Parameters: nlp.batch_size and components.transformer.max_batch_size no effect on VRAM usage

[whranclast]

Hello everyone,

I am currently experiencing constantly GPU OOM issues as some of my data that I train/test with grew significantly. I've experimented with adjusting the batch_size and max_batch_size to very low number (4 and 32 accordingly), however absolutely no change in the usage of VRAM. I currently training on NVIDIA A10G with 24GB of VRAM. The only way to reduce the usage of VRAM is to lower significantly my test set.

As I've read here #8600 I've experimented also with max_lenght but that did nothing for me.

I've also read this interesting response by @mbru that no matter the config settings, the entire dev corpus would be loaded into the VRAM at least once. There must be a workaround on that as I imagine people are training with much more data than me. My total amount of data is around 60 000 articles, on average with length of 800-1000 characters, where I have some that are 16 000 characters but also some that are 200 characters.

I also tried with manually lowering the max_batch_size from pytorch with os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:124" and what did this cause was after my GPU went to 20GB of usage, it dropped to 4GB of usage and just stayed there, however if I use again the entire test set I still run out of memory.

I've experimented quite a bit but unfortunately whatever I change on the config had no effect. I can also say that I am using a fine-tuned transformer based on transformer = SentenceTransformer('nli-distilroberta-base-v2')

Some more information about my system:
Python version: 3.9.14
Spacy version: 3.6.0
Operating system: Ubuntu 22.04

This is the config file I am currently working with:

train = null
dev = null
vectors = null
init_tok2vec = null

[system]
gpu_allocator = "pytorch"
seed = 0

[nlp]
lang = "en"
pipeline = ["transformer","ner"]
batch_size = 4
disabled = []
before_creation = null
after_creation = null
after_pipeline_creation = null
tokenizer = {"@tokenizers":"spacy.Tokenizer.v1"}

[components]

[components.ner]
factory = "ner"
incorrect_spans_key = null
moves = null
scorer = {"@scorers":"spacy.ner_scorer.v1"}
update_with_oracle_cut_size = 128

[components.ner.model]
@architectures = "spacy.TransitionBasedParser.v2"
state_type = "ner"
extra_state_tokens = false
hidden_width = 64
maxout_pieces = 2
use_upper = false
nO = null

[components.ner.model.tok2vec]
@architectures = "spacy-transformers.TransformerListener.v1"
grad_factor = 1.0
pooling = {"@layers":"reduce_mean.v1"}
upstream = "*"

[components.transformer]
factory = "transformer"
max_batch_items = 32
set_extra_annotations = {"@annotation_setters":"spacy-transformers.null_annotation_setter.v1"}

[components.transformer.model]
@architectures = "spacy-transformers.TransformerModel.v3"
name = "transformer/current/transformer_model"
mixed_precision = true

[components.transformer.model.get_spans]
@span_getters = "spacy-transformers.strided_spans.v1"
window = 128
stride = 96

[components.transformer.model.tokenizer_config]
use_fast = true

[corpora]

[corpora.dev]
@readers = "spacy.Corpus.v1"
path = ${paths.dev}
# max_length = 1200
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[corpora.train]
@readers = "spacy.Corpus.v1"
path = ${paths.train}
max_length = 0
gold_preproc = false
limit = 0
augmenter = null

[training]
accumulate_gradient = 3
dev_corpus = "corpora.dev"
train_corpus = "corpora.train"
seed = ${system.seed}
gpu_allocator = ${system.gpu_allocator}
dropout = 0.1
patience = 8000
max_epochs = 0
max_steps = 30000
eval_frequency = 800
frozen_components = []
before_to_disk = null
annotating_components = []

[training.batcher]
@batchers = "spacy.batch_by_padded.v1"
discard_oversize = true
size = 256
buffer = 128
get_length = null

[training.logger]
@loggers = "spacy.ConsoleLogger.v1"
progress_bar = false

[training.optimizer]
@optimizers = "Adam.v1"
beta1 = 0.9
beta2 = 0.999
L2_is_weight_decay = true
L2 = 0.01
grad_clip = 1.0
use_averages = false
eps = 0.00000001

[training.optimizer.learn_rate]
@schedules = "warmup_linear.v1"
warmup_steps = 250
total_steps = 30000
initial_rate = 0.00005

[training.score_weights]
ents_f = 1
ents_p = 0
ents_r = 0
ents_per_type = null

[pretraining]

[initialize]
vectors = null
init_tok2vec = ${paths.init_tok2vec}
vocab_data = null
lookups = null
before_init = null
after_init = null

[initialize.components]

[initialize.tokenizer]