How to avoid "CUDA out of memory" in PyTorch

Question

I think it's a pretty common message for PyTorch users with low GPU memory:

RuntimeError: CUDA out of memory. Tried to allocate X MiB (GPU X; X GiB total capacity; X GiB already allocated; X MiB free; X cached)

I tried to process an image by loading each layer to GPU and then loading it back:

for m in self.children():
    m.cuda()
    x = m(x)
    m.cpu()
    torch.cuda.empty_cache()

But it doesn't seem to be very effective. I'm wondering is there any tips and tricks to train large deep learning models while using little GPU memory.

Answer 1

While training large deep learning models while using little GPU memory, you can mainly use two ways (apart from the ones discussed in other answers) to avoid CUDA out of memory error.

1. Use Automatic Mixed Precision (AMP) training i.e. use fp16. Mixed precision, in most cases, will retain your model's accuracy, while using less memory and time. Check AMP docs

I did a little experiment of fine-tuning ResNet101 (layer4+fc layers) with AMP for 5 epochs on NVIDIA GeForce RTX 2060 SUPER (Turing), these are the results. For larger workloads, you'll see bigger benefits. By default, FP32 training is used in PyTorch. AMP uses less memory and speeds up the training while maintaining accuracy.

It requires minor changes in your training loop. Check the pytorch docs.

def train(n_epochs, loaders, model, optimizer, criterion, use_amp=False):
    scaler = torch.cuda.amp.GradScaler(enabled=use_amp)  #1 initialize gradient scaler
    
    for epoch in range(1, n_epochs+1):  # epoch
        train_loss, valid_loss = 0.0, 0.0; torch.cuda.synchronize(); start_time = time.time()
        
        model.train()  # set model to training mode
        for batch_idx, (data, target) in enumerate(loaders['train']):  # training iteration
            data, target = data.to(device), target.to(device)
            optimizer.zero_grad()  # zero the accumulated gradients
            with torch.cuda.amp.autocast(enabled=use_amp):  #2 use AMP context
                outputs = model(data)  # forward pass
                loss = criterion(outputs, target)
            scaler.scale(loss).backward()  #3 call backward pass on scaled loss
            scaler.step(optimizer)  #4 unscale gradients, do weight update if not Infs or NaNs
            scaler.update()  #5 update scale factor for next iteration
            train_loss += ((1 / (batch_idx + 1)) * (loss.item() - train_loss))
            
        model.eval()  # set model to evaluation mode
        for batch_idx, (data, target) in enumerate(loaders['valid']):  # validation iteration
            data, target = data.to(device), target.to(device)
            with torch.no_grad():
                with torch.cuda.amp.autocast(enabled=use_amp):  # AMP
                    outputs = model(data)
                    loss = criterion(outputs, target)
                valid_loss += ((1 / (batch_idx + 1)) * (loss.item() - valid_loss))
                
        torch.cuda.synchronize(); end_time = time.time(); total_time = round((end_time - start_time)/60, 2)
        print(f'Epoch: {epoch} \tTraining Loss: {train_loss:.3f} \tValidation Loss: {valid_loss:.3f} \tTime: {total_time}min')
    return model

2. Use gradient accumulation to accumulate gradient over batches to replicate larger batch sizes even when your GPU RAM is not large enough to accommodate large batch sizes. It utilizes that fact that PyTorch, by default, accumulates the gradients unless cleared. For example, If your GPU allows batch size of 1 and you call optimizer.zero_grad() in alternate iteration, then your effective batch size becomes 2. Read more on pytorch forums.

# some code
# Initialize dataset with batch size 10
opt.zero_grad()
for i, (input, target) in enumerate(dataset):
    pred = net(input)
    loss = crit(pred, target)
    # one graph is created here
    loss.backward()
    # graph is cleared here
    if (i+1)%10 == 0:
        # every 10 iterations of batches of size 10
        opt.step()
        opt.zero_grad()

Answer 2

TRY NOT USING REFINER.

None of solutions worked for me beside reducing num_inference_steps to 15. Also generating without refiner. You have to adjust parameters and find out what fits for you.

Answer 3

In the case where right of the bat, before epoch 1 starts, we get the out of memory error,

torch.cuda.empty_cache()
gc.collect()

couple with lower the batch_size may work in some case, as noted by previous answers. In my case it was not enough. I did 2 more things:

import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:1024"

Here you can adjust 1024 to a desired size.

I adjusted the size of the images I was introducing to the network, in the dataset class, particularly in the __getitem__() method:

def __getitem__(self, i_dex, resize_=(320,480)):
        transforms_ = transforms.Compose([
                                        transforms.PILToTensor(),
                        transforms.ConvertImageDtype(torch.float32),
                                        ])
        im_ = Image_.open(self.data_paths[i_dex])
        if im_.mode !='RGB':
            im_ = im_.convert('RGB')
        im_ = im_.resize(resize_)
      
        return transforms_(im_), labels[i_dex]

and reduced the batch_size from 40 to 20. Before resizing the maximum batch_size I was able to run was 4. This is very important for contrastive learning models like the SimCLR where the batch size must be larger (256 or more) such that the model learns from multiple contrastive augmentation image pairs.

Edits: Repeating the process above several times, I was able to train the model on a batch size of 400 eventually.

To monitor GPU resources you can use something like glances . This makes things easier while adjusting parameters.

Answer 4

Though not relevant to the original question, I faced the same issue while using https://github.com/oobabooga/text-generation-webui Bing search results in this particular SO page as the top result. I resolved this by increasing the GPU memory:

Answer 5

Send the batches to CUDA iteratively, and make small batch sizes. Don't send all your data to CUDA at once in the beginning. Rather, do it as follows:

for e in range(epochs):
    for images, labels in train_loader:   
        if torch.cuda.is_available():
            images, labels = images.cuda(), labels.cuda()   
        # blablabla  

Answer 6

I see no one advice wait after collection of garbage. If nothing help you you can try wait befor garbage collected. Try this:

import torch
import time
import gc
from pynvml import nvmlInit, nvmlDeviceGetHandleByIndex, nvmlDeviceGetMemoryInfo

def clear_gpu_memory():
    torch.cuda.empty_cache()
    gc.collect()
    del variables

def wait_until_enough_gpu_memory(min_memory_available, max_retries=10, sleep_time=5):
    nvmlInit()
    handle = nvmlDeviceGetHandleByIndex(torch.cuda.current_device())

    for _ in range(max_retries):
        info = nvmlDeviceGetMemoryInfo(handle)
        if info.free >= min_memory_available:
            break
        print(f"Waiting for {min_memory_available} bytes of free GPU memory. Retrying in {sleep_time} seconds...")
        time.sleep(sleep_time)
    else:
        raise RuntimeError(f"Failed to acquire {min_memory_available} bytes of free GPU memory after {max_retries} retries.")

# Usage example
min_memory_available = 2 * 1024 * 1024 * 1024  # 2GB
clear_gpu_memory()
wait_until_enough_gpu_memory(min_memory_available)

Answer 7

Might seem too simplistic but it worked for me; I just closed my VScode and opened it again and then restarted and ran all the cells.

Answer 8

If you are working with images, just reduce the input image shape. For example, if you are using 512x512, try 256x256. It worked for me!

Answer 9

If you are done training and just want to test with an image, make sure to add a with torch.no_grad() and m.eval() at the beginning:

with torch.no_grad():
  for m in self.children():
    m.cuda()
    m.eval()
    x = m(x)
    m.cpu()
    torch.cuda.empty_cache()

This may seem obvious but it worked on my case. I was trying to use BERT to transform sentences into an embbeding representation. As BERT is a pre-trained model I didn't need to save all the gradients, and they were consuming all the GPU's memory.

Answer 10

I faced the same problem and resolved it by degrading the PyTorch version from 1.10.1 to 1.8.1 with code 11.3. In my case, I am using GPU RTX 3060, which works only with Cuda version 11.3 or above, and when I installed Cuda 11.3, it came with PyTorch 1.10.1. So I degraded the PyTorch version, and now it is working fine.

$ pip3 install torch==1.8.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html

2- You can check by reducing train batch size also.

Answer 11

Although

import torch
torch.cuda.empty_cache()

provides a good alternative for clearing the occupied cuda memory and we can also manually clear the not in use variables by using,

import gc
del variables
gc.collect()

But still after using these commands, the error might appear again because pytorch doesn't actually clears the memory instead clears the reference to the memory occupied by the variables. So reducing the batch_size after restarting the kernel and finding the optimum batch_size is the best possible option (but sometimes not a very feasible one).

Another way to get a deeper insight into the alloaction of memory in gpu is to use:

torch.cuda.memory_summary(device=None, abbreviated=False)

wherein, both the arguments are optional. This gives a readable summary of memory allocation and allows you to figure the reason of CUDA running out of memory and restart the kernel to avoid the error from happening again (Just like I did in my case).

Passing the data iteratively might help but changing the size of layers of your network or breaking them down would also prove effective (as sometimes the model also occupies a significant memory for example, while doing transfer learning).

Answer 12

I meet the same error, and my GPU is GTX1650 with 4g video memory and 16G ram. It worked for me when I reduce the batch_size to 3. Hope this can help you

Answer 13

As long as you don't cross a batch size of 32, you will be fine. Just remember to refresh or restart runtime or else even if you reduce the batch size, you will encounter the same error. I set my batch size to 16, it reduces zero gradients from occurring during my training and the model matches the true function much better. Rather than using a batch size of 4 or 8 which causes the training loss to fluctuate than

Answer 14

There is now a pretty awesome library which makes this very simple: https://github.com/rentruewang/koila

pip install koila

in your code, simply wrap the input with lazy:

from koila import lazy
input = lazy(input, batch=0)

Answer 15

Most things are covered, still will add a little.

If torch gives error as "Tried to allocate 2 MiB" etc. it is a mis-leading message. Actually, CUDA runs out of total memory required to train the model. You can reduce the batch size. Say, even if batch size of 1 is not working (happens when you train NLP models with massive sequences), try to pass lesser data, this will help you confirm that your GPU does not have enough memory to train the model.

Also, Garbage collection and cleaning cache part has to be done again, if you want to re-train the model.

Answer 16

I would recommend using mixed precision training with PyTorch. It can make training way faster and consume less memory.

Take a look at https://spell.ml/blog/mixed-precision-training-with-pytorch-Xuk7YBEAACAASJam.

Answer 17

Although this seems bizarre what I found is there are many sessions running in the background for collab even if we factory reset runtime or we close the tab. I conquered this by clicking on "Runtime" from the menu and then selecting "Manage Sessions". I terminated all the unwanted sessions and I was good to go.

Answer 18

Follow these steps:

1. Reduce train,val,test data
2. Reduce batch size {eg. 16 or 32}
3. Reduce number of model parameters {eg. less than million}

In my case, when I am training common voice dataset in kaggle kernels the same error raises. I delt with reducing training dataset to 20000,batch size to 16 and model parameter to 112K.

Answer 19

There are ways to avoid, but it certainly depends on your GPU memory size:

1. Loading the data in GPU when unpacking the data iteratively,

features, labels in batch:
   features, labels = features.to(device), labels.to(device)

2. Using FP_16 or single precision float dtypes.
3. Try reducing the batch size if you ran out of memory.
4. Use .detach() method to remove tensors from GPU which are not needed.

If all of the above are used properly, PyTorch library is already highly optimizer and efficient.

Answer 20

I have the same error but fix it by resize my images from ~600 to 100 using the lines:

import torchvision.transforms as transforms
transform = transforms.Compose([
    transforms.Resize((100, 100)), 
    transforms.ToTensor()
])

Answer 21

Try not drag your grads too far.

I got the same error when I tried to sum up loss in all batches.

loss =  self.criterion(pred, label)

total_loss += loss

Then I use loss.item instead of loss which requires grads, then solved the problem

loss =  self.criterion(pred, label)

total_loss += loss.item()

The solution below is credited to yuval reina in the kaggle question

This error is related to the GPU memory and not the general memory => @cjinny comment might not work.
Do you use TensorFlow/Keras or Pytorch?
Try using a smaller batch size.
If you use Keras, Try to decrease some of the hidden layer sizes.
If you use Pytorch:
do you keep all the training data on the GPU all the time?
make sure you don't drag the grads too far
check the sizes of you hidden layer

Answer 22

Best way would be lowering down the batch size. Usually it works. Otherwise try this:

import gc

del variable #delete unnecessary variables 
gc.collect()

Answer 23

Implementation:

1. Feed the image into gpu batch by batch.

2. Using a small batch size during training or inference.

3. Resize the input images with a small image size.

Technically:

1. Most networks are over parameterized, which means they are too large for the learning tasks. So finding an appropriate network structure can help:

a. Compact your network with techniques like model compression, network pruning and quantization.

b. Directly using a more compact network structure like mobileNetv1/2/3.

c. Network architecture search(NAS).

Answer 24

Just reduce the batch size, and it will work. While I was training, it gave following error:

CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 10.76 GiB total capacity; 4.29 GiB already allocated; 10.12 MiB free; 4.46 GiB reserved in total by PyTorch)

And I was using batch size of 32. So I just changed it to 15 and it worked for me.