Performance and Scalability: How To Fit a Bigger Model and Train It Faster

Or how to escape the dreaded “RuntimeError: CUDA error: out of memory” error.

Training ever larger models can become challenging even on modern GPUs. Due to their immense size we often run out of GPU memory and training can take very long. In this section we have a look at a few tricks to reduce the memory footprint and speed up training for large models and how they are integrated in the Trainer and 🤗 Accelerate. Before we start make sure you have installed the following libraries:

pip install transformers datasets accelerate nvidia-ml-py3

The nvidia-ml-py3 library allows us to monitor the memory usage of the models from within Python. You might be familiar with the nvidia-smi command in the terminal - this library allows to access the same information in Python directly.

Then we create some dummy data. We create random token IDs between 100 and 30000 and binary labels for a classifier. In total we get 512 sequences each with length 512 and store them in a Dataset with PyTorch format.

import numpy as np
from datasets import Dataset


seq_len, dataset_size = 512, 512
dummy_data = {
    "input_ids": np.random.randint(100, 30000, (dataset_size, seq_len)),
    "labels": np.random.randint(0, 1, (dataset_size)),
}
ds = Dataset.from_dict(dummy_data)
ds.set_format("pt")

We want to print some summary statistics for the GPU utilization and the training run with the Trainer. We setup a two helper functions to do just that:

from pynvml import *


def print_gpu_utilization():
    nvmlInit()
    handle = nvmlDeviceGetHandleByIndex(0)
    info = nvmlDeviceGetMemoryInfo(handle)
    print(f"GPU memory occupied: {info.used//1024**2} MB.")


def print_summary(result):
    print(f"Time: {result.metrics['train_runtime']:.2f}")
    print(f"Samples/second: {result.metrics['train_samples_per_second']:.2f}")
    print_gpu_utilization()

Let’s verify that we start with a free GPU memory:

>>> print_gpu_utilization()
GPU memory occupied: 0 MB.

That looks good: the GPU memory is not occupied as we would expect before we load any models. If that’s not the case on your machine make sure to stop all processes that are using GPU memory. However, not all free GPU memory can be used by the user. When a model is loaded to the GPU also the kernels are loaded which can take up 1-2GB of memory. To see how much it is we load a tiny tensor into the GPU which triggers the kernels to be loaded as well.

>>> import torch


>>> torch.ones((1, 1)).to("cuda")
>>> print_gpu_utilization()
GPU memory occupied: 1343 MB.

We see that the kernels alone take up 1.3GB of GPU memory. Now let’s see how much space the model uses.

Load Model

First, we load the bert-large-uncased model. We load the model weights directly to the GPU so that we can check how much space just weights use.

>>> from transformers import AutoModelForSequenceClassification


>>> model = AutoModelForSequenceClassification.from_pretrained("bert-large-uncased").to("cuda")
>>> print_gpu_utilization()
GPU memory occupied: 2631 MB.

We can see that the model weights alone take up 1.3 GB of the GPU memory. The exact number depends on the specific GPU you are using. Note that on newer GPUs a model can sometimes take up more space since the weights are loaded in an optimized fashion that speeds up the usage of the model. Now we can also quickly check if we get the same result as with nvidia-smi CLI:

nvidia-smi

Tue Jan 11 08:58:05 2022       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03    Driver Version: 460.91.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  On   | 00000000:00:04.0 Off |                    0 |
| N/A   37C    P0    39W / 300W |   2631MiB / 16160MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                            
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      3721      C   ...nvs/codeparrot/bin/python     2629MiB |
+-----------------------------------------------------------------------------+

We get the same number as before and you can also see that we are using a V100 GPU with 16GB of memory. So now we can start training the model and see how the GPU memory consumption changes. First, we set up a few standard training arguments that we will use across all our experiments:

default_args = {
    "output_dir": "tmp",
    "evaluation_strategy": "steps",
    "num_train_epochs": 1,
    "log_level": "error",
    "report_to": "none",
}

Vanilla Training

As a first experiment we will use the Trainer and train the model without any further modifications and a batch size of 4:

from transformers import TrainingArguments, Trainer, logging

logging.set_verbosity_error()


training_args = TrainingArguments(per_device_train_batch_size=4, **default_args)
trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 57.82
Samples/second: 8.86
GPU memory occupied: 14949 MB.

We see that already a relatively small batch size almost fills up our GPU’s entire memory. However, a larger batch size can often result in faster model convergence or better end performance. So ideally we want to tune the batch size to our model’s needs and not to the GPU limitations. A simple trick to effectively train larger batch size is gradient accumulation.

Gradient Accumulation

The idea behind gradient accumulation is to instead of calculating the gradients for the whole batch at once to do it in smaller steps. The way we do that is to calculate the gradients iteratively in smaller batches by doing a forward and backward pass through the model and accumulating the gradients in the process. When enough gradients are accumulated we run the model’s optimization step. This way we can easily increase the overall batch size to numbers that would never fit into the GPU’s memory. In turn, however, the added forward and backward passes can slow down the training a bit.

We can use gradient accumulation in the Trainer by simply adding the gradient_accumulation_steps argument to TrainingArguments. Let’s see how it impacts the models memory footprint:

training_args = TrainingArguments(per_device_train_batch_size=1, gradient_accumulation_steps=4, **default_args)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 66.03
Samples/second: 7.75
GPU memory occupied: 8681 MB.

We can see that the memory footprint was dramatically reduced at the cost of being only slightly slower than the vanilla run. Of course, this would change as you increase the number of accumulation steps. In general you would want to max out the GPU usage as much as possible. So in our case, the batch_size of 4 was already pretty close to the GPU’s limit. If we wanted to train with a batch size of 64 we should not use per_device_train_batch_size=1 and gradient_accumulation_steps=64 but instead per_device_train_batch_size=4 and gradient_accumulation_steps=16 which has the same effective batch size while making better use of the available GPU resources.

Next we have a look at another trick to save a little bit more GPU memory called gradient checkpointing.

Gradient Checkpointing

Even when we set the batch size to 1 and use gradient accumulation we can still run out of memory when working with large models. In order to compute the gradients during the backward pass all activations from the forward pass are normally saved. This can create a big memory overhead. Alternatively, one could forget all activations during the forward pass and recompute them on demand during the backward pass. This would however add a significant computational overhead and slow down training.

Gradient checkpointing strikes a compromise between the two approaches and saves strategically selected activations throughout the computational graph so only a fraction of the activations need to be re-computed for the gradients. See this great article explaining the ideas behind gradient checkpointing.

To enable gradient checkpointing in the Trainer we only need ot pass it as a flag to the TrainingArguments. Everything else is handled under the hood:

training_args = TrainingArguments(
    per_device_train_batch_size=1, gradient_accumulation_steps=4, gradient_checkpointing=True, **default_args
)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 85.47
Samples/second: 5.99
GPU memory occupied: 6775 MB.

We can see that this saved some more memory but at the same time training became a bit slower. A general rule of thumb is that gradient checkpointing slows down training by about 20%. Let’s have a look at another method with which we can regain some speed: mixed precision training.

FP16 Training

The idea of mixed precision training is that no all variables need to be stored in full (32-bit) floating point precision. If we can reduce the precision the variales and their computations are faster. The main advantage comes from saving the activations in half (16-bit) precision. Although the gradients are also computed in half precision they are converted back to full precision for the optimization step so no memory is saved here. Since the model is present on the GPU in both 16-bit and 32-bit precision this can use more GPU memory (1.5x the original model is on the GPU), especially for small batch sizes. Since some computations are performed in full and some in half precision this approach is also called mixed precision training. Enabling mixed precision training is also just a matter of setting the fp16 flag to True:

training_args = TrainingArguments(per_device_train_batch_size=4, fp16=True, **default_args)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 27.46
Samples/second: 18.64
GPU memory occupied: 13939 MB.

We can see that this is almost twice as fast as the vanilla training. Let’s add it to the mix of the previous methods:

training_args = TrainingArguments(
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    fp16=True,
    **default_args,
)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 50.76
Samples/second: 10.09
GPU memory occupied: 7275 MB.

We can see that with these tweaks we use about half the GPU memory as at the beginning while also being slightly faster. But we are not done, yet! There is another area where we can save GPU memory: the optimizer.

Optimizer

The most common optimizer used to train transformer model is Adam or AdamW (Adam with weight decay). Adam achieves good convergence by storing the rolling average of the previous gradients which, however, adds an additional memory footprint of the order of the number of model parameters. One remedy to this is to use an alternative optimizer such as Adafactor.

Adafactor

Instead of keeping the rolling average for each element in the weight matrices Adafactor only stores aggregated information (row- and column-wise sums of the rolling averages) which reduces the footprint considerably. One downside of Adafactor is that in some instances convergence can be slower than Adam’s so some experimentation is advised here. We can use Adafactor simply by setting optim="adafactor":

training_args = TrainingArguments(per_device_train_batch_size=4, optim="adafactor", **default_args)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 64.31
Samples/second: 7.96
GPU memory occupied: 12295 MB.

We can see that this saves a few more GB on the GPU. Let’s see how it looks when we add it to the other methods we introduced earlier:

training_args = TrainingArguments(
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    fp16=True,
    optim="adafactor",
    **default_args,
)

trainer = Trainer(model=model, args=training_args, train_dataset=ds)
result = trainer.train()
print_summary(result)

Time: 56.54
Samples/second: 9.06
GPU memory occupied: 4847 MB.

We went from 15 GB memory usage to 5 GB - a 3x improvement while maintaining the throughput! However, as mentioned before, the convergence of Adafactor can be worse than Adam. There is an alternative to Adafactor called 8-bit Adam that takes a slightly different approach.

8-bit Adam

Instead of aggregating optimizer states like Adafactor, 8-bit Adam keeps the full state and quantizes it. Quantization means that it stores the state with lower precision and dequantizes it only for the optimization. This is similar to the idea behind FP16 training where using variables with lower precision saves memory.

In contrast to the previous approaches is this one not integrated into the Trainer as a simple flag. We need to install the 8-bit optimizer and then pass it as a custom optimizer to the Trainer. Follow the installation guide in the Github repo to install the bitsandbytes library that implements the 8-bit Adam optimizer.

Once installed, we just need to initialize the the optimizer. Although this looks like a considerable amount of work it actually just involves two steps: first we need to group the model’s parameters into two groups where to one group we apply weight decay and to the other we don’t. Usually, biases and layer norm parameters are not weight decayed. Then in a second step we just do some argument housekeeping to use the same parameters as the previously used AdamW optimizer.

import bitsandbytes as bnb
from torch import nn
from transformers.trainer_pt_utils import get_parameter_names

training_args = TrainingArguments(per_device_train_batch_size=4, **default_args)

decay_parameters = get_parameter_names(model, [nn.LayerNorm])
decay_parameters = [name for name in decay_parameters if "bias" not in name]
optimizer_grouped_parameters = [
    {
        "params": [p for n, p in model.named_parameters() if n in decay_parameters],
        "weight_decay": training_args.weight_decay,
    },
    {
        "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
        "weight_decay": 0.0,
    },
]

optimizer_kwargs = {
    "betas": (training_args.adam_beta1, training_args.adam_beta2),
    "eps": training_args.adam_epsilon,
}
optimizer_kwargs["lr"] = training_args.learning_rate
adam_bnb_optim = bnb.optim.Adam8bit(
    optimizer_grouped_parameters,
    betas=(training_args.adam_beta1, training_args.adam_beta2),
    eps=training_args.adam_epsilon,
    lr=training_args.learning_rate,
)

We can now pass the custom optimizer as an argument to the Trainer:

trainer = Trainer(model=model, args=training_args, train_dataset=ds, optimizers=(adam_bnb_optim, None))
result = trainer.train()
print_summary(result)

Time: 55.95
Samples/second: 9.15
GPU memory occupied: 13085 MB.

We can see that we get a similar memory improvement as with Adafactor while keeping the full rolling average of the gradients. Let’s repeat the experiment with the full settings:

training_args = TrainingArguments(
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    fp16=True,
    **default_args,
)

trainer = Trainer(model=model, args=training_args, train_dataset=ds, optimizers=(adam_bnb_optim, None))
result = trainer.train()
print_summary(result)

Time: 49.46
Samples/second: 10.35
GPU memory occupied: 5363 MB.

Again, we get about a 3x memory improvement and even slightly higher throughput as using Adafactor. So we have seen how we can optimize the memory footprint of large models. The following plot summarizes all our experiments:

Using 🤗 Accelerate

So far we have used the Trainer to run the experiments but a more flexible alternative to that approach is to use 🤗 Accelerate. With 🤗 Accelerate you have full control over the training loop and can essentially write the loop in pure PyTorch with some minor modifications. In turn it allows you to easily scale across different infrastructures such as CPUs, GPUs, TPUs, or distributed multi-GPU setups without changing any code. Let’s see what it takes to implement all of the above tweaks in 🤗 Accelerate. We can still use the TrainingArguments to wrap the training settings:

training_args = TrainingArguments(
    per_device_train_batch_size=1,
    gradient_accumulation_steps=4,
    gradient_checkpointing=True,
    fp16=True,
    **default_args,
)

The full example training loop with 🤗 Accelerate is only a handful of lines of code long:

from accelerate import Accelerator
from torch.utils.data.dataloader import DataLoader

dataloader = DataLoader(ds, batch_size=training_args.per_device_train_batch_size)

if training_args.gradient_checkpointing:
    model.gradient_checkpointing_enable()

accelerator = Accelerator(fp16=training_args.fp16)
model, optimizer, dataloader = accelerator.prepare(model, adam_bnb_optim, dataloader)

model.train()
for step, batch in enumerate(dataloader, start=1):
    loss = model(**batch).loss
    loss = loss / training_args.gradient_accumulation_steps
    accelerator.backward(loss)
    if step % training_args.gradient_accumulation_steps == 0:
        optimizer.step()
        optimizer.zero_grad()

First we wrap the dataset in a DataLoader. Then we can enable gradient checkpointing by calling the model’s gradient_checkpointing_enable() method. When we initialize the Accelerator we can specifiy if we want to use mixed precision training and it will take care of it for us in the prepare call. During the prepare call the dataloader will also be distributed across workers should we use multiple GPUs. We use the same 8-bit optimizer from the earlier experiments.

Finally, we can write the main training loop. Note that the backward call is handled by 🤗 Accelerate. We can also see how gradient accumulation works: we normalize the loss so we get the average at the end of accumulation and once we have enough steps we run the optimization. Now the question is: does this use the same amount of memory as the previous steps? Let’s check:

>>> print_gpu_utilization()
GPU memory occupied: 5363 MB.

Indeed it does. Implementing these optimization techniques with 🤗 Accelerate only takes a handful of lines of code and comes with the benefit of more flexiblity in the training loop.

Now, let’s take a step back and discuss what we should optimize for when scaling the training of large models.

How to scale

When we train models there are a two aspects we want to optimize at the same time:

• Data throughput/training time
• Model performance

We have seen that each method changes the memory usage and throughput. In general we want to maximize the throughput (samples/second) to minimize the training cost. This is generally achieved by utilizing the GPU as much as possible and thus filling GPU memory to its limit. For example, as mentioned earlier, we only employ gradient accumulation when we want to use a batch size beyond the size of the GPU memory. If the desired batch size fits into memory then there is no reason to apply gradient accumulation which will only slow down training.

The second objective is model performance. Just because we can does not mean we should use a large batch size. As part of hyperparameter tuning you should determine which batch size yields the best result and then optimize the throughput accordingly.

Sometimes, even when applying all the above tweaks the throughput on a given GPU might still not be good enough. One easy solution is to change the type of GPU. For example switching from let’s say a K80 (which you typically get on Google Colab) to a fancier GPU such as the V100 or A100. Although they are more expensive they are usually more cost effective than cheaper GPUs due to their larger memory and faster architecture. For some applications, such as pretraining, this might still not be fast enough. In this case you want to scale your experiment to several GPUs.

Multi-GPU Training

If your model fits on a single GPU scaling to many GPUs can be achieved fairly easily with data parallelism. The idea is very similar to gradient accumulation with the distinction that instead of running the forward and backward passes during the accumulation in sequence on a single machine they are performed in parallel on multiple machines. So each GPU gets a small batch, runs the forward and backward passes and then the gradients from all machines are aggregated and the model is optimized. You can combine this with all the methods we described before. For example, if you have 4 GPUs and use per_device_train_batch_size=12 and gradient_accumulation_steps=3 you will have an effective batch size of 4*12*3=144.

The Trainer allows for distributed training and if you execute your Trainer training script on a machine with multiple GPUs it will automatically utilize all of them, hence the name per_device_train_batch_size. In 🤗 Accelerate you can configure the infrastructure setup with the following command:

accelerate config

Until now we have opperated under the assumption that we can fit the model onto a single GPU without or with the introduced tricks . But what if this is not possible? We still have a few tricks up our sleeves!

What if my model still does not fit?

If the model does not fit on a single GPU with all the mentioned tricks there are still more methods we can apply although life starts to get a bit more complicated. This usually involves some form of pipeline or tensor parallelism where the model itself is distributed across several GPUs. One can also make use of DeepSpeed which implements some of these parallelism strategies along with some more optimization to reduce the memory footprint such as partitioning the optimizer states. You can read more about this in the “Model Parallelism” section.

This concludes the practical part of this guide for scaling the training of large models. The following section goes into more details on some of the aspects discussed above.

Further discussions

This section gives brief ideas on how to make training faster and support bigger models. Later sections will expand, demonstrate and elucidate each of these.

Faster Training

Hardware:

• fast connectivity between GPUs
  • intra-node: NVLink
  • inter-node: Infiniband / Intel OPA

Software:

• Data Parallel / Distributed Data Parallel
• fp16 (autocast caching)

Bigger Models

Hardware:

• bigger GPUs
• more GPUs
• more CPU and NVMe (offloaded to by DeepSpeed-Infinity)

Software:

• Model Scalability (ZeRO and 3D Parallelism)
• Low-memory Optimizers
• fp16/bf16 (smaller data/faster throughput)
• tf32 (faster throughput)
• Gradient accumulation
• Gradient checkpointing
• Sparsity

Hardware

Power and Cooling

If you bought an expensive high end GPU make sure you give it the correct power and sufficient cooling.

Power:

Some high end consumer GPU cards have 2 and sometimes 3 PCI-E 8-Pin power sockets. Make sure you have as many independent 12V PCI-E 8-Pin cables plugged into the card as there are sockets. Do not use the 2 splits at one end of the same cable (also known as pigtail cable). That is if you have 2 sockets on the GPU, you want 2 PCI-E 8-Pin cables going from your PSU to the card and not one that has 2 PCI-E 8-Pin connectors at the end! You won’t get the full performance out of your card otherwise.

Each PCI-E 8-Pin power cable needs to be plugged into a 12V rail on the PSU side and can supply up to 150W of power.

Some other cards may use a PCI-E 12-Pin connectors, and these can deliver up to 500-600W of power.

Low end cards may use 6-Pin connectors, which supply up to 75W of power.

Additionally you want the high-end PSU that has stable voltage. Some lower quality ones may not give the card the stable voltage it needs to function at its peak.

And of course the PSU needs to have enough unused Watts to power the card.

Cooling:

When a GPU gets overheated it would start throttling down and will not deliver full performance. And it will shutdown if it gets too hot.

It’s hard to tell the exact best temperature to strive for when a GPU is heavily loaded, but probably anything under +80C is good, but lower is better - perhaps 70-75C is an excellent range to be in. The throttling down is likely to start at around 84-90C. But other than throttling performance a prolonged very higher temperature is likely to reduce the lifespan of a GPU.

Multi-GPU Connectivity

If you use multiple GPUs the way cards are inter-connected can have a huge impact on the total training time.

If the GPUs are on the same physical node, you can run:

nvidia-smi topo -m

and it will tell you how the GPUs are inter-connected.

On a machine with dual-GPU and which are connected with NVLink, you will most likely see something like:

        GPU0    GPU1    CPU Affinity    NUMA Affinity
GPU0     X      NV2     0-23            N/A
GPU1    NV2      X      0-23            N/A

on a different machine w/o NVLink we may see:

        GPU0    GPU1    CPU Affinity    NUMA Affinity
GPU0     X      PHB     0-11            N/A
GPU1    PHB      X      0-11            N/A

The report includes this legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

So the first report NV2 tells us the GPUs are interconnected with 2 NVLinks, and the second report PHB we have a typical consumer-level PCIe+Bridge setup.

Check what type of connectivity you have on your setup. Some of these will make the communication between cards faster (e.g. NVLink), others slower (e.g. PHB).

Depending on the type of scalability solution used, the connectivity speed could have a major or a minor impact. If the GPUs need to sync rarely, as in DDP, the impact of a slower connection will be less significant. If the GPUs need to send messages to each other often, as in ZeRO-DP, then faster connectivity becomes super important to achieve faster training.

NVlink

NVLink is a wire-based serial multi-lane near-range communications link developed by Nvidia.

Each new generation provides a faster bandwidth, e.g. here is a quote from Nvidia Ampere GA102 GPU Architecture:

Third-Generation NVLink® GA102 GPUs utilize NVIDIA’s third-generation NVLink interface, which includes four x4 links, with each link providing 14.0625 GB/sec bandwidth in each direction between two GPUs. Four links provide 56.25 GB/sec bandwidth in each direction, and 112.5 GB/sec total bandwidth between two GPUs. Two RTX 3090 GPUs can be connected together for SLI using NVLink. (Note that 3-Way and 4-Way SLI configurations are not supported.)

So the higher X you get in the report of NVX in the output of nvidia-smi topo -m the better. The generation will depend on your GPU architecture.

Let’s compare the execution of a gpt2 language model training over a small sample of wikitext.

The results are:

You can see that NVLink completes the training ~23% faster.

In the second benchmark we use NCCL_P2P_DISABLE=1 to tell the GPUs not to use NVLink.

Here is the full benchmark code and outputs:

# DDP w/ NVLink

rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train \
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200

{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}

# DDP w/o NVLink

rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 NCCL_P2P_DISABLE=1 python -m torch.distributed.launch \
--nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py --model_name_or_path gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 --do_train
--output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200

{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}

Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (NV2 in nvidia-smi topo -m) Software: pytorch-1.8-to-be + cuda-11.0 / transformers==4.3.0.dev0

Software

Model Scalability

When you can’t fit a model into the available GPU memory, you need to start using a solution that allows you to scale a large model to use multiple GPUs in parallel.

For indepth details on ZeRO and various other model parallelism protocols please see: Model Parallelism

Anatomy of Model's Operations

Transformers architecture includes 3 main groups of operations grouped below by compute-intensity.

1. Tensor Contractions

   Linear layers and components of Multi-Head Attention all do batched matrix-matrix multiplications. These operations are the most compute-intensive part of training a transformer.

2. Statistical Normalizations

   Softmax and layer normalization are less compute-intensive than tensor contractions, and involve one or more reduction operations, the result of which is then applied via a map.

3. Element-wise Operators

   These are the remaining operators: biases, dropout, activations, and residual connections. These are the least compute-intensive operations.

This knowledge can be helpful to know when analyzing performance bottlenecks.

This summary is derived from Data Movement Is All You Need: A Case Study on Optimizing Transformers 2020

Anatomy of Model's Memory

The components on GPU memory are the following:

1. model weights
2. optimizer states
3. gradients
4. forward activations saved for gradient computation
5. temporary buffers
6. functionality-specific memory

A typical model trained in mixed precision with AdamW requires 18 bytes per model parameter plus activation memory.

For inference there are no optimizer states and gradients, so we can subtract those. And thus we end up with 6 bytes per model parameter for mixed precision inference, plus activation memory.

Let’s look at the details.

Model Weights

• 4 bytes * number of parameters for fp32 training
• 6 bytes * number of parameters for mixed precision training

Optimizer States

• 8 bytes * number of parameters for normal AdamW (maintains 2 states)
• 2 bytes * number of parameters for 8-bit AdamW optimizers like bitsandbytes
• 4 bytes * number of parameters for optimizers like SGD (maintains only 1 state)

Gradients

• 4 bytes * number of parameters for either fp32 or mixed precision training

Forward Activations

• size depends on many factors, the key ones being sequence length, hidden size and batch size.

There are the input and output that are being passed and returned by the forward and the backward functions and the forward activations saved for gradient computation.

Temporary Memory

Additionally there are all kinds of temporary variables which get released once the calculation is done, but in the moment these could require additional memory and could push to OOM. Therefore when coding it’s crucial to think strategically about such temporary variables and sometimes to explicitly free those as soon as they are no longer needed.

Functionality-specific memory

Then your software could have special memory needs. For example, when generating text using beam search, the software needs to maintain multiple copies of inputs and outputs.

forward vs backward Execution Speed

For convolutions and linear layers there are 2x flops in the backward compared to the forward, which generally translates into ~2x slower (sometimes more, because sizes in the backward tend to be more awkward). Activations are usually bandwidth-limited, and it’s typical for an activation to have to read more data in the backward than in the forward (e.g. activation forward reads once, writes once, activation backward reads twice, gradOutput and output of the forward, and writes once, gradInput).

Floating Data Types

Here are the commonly used floating point data types choice of which impacts both memory usage and throughput:

• fp32 (float32)
• fp16 (float16)
• bf16 (bfloat16)
• tf32 (CUDA internal data type)

Here is a diagram that shows how these data types correlate to each other.

(source: NVIDIA Blog)

While fp16 and fp32 have been around for quite some time, bf16 and tf32 are only available on the Ampere architecture GPUS. TPUs support bf16 as well.

fp16

AMP = Automatic Mixed Precision

If we look at what’s happening with FP16 training (mixed precision) we have:

• the model has two copies in memory: one in half-precision for the forward/backward computations and one in full precision - no memory saved here
• the forward activations saved for gradient computation are in half-precision - memory is saved here
• the gradients are computed in half-precision but converted to full-precision for the update, no saving there
• the optimizer states are in full precision as all the updates are done in full-precision

So the savings only happen for the forward activations saved for the backward computation, and there is a slight overhead because the model weights are stored both in half- and full-precision.

In 🤗 Transformers fp16 mixed precision is enabled by passing --fp16 to the 🤗 Trainer.

Now let’s look at a simple text-classification fine-tuning on 2 GPUs (I’m giving the command for reference):

export BS=16
python -m torch.distributed.launch \
    --nproc_per_node 2 examples/pytorch/text-classification/run_glue.py \
    --model_name_or_path bert-base-cased \
    --task_name mrpc \
    --do_train \
    --do_eval \
    --max_seq_length 128 \
    --per_device_train_batch_size $BS \
    --learning_rate 2e-5 \
    --num_train_epochs 3.0 \
    --output_dir /tmp/mrpc \
    --overwrite_output_dir \
    --fp16

Since the only savings we get are in the model activations saved for the backward passed, it’s logical that the bigger those activations are, the bigger the saving will be. If we try different batch sizes, I indeed get (this is with nvidia-smi so not completely reliable as said above but it will be a fair comparison):

So there is only a real memory saving if we train at a high batch size (and it’s not half) and at batch sizes lower than 8, you actually get a bigger memory footprint (because of the overhead mentioned above). The gain for FP16 training is that in each of those cases, the training with the flag --fp16 is twice as fast, which does require every tensor to have every dimension be a multiple of 8 (examples pad the tensors to a sequence length that is a multiple of 8).

Summary: FP16 with apex or AMP will only give you some memory savings with a reasonably high batch size.

Additionally, under mixed precision when possible, it’s important that the batch size is a multiple of 8 to efficiently use tensor cores.

Note that in some situations the speed up can be as big as 5x when using mixed precision. e.g. we have observed that while using Megatron-Deepspeed.

Some amazing tutorials to read on mixed precision:

• @sgugger wrote a great explanation of mixed precision here
• Aleksey Bilogur’s A developer-friendly guide to mixed precision training with PyTorch

You can also see a variety of benchmarks on fp16 vs other precisions: RTX-3090 and A100.

fp16 caching

pytorch autocast which performs AMP include a caching feature, which speed things up by caching fp16-converted values. Here is the full description from this comment:

Autocast maintains a cache of the FP16 casts of model parameters (leaves). This helps streamline parameter reuse: if the same FP32 param is used in several different FP16list ops, like several matmuls, instead of re-casting the param to FP16 on entering each matmul, the cast will occur on the first matmul, the casted FP16 copy will be cached, and for all later matmuls the FP16 copy will be reused. The cache is maintained only within a particular outermost autocast context. When you exit the autocast context the cache is dropped. For recommended usage, in which autocast wraps the forward pass, and then you exit the context before calling backward(), this means the cache only lasts the duration of the forward pass each iteration, and will be rebuilt next iteration. (The cache of FP16-casted copies MUST be rebuilt each iteration. The FP32 parameters get updated by the optimizer, so the FP16 copies must be recreated, otherwise the FP16 values will be stale.)

fp16 Inference

While normally inference is done with fp16/amp as with training, it’s also possible to use the full fp16 mode without using mixed precision. This is especially a good fit if the pretrained model weights are already in fp16. So a lot less memory is used: 2 bytes per parameter vs 6 bytes with mixed precision!

How good the results this will deliver will depend on the model. If it can handle fp16 without overflows and accuracy issues, then it’ll definitely better to use the full fp16 mode.

For example, LayerNorm has to be done in fp32 and recent pytorch (1.10+) has been fixed to do that regardless of the input types, but earlier pytorch versions accumulate in the input type which can be an issue.

In 🤗 Transformers the full fp16 inference is enabled by passing --fp16_full_eval to the 🤗 Trainer.

bf16

If you own Ampere or newer hardware you can start using bf16 for your training and evaluation. While bf16 has a worse precision than fp16, it has a much much bigger dynamic range. Therefore, if in the past you were experiencing overflow issues while training the model, bf16 will prevent this from happening most of the time. Remember that in fp16 the biggest number you can have is 65535 and any number above that will overflow. A bf16 number can be as large as 3.39e+38 (!) which is about the same as fp32 - because both have 8-bits used for the numerical range.

Automatic Mixed Precision (AMP) is the same as with fp16, except it’ll use bf16.

Thanks to the fp32-like dynamic range with bf16 mixed precision loss scaling is no longer needed.

If you have tried to finetune models pre-trained under bf16 mixed precision (e.g. T5) it’s very likely that you have encountered overflow issues. Now you should be able to finetune those models without any issues.

That said, also be aware that if you pre-trained a model in bf16, it’s likely to have overflow issues if someone tries to finetune it in fp16 down the road. So once started on the bf16-mode path it’s best to remain on it and not switch to fp16.

In 🤗 Transformers bf16 mixed precision is enabled by passing --bf16 to the 🤗 Trainer.

If you use your own trainer, this is just:

from torch.cuda.amp import autocast
with autocast(dtype=torch.bfloat16):
    loss, outputs = ...

If you need to switch a tensor to bf16, it’s just: t.to(dtype=torch.bfloat16)

Here is how you can check if your setup supports bf16:

python -c 'import transformers; print(f"BF16 support is {transformers.file_utils.is_torch_bf16_available()}")'

On the other hand bf16 has a much worse precision than fp16, so there are certain situations where you’d still want to use fp16 and not bf16.

You can also see a variety of benchmarks on bf16 vs other precisions: RTX-3090 and A100.

bf16 Inference

Same as with fp16, you can do inference in either the mixed precision bf16 or using the full bf16 mode. The same caveats apply. For details see fp16 Inference.

In 🤗 Transformers the full bf16 inference is enabled by passing --bf16_full_eval to the 🤗 Trainer.

tf32

The Ampere hardware uses a magical data type called tf32. It has the same numerical range as fp32 (8-bits), but instead of 23 bits precision it has only 10 bits (same as fp16). In total it uses only 19 bits.

It’s magical in the sense that you can use the normal fp32 training and/or inference code and by enabling tf32 support you can get up to 3x throughput improvement. All you need to do is to add this to your code:

import torch
torch.backends.cuda.matmul.allow_tf32 = True

When this is done CUDA will automatically switch to using tf32 instead of fp32 where it’s possible. This, of course, assumes that the used GPU is from the Ampere series.

Like all cases with reduced precision this may or may not be satisfactory for your needs, so you have to experiment and see. According to NVIDIA research the majority of machine learning training shouldn’t be impacted and showed the same perplexity and convergence as the fp32 training.

If you’re already using fp16 or bf16 mixed precision it may help with the throughput as well.

You can enable this mode in the 🤗 Trainer with --tf32, or disable it with --tf32 0 or --no_tf32. By default the PyTorch default is used.

Note: tf32 mode is internal to CUDA and can’t be accessed directly via tensor.to(dtype=torch.tf32) as torch.tf32 doesn’t exit.

Note: you need torch>=1.7 to enjoy this feature.

You can also see a variety of benchmarks on tf32 vs other precisions: RTX-3090 and A100.

Gradient Accumulation

Since gradient accumulation essentially is identical to having a larger batch size, just as with the larger batch size here you are likely to see a 20-30% speedup due to the optimizer running less often. For example, see benchmarks for RTX-3090 and A100.

To activate this feature in 🤗 Trainer add --gradient_accumulation_steps 4 to its arguments (experiment with the value to get the best performance).

It’s important to remember that using gradient accumulation you may end up with a much larger effective batch size, so you may need to adjust the learning rate, its warm up and for very short datasets it’ll impact the loss as the training will end up doing less steps than normal.

Gradient Checkpointing

One way to use significantly less GPU memory is to enabled “Gradient Checkpointing” (also known as “activation checkpointing”). When enabled, a lot of memory can be freed at the cost of small decrease in the training speed due to recomputing parts of the graph during back-propagation. The slowdown will depend on the model but quite often it is around 20-30%.

This technique was first shared in the paper: Training Deep Nets with Sublinear Memory Cost. The paper will also give you the exact details on the savings, but it’s in the ballpark of O(sqrt(n)), where n is the number of feed-forward layers.

To activate this feature in 🤗 Transformers for models that support it, use:

model.gradient_checkpointing_enable()

or add --gradient_checkpointing to the Trainer arguments.

Batch sizes

One gets the most efficient performance when batch sizes and input/output neuron counts are divisible by a certain number, which typically starts at 8, but can be much higher as well. That number varies a lot depending on the specific hardware being used and the dtype of the model.

For example for fully connected layers (which correspond to GEMMs), NVIDIA provides recommendations for input/output neuron counts and batch size.

Tensor Core Requirements define the multiplier based on the dtype and the hardware. For example, for fp16 a multiple of 8 is recommended, but on A100 it’s 64!

For parameters that are small, there is also Dimension Quantization Effects to consider, this is where tiling happens and the right multiplier can have a significant speedup.

Additionally, as explained in the Gradient Accumulation section, the bigger the batch size the less often the optimizer is run, the faster the training is (considering the same dataset length). See benchmarks for RTX-3090 and A100.

DP vs DDP

DistributedDataParallel (DDP) is typically faster than DataParallel (DP), but it is not always the case:

• while DP is python threads-based, DDP is multiprocess-based - and as such it has no python threads limitations, such as GIL
• on the other hand a slow inter-connectivity between the GPU cards could lead to an actual slower outcome with DDP

Here are the main differences in the inter-GPU communication overhead between the two modes:

DDP:

• At the start time the main process replicates the model once from gpu 0 to the rest of gpus
• Then for each batch:
  1. each gpu consumes each own mini-batch of data directly
  2. during backward, once the local gradients are ready, they are then averaged across all processes

DP:

For each batch:

1. gpu 0 reads the batch of data and then sends a mini-batch to each gpu
2. replicates the up-to-date model from gpu 0 to each gpu
3. runs forward and sends output from each gpu to gpu 0, computes loss
4. scatters loss from gpu 0 to all gpus, runs backward
5. sends gradients from each gpu to gpu 0 and averages those

The only communication DDP performs per batch is sending gradients, whereas DP does 5 different data exchanges per batch.

DP copies data within the process via python threads, whereas DDP copies data via torch.distributed.

Under DP gpu 0 performs a lot more work than the rest of the gpus, thus resulting in under-utilization of gpus.

You can use DDP across multiple machines, but this is not the case with DP.

There are other differences between DP and DDP but they aren’t relevant to this discussion.

If you want to go really deep into understanding these 2 modes, this article is highly recommended, as it has great diagrams, includes multiple benchmarks and profiler outputs on various hardware, explains all the nuances that you may need to know.

Let’s look at an actual benchmark:

Analysis:

Here DP is ~10% slower than DDP w/ NVlink, but ~15% faster than DDP w/o NVlink

The real difference will depend on how much data each GPU needs to sync with the others - the more there is to sync, the more a slow link will slow down the total runtime.

Here is the full benchmark code and outputs:

NCCL_P2P_DISABLE=1 was used to disable the NVLink feature on the corresponding benchmark.

# DP
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
python examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200

{'train_runtime': 110.5948, 'train_samples_per_second': 1.808, 'epoch': 0.69}

# DDP w/ NVlink
rm -r /tmp/test-clm; CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200

{'train_runtime': 101.9003, 'train_samples_per_second': 1.963, 'epoch': 0.69}

# DDP w/o NVlink
rm -r /tmp/test-clm; NCCL_P2P_DISABLE=1 CUDA_VISIBLE_DEVICES=0,1 \
python -m torch.distributed.launch --nproc_per_node 2 examples/pytorch/language-modeling/run_clm.py \
--model_name_or_path gpt2 --dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --output_dir /tmp/test-clm --per_device_train_batch_size 4 --max_steps 200

{'train_runtime': 131.4367, 'train_samples_per_second': 1.522, 'epoch': 0.69}

Hardware: 2x TITAN RTX 24GB each + NVlink with 2 NVLinks (NV2 in nvidia-smi topo -m) Software: pytorch-1.8-to-be + cuda-11.0 / transformers==4.3.0.dev0

DataLoader

One of the important requirements to reach great training speed is the ability to feed the GPU at the maximum speed it can handle. By default everything happens in the main process and it might not be able to read the data from disk fast enough, and thus create a bottleneck, leading to GPU under-utilization.

• DataLoader(pin_memory=True, ...) which ensures that the data gets preloaded into the pinned memory on CPU and typically leads to much faster transfers from CPU to GPU memory.
• DataLoader(num_workers=4, ...) - spawn several workers to pre-load data faster - during training watch the GPU utilization stats and if it’s far from 100% experiment with raising the number of workers. Of course, the problem could be elsewhere so a very big number of workers won’t necessarily lead to a better performance.

Faster optimizer

pytorch-nightly introduced torch.optim._multi_tensor which should significantly speed up the optimizers for situations with lots of small feature tensors. It should eventually become the default, but if you want to experiment with it sooner and don’t mind using the bleed-edge, see: https://github.com/huggingface/transformers/issues/9965

Sparsity

Mixture of Experts

Quite a few of the recent papers reported a 4-5x training speedup and a faster inference by integrating Mixture of Experts (MoE) into the Transformer models.

Since it has been discovered that more parameters lead to better performance, this technique allows to increase the number of parameters by an order of magnitude without increasing training costs.

In this approach every other FFN layer is replaced with a MoE Layer which consists of many experts, with a gated function that trains each expert in a balanced way depending on the input token’s position in a sequence.

(source: GLAM)

You can find exhaustive details and comparison tables in the papers listed at the end of this section.

The main drawback of this approach is that it requires staggering amounts of GPU memory - almost an order of magnitude larger than its dense equivalent. Various distillation and approaches are proposed to how to overcome the much higher memory requirements.

There is direct trade-off though, you can use just a few experts with a 2-3x smaller base model instead of dozens or hundreds experts leading to a 5x smaller model and thus increase the training speed moderately while increasing the memory requirements moderately as well.

Most related papers and implementations are built around Tensorflow/TPUs:

• GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
• Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
• GLaM: Generalist Language Model (GLaM)

And for Pytorch DeepSpeed has built one as well: DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale, Mixture of Experts - blog posts: 1, 2 and specific deployment with large transformer-based natural language generation models: blog post, Megatron-Deepspeed branch.

Efficient Software Prebuilds

PyTorch’s pip and conda builds come prebuit with the cuda toolkit which is enough to run PyTorch, but it is insufficient if you need to build cuda extensions.

At times it may take an additional effort to pre-build some components, e.g., if you’re using libraries like apex that don’t come pre-compiled. In other situations figuring out how to install the right cuda toolkit system-wide can be complicated. To address these users’ needs PyTorch and NVIDIA release a new version of NGC docker container which already comes with everything prebuilt and you just need to install your programs on it and it will run out of the box.

This approach is also useful if you want to tweak the pytorch source and/or make a new customized build.

To find the docker image version you want start here, choose one of the latest monthly releases. Go into the release’s notes for the desired release, check that the environment’s components are matching your needs (including NVIDIA Driver requirements!) and then at the very top of that document go to the corresponding NGC page. If for some reason you get lost, here is the index of all PyTorch NGC images.

Next follow the instructions to download and deploy the docker image.

Contribute

This document is far from being complete and a lot more needs to be added, so if you have additions or corrections to make please don’t hesitate to open a PR or if you aren’t sure start an Issue and we can discuss the details there.

When making contributions that A is better than B, please try to include a reproducible benchmark and/or a link to the source of that information (unless it comes directly from you).