multi gpu use?

[ganeshkrishnan1]

I am running out of memory on Tesla T4. I have 4 of them though and I usually use accelerator for multigpu setup. How can I use them for angle semantic similarity?

[SeanLee97]

Do you use it for training or inference?

[ganeshkrishnan1]

I used it for training. It looks like the script does use multiple gpu but it runs out of memory due to high batch size. I will close this ticket

[ganeshkrishnan1]

I ran this on lower batch count and I can see the trainer never uses more than 1 GPU

---

[ganeshkrishnan1]

I used the example provided and also put accelerator but both of them fails to use more than 1 GPU. Any suggestions?

[SeanLee97]

hi @ganeshkrishnan1 , could you provide the training script?

[SeanLee97]

Here is one example that can be successfully run on multi gpus:

CUDA_VISIBLE_DEVICES=0,1 WANDB_MODE=disabled torchrun --nproc_per_node=2 --master_port=2345 train_cli.py \
--model_name_or_path mixedbread-ai/mxbai-embed-large-v1 \
--train_name_or_path ./snli_5k.jsonl --save_dir mxbai-snli-ckpts \
--w1 0. --w2 20.0 --w3 1.0 --angle_tau 20.0 --learning_rate 3e-6 --maxlen 64 \
--pooling_strategy cls \
--epochs 1 \
--batch_size 32 \
--logging_steps 100 \
--warmup_steps 200 \
--save_steps 1000 --seed 42 --gradient_accumulation_steps 2 --fp16 1 --torch_dtype 'float32'

train_cli.py is from: https://github.com/SeanLee97/AnglE/blob/main/angle_emb/train_cli.py

data format:

$ head -3 snli_5k.jsonl

{"text": "A person on a horse jumps over a broken down airplane.", "positive": "A person is outdoors, on a horse.", "negative": "A person is at a diner, ordering an omelette."}
{"text": "Children smiling and waving at camera", "positive": "There are children present", "negative": "The kids are frowning"}
{"text": "A boy is jumping on skateboard in the middle of a red bridge.", "positive": "The boy does a skateboarding trick.", "negative": "The boy skates down the sidewalk."}

[ganeshkrishnan1]

This is my python code:
I experimented with accelerator, then torch distributed and also added to(device).
I will try with your method and see if it works out with 4 gpus.

from sentence_transformers import InputExample, losses, SentenceTransformer
from torch import optim
from sentence_transformers import SentenceTransformer, models, losses
import torch
from datasets import load_dataset,Dataset, DatasetDict
from angle_emb import AnglE, AngleDataTokenizer
from accelerate import FullyShardedDataParallelPlugin, Accelerator
from torch.distributed.fsdp.fully_sharded_data_parallel import FullOptimStateDictConfig, FullStateDictConfig
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, StateDictType, FullStateDictConfig


fsdp_plugin = FullyShardedDataParallelPlugin(
    state_dict_config=FullStateDictConfig(offload_to_cpu=True, rank0_only=False),
    optim_state_dict_config=FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=False),
)

accelerator = Accelerator(fsdp_plugin=fsdp_plugin)
device = accelerator.device

train_json_file = './brand_search_term_con_new.json_9.json'  # JSON file for training data
full_dataset  = load_dataset('json', data_files=train_json_file,  split='train')
desired_test_size = 5000
# # Calculate the training set size
train_size = len(full_dataset) - desired_test_size
# # Split the dataset into training and evaluation sets
split_datasets = full_dataset.train_test_split(test_size=desired_test_size, train_size=train_size)
dataset_dict = DatasetDict({
    'train': split_datasets['train'],
    'test': split_datasets['test']
})

# 1. load pretrained model
angle = AnglE.from_pretrained('SeanLee97/angle-bert-base-uncased-nli-en-v1', max_length=512, pooling_strategy='cls',device_map = 'auto').to(device)
# # 3. transform data
train_ds = dataset_dict['train'].shuffle().map(AngleDataTokenizer(angle.tokenizer, angle.max_length), num_proc=16)
test_ds = dataset_dict['test'].map(AngleDataTokenizer(angle.tokenizer, angle.max_length), num_proc=16)
# angle, train_ds, test_ds = accelerator.prepare(angle, train_ds, test_ds)
angle.to(device)
# # 4. fit
angle.fit(
    train_ds=train_ds,
    valid_ds=test_ds,
    output_dir='trainedmodel/aihello-model',
    batch_size=8,
    epochs=2,
    learning_rate=2e-5,
    save_steps=5000,
    eval_steps=5000,
    warmup_steps=100,
    gradient_accumulation_steps=4,
    loss_kwargs={
        'w1': 1.0,
        'w2': 1.0,
        'w3': 1.0,
        'cosine_tau': 20,
        'ibn_tau': 20,
        'angle_tau': 1.0
    },
    fp16=True,
    logging_steps=100
)
corrcoef, accuracy = angle.evaluate(test_ds, device=angle.device)
print('corrcoef:', corrcoef)

[ganeshkrishnan1]

The shell script worked and I got the checkpoint as well with multiple GPUs.

Python code didn't use the multiple GPU though.

[SeanLee97]

I haven't tried multiGPU in python code, just used it supported by Transformers Trainer.

BTW, here are some tips to improve the model:

1. if your dataset is FormatA: {'text1': "", "text2": "", "label": float or int}, it is better slightly increase weight for w1.

2. if your dataset is FormatB: {'text': "", "positive": "", "negative": ""}, the suggested parameters are w1=0, w2=20, w3=1.0, angle_tau=20.0

[ganeshkrishnan1]

Thanks for the tip about the w. I am using DataFormat C.

eg
{"text": "Cool Spot 11x11 Pop-Up Instant Gazebo Tent with Mosquito Netting Outdoor Canopy Shelter with 121 Square Feet of Shade by COOS BAY (Beige)", "positive": "outdoor tent canopy"}

Should I use the same as B?

[SeanLee97]

DataFormats.C is okay. However, DataFormats.B is recommended since it can improve performance more significantly.

BTW, here are the tips, we will push it in the next version.

[ganeshkrishnan1]

Negative is very hard to generate from unlabelled text for DataSet B. We have "product title" -> "search term" as positive correlation but there is no correct way to generate negative

Like you mentioned, the performance of Dataset C on training from sample was not as good as I wanted it to be. I am running the trainer on our whole dataset of 200m records and report back on performance (~15 days)

[SeanLee97]

Negative is very hard to generate from unlabelled text for DataSet B. We have "product title" -> "search term" as positive correlation but there is no correct way to generate negative

Like you mentioned, the performance of Dataset C on training from sample was not as good as I wanted it to be. I am running the trainer on our whole dataset of 200m records and report back on performance (~15 days)

For such large datasets, it is better to specify a small learning_rate such as 1e-6, and specify --fixed_teacher_name_or_path to alleviate information forgetting.