
from collie.models.llama_colossalai import ModelArgs, get_7B_llama, load_state_dict
from collie.models.llama_tokenizer import HFLikeTokenizer, Tokenizer
from collie.trainer.colossalai_trainer import ColossalaiTrainer, TrainerArgs
from torch.utils.data import DataLoader

tokenizer = HFLikeTokenizer(tokenizer=Tokenizer(model_path='./tokenizer.model'))
model_args = ModelArgs()
model_args.pp_size = 8 # your pipeline parallel degree
model = get_7B_llama(model_args=model_args)
state_dict = load_state_dict(
    protocol="file", 
    source="raw", 
    file_folder="./llama/7B",  # where consolidated.00.pth in it
    model_args=model_args)
model.load_state_dict(state_dict)
train_sample = {
    "input_ids": tokenizer("CoLLiE is a python package for training large language models", return_tensors="pt")["input_ids"].long()
}, tokenizer("CoLLiE is a python package for training large language models", return_tensors="pt")["input_ids"].long()
eval_sample = {
    "input_ids": tokenizer("CoLLiE is a python package for", return_tensors="pt")["input_ids"].long()
}, tokenizer("CoLLiE is a python package for", return_tensors="pt")["input_ids"].long()

train_dataloader = DataLoader(
    [train_sample for _ in range(1000)],
    batch_size=1
)
eval_dataloader = DataLoader(
    [eval_sample for _ in range(1000)],
    batch_size=1
)
def compute_metrics(batch, generated_batch, epoch, step):
    print("\n")
    print(tokenizer.decode(generated_batch[0]["input_ids"].tolist()))
    print("\n")
    
trainer = ColossalaiTrainer(model=model,
                            train_dataloader=train_dataloader,
                            eval_dataloader=eval_dataloader,
                            tokenizer=tokenizer,
                            compute_metrics=compute_metrics)
trainer.train()

To run tensor parallel with tensor parallel and inplace:

# Command: $torchrun --nproc_per_node 2 train.py ./example/config/tensor_args.yaml
import collie
from collie.models import llama
import torch
from torch.utils.data import Dataset, DataLoader
from datasets import load_dataset, load_from_disk
from transformers import HfArgumentParser
from collie.arguments import ModelArguments, DataArguments, CollieArguments
import os


def collate_fn(batch, tknz, max_len=512):
    text = [e['text'] for e in batch]
    tknz_batch = tknz(
        text,
        max_length=max_len,
        padding='max_length',
        truncation=True,
        return_tensors='pt'
    )
    eos_tensors = torch.tensor([tknz.eos_token_id] * tknz_batch['input_ids'].shape[0]).unsqueeze(1)
    tknz_batch['input_ids'] = torch.cat((tknz_batch['input_ids'][:, :max_len-1], eos_tensors), dim=1)
    tknz_batch['attention_mask'] = tknz_batch['attention_mask'][:, :max_len]
    return {
        'input_ids': tknz_batch['input_ids'],
        'attention_mask': tknz_batch['attention_mask'],
        'labels': tknz_batch['input_ids']
    }


def compute_metrics(all_pred, eval_dataset):
    print(len(all_pred), len(eval_dataset))
    return {'my_metric': len(all_pred[0]) - len(eval_dataset[0])}  # dummy metric


local_rank, world_size = llama.setup_model_parallel()
parser = HfArgumentParser((ModelArguments, DataArguments, CollieArguments))
if sys.argv[-1].endswith(".yaml"):
    model_args, data_args, collie_args = parser.parse_yaml_file(yaml_file=os.path.abspath(sys.argv[-1]))
else:
    model_args, data_args, collie_args = parser.parse_args_into_dataclasses()

model, tokenizer = llama.load_model(
    ckpt_dir=model_args.model_name_or_path,  # 7B, 13B, 30B, 65B
    tokenizer_path=os.path.join(model_args.llama_dir, 'tokenizer.model'),
    local_rank=local_rank,
    world_size=world_size,
    froze_embeddings=False,
    zero=False,
    tensor_parallel=True,
    pipeline_parallel=False,
    max_batch_size=collie_args.per_device_train_batch_size,
    max_seq_len=data_args.max_length,
)
tokenizer.pad_token_id = 0

dataset = load_from_disk(data_args.data_dir)['train'].select(range(1000))
train_dataloader = DataLoader(
    dataset,
    batch_size=collie_args.per_device_train_batch_size,
    collate_fn=lambda x: collate_fn(x, tokenizer, max_len=data_args.max_length)
)
eval_dataloader = DataLoader(
    dataset.select(range(4)),
    batch_size=collie_args.per_device_train_batch_size,
    collate_fn=lambda x: collate_fn(x, tokenizer, max_len=data_args.max_length)
)

trainer = collie.trainer.InplaceTensorTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataloader=train_dataloader,
    eval_dataloader=eval_dataloader,
    eval_dataset=dataset.select(range(4)),
    collie_args=collie_args,
    compute_metrics=compute_metrics,
)
trainer.train()

For more examples, check CoLLiE Examples.

Features

1. Utilizes Colossal-AI's pipeline parallelism mode (requires installing Colossal-AI).
2. Replaces the original attention structure with flash-attention (requires installing flash-attn).
3. Replaces the original dense layer with fused-dense (requires installing fused_dense_lib).
4. Replaces the original rotary position encoding with rotary_emb (requires installing rotary_emb).
5. Employs fairscale's tensor parallelism mode (requires installing fairscale).
6. Utilizes deepspeed's ZeRO (requires installing deepspeed).
7. Support Inplace SGD.

Roadmap

• Utilizes Colossal-AI's pipeline parallelism.
• Utilizes FairScale's tensor parallelism.
• Utilizes Deepspeed's ZeRO.
• Implement Inplace SGD.
• Reimplement LlaMA with Colossal-AI APIs.
• Speed Benchmark.
• Support Colossal-AI's tensor parallelism and ZeRO CPU offload.
• Add more examples.