# Performance Tuning Guide¶

Author: Szymon Migacz

Performance Tuning Guide is a set of optimizations and best practices which can accelerate training and inference of deep learning models in PyTorch. Presented techniques often can be implemented by changing only a few lines of code and can be applied to a wide range of deep learning models across all domains.

## General optimizations¶

### Enable async data loading and augmentation¶

torch.utils.data.DataLoader supports asynchronous data loading and data augmentation in separate worker subprocesses. The default setting for DataLoader is num_workers=0, which means that the data loading is synchronous and done in the main process. As a result the main training process has to wait for the data to be available to continue the execution.

Setting num_workers > 0 enables asynchronous data loading and overlap between the training and data loading. num_workers should be tuned depending on the workload, CPU, GPU, and location of training data.

DataLoader accepts pin_memory argument, which defaults to False. When using a GPU it’s better to set pin_memory=True, this instructs DataLoader to use pinned memory and enables faster and asynchronous memory copy from the host to the GPU.

### Disable gradient calculation for validation or inference¶

PyTorch saves intermediate buffers from all operations which involve tensors that require gradients. Typically gradients aren’t needed for validation or inference. torch.no_grad() context manager can be applied to disable gradient calculation within a specified block of code, this accelerates execution and reduces the amount of required memory. torch.no_grad() can also be used as a function decorator.

### Disable bias for convolutions directly followed by a batch norm¶

torch.nn.Conv2d() has bias parameter which defaults to True (the same is true for Conv1d and Conv3d ).

If a nn.Conv2d layer is directly followed by a nn.BatchNorm2d layer, then the bias in the convolution is not needed, instead use nn.Conv2d(..., bias=False, ....). Bias is not needed because in the first step BatchNorm subtracts the mean, which effectively cancels out the effect of bias.

This is also applicable to 1d and 3d convolutions as long as BatchNorm (or other normalization layer) normalizes on the same dimension as convolution’s bias.

Models available from torchvision already implement this optimization.

### Use parameter.grad = None instead of model.zero_grad() or optimizer.zero_grad()¶

Instead of calling:

model.zero_grad()
# or
optimizer.zero_grad()


to zero out gradients, use the following method instead:

for param in model.parameters():
param.grad = None


The second code snippet does not zero the memory of each individual parameter, also the subsequent backward pass uses assignment instead of addition to store gradients, this reduces the number of memory operations.

Setting gradient to None has a slightly different numerical behavior than setting it to zero, for more details refer to the documentation.

Alternatively, starting from PyTorch 1.7, call model or optimizer.zero_grad(set_to_none=True).

### Fuse pointwise operations¶

Pointwise operations (elementwise addition, multiplication, math functions - sin(), cos(), sigmoid() etc.) can be fused into a single kernel to amortize memory access time and kernel launch time.

PyTorch JIT can fuse kernels automatically, although there could be additional fusion opportunities not yet implemented in the compiler, and not all device types are supported equally.

Pointwise operations are memory-bound, for each operation PyTorch launches a separate kernel. Each kernel loads data from the memory, performs computation (this step is usually inexpensive) and stores results back into the memory.

Fused operator launches only one kernel for multiple fused pointwise ops and loads/stores data only once to the memory. This makes JIT very useful for activation functions, optimizers, custom RNN cells etc.

In the simplest case fusion can be enabled by applying torch.jit.script decorator to the function definition, for example:

@torch.jit.script
def fused_gelu(x):
return x * 0.5 * (1.0 + torch.erf(x / 1.41421))


Refer to TorchScript documentation for more advanced use cases.

### Enable channels_last memory format for computer vision models¶

PyTorch 1.5 introduced support for channels_last memory format for convolutional networks. This format is meant to be used in conjunction with AMP to further accelerate convolutional neural networks with Tensor Cores.

Support for channels_last is experimental, but it’s expected to work for standard computer vision models (e.g. ResNet-50, SSD). To convert models to channels_last format follow Channels Last Memory Format Tutorial. The tutorial includes a section on converting existing models.

### Checkpoint intermediate buffers¶

Buffer checkpointing is a technique to mitigate the memory capacity burden of model training. Instead of storing inputs of all layers to compute upstream gradients in backward propagation, it stores the inputs of a few layers and the others are recomputed during backward pass. The reduced memory requirements enables increasing the batch size that can improve utilization.

Checkpointing targets should be selected carefully. The best is not to store large layer outputs that have small re-computation cost. The example target layers are activation functions (e.g. ReLU, Sigmoid, Tanh), up/down sampling and matrix-vector operations with small accumulation depth.

PyTorch supports a native torch.utils.checkpoint API to automatically perform checkpointing and recomputation.

### Disable debugging APIs¶

Many PyTorch APIs are intended for debugging and should be disabled for regular training runs:

## CPU specific optimizations¶

### Utilize Non-Uniform Memory Access (NUMA) Controls¶

NUMA or non-uniform memory access is a memory layout design used in data center machines meant to take advantage of locality of memory in multi-socket machines with multiple memory controllers and blocks. Generally speaking, all deep learning workloads, training or inference, get better performance without accessing hardware resources across NUMA nodes. Thus, inference can be run with multiple instances, each instance runs on one socket, to raise throughput. For training tasks on single node, distributed training is recommended to make each training process run on one socket.

In general cases the following command executes a PyTorch script on cores on the Nth node only, and avoids cross-socket memory access to reduce memory access overhead.

# numactl --cpunodebind=N --membind=N python <pytorch_script>


More detailed descriptions can be found here.

### Utilize OpenMP¶

OpenMP is utilized to bring better performance for parallel computation tasks. OMP_NUM_THREADS is the easiest switch that can be used to accelerate computations. It determines number of threads used for OpenMP computations. CPU affinity setting controls how workloads are distributed over multiple cores. It affects communication overhead, cache line invalidation overhead, or page thrashing, thus proper setting of CPU affinity brings performance benefits. GOMP_CPU_AFFINITY or KMP_AFFINITY determines how to bind OpenMP* threads to physical processing units. Detailed information can be found here.

With the following command, PyTorch run the task on N OpenMP threads.

# export OMP_NUM_THREADS=N


Typically, the following environment variables are used to set for CPU affinity with GNU OpenMP implementation. OMP_PROC_BIND specifies whether threads may be moved between processors. Setting it to CLOSE keeps OpenMP threads close to the primary thread in contiguous place partitions. OMP_SCHEDULE determines how OpenMP threads are scheduled. GOMP_CPU_AFFINITY binds threads to specific CPUs.

# export OMP_SCHEDULE=STATIC
# export OMP_PROC_BIND=CLOSE
# export GOMP_CPU_AFFINITY="N-M"


### Intel OpenMP Runtime Library (libiomp)¶

By default, PyTorch uses GNU OpenMP (GNU libgomp) for parallel computation. On Intel platforms, Intel OpenMP Runtime Library (libiomp) provides OpenMP API specification support. It sometimes brings more performance benefits compared to libgomp. Utilizing environment variable LD_PRELOAD can switch OpenMP library to libiomp:



### Train a model on CPU with PyTorch DistributedDataParallel(DDP) functionality¶

For small scale models or memory-bound models, such as DLRM, training on CPU is also a good choice. On a machine with multiple sockets, distributed training brings a high-efficient hardware resource usage to accelerate the training process. Torch-ccl, optimized with Intel(R) oneCCL (collective commnications library) for efficient distributed deep learning training implementing such collectives like allreduce, allgather, alltoall, implements PyTorch C10D ProcessGroup API and can be dynamically loaded as external ProcessGroup. Upon optimizations implemented in PyTorch DDP moduel, torhc-ccl accelerates communication operations. Beside the optimizations made to communication kernels, torch-ccl also features simultaneous computation-communication functionality.

## GPU specific optimizations¶

### Enable cuDNN auto-tuner¶

NVIDIA cuDNN supports many algorithms to compute a convolution. Autotuner runs a short benchmark and selects the kernel with the best performance on a given hardware for a given input size.

For convolutional networks (other types currently not supported), enable cuDNN autotuner before launching the training loop by setting:

torch.backends.cudnn.benchmark = True

• the auto-tuner decisions may be non-deterministic; different algorithm may be selected for different runs. For more details see PyTorch: Reproducibility
• in some rare cases, such as with highly variable input sizes, it’s better to run convolutional networks with autotuner disabled to avoid the overhead associated with algorithm selection for each input size.

### Avoid unnecessary CPU-GPU synchronization¶

Avoid unnecessary synchronizations, to let the CPU run ahead of the accelerator as much as possible to make sure that the accelerator work queue contains many operations.

When possible, avoid operations which require synchronizations, for example:

• print(cuda_tensor)
• cuda_tensor.item()
• memory copies: tensor.cuda(), cuda_tensor.cpu() and equivalent tensor.to(device) calls
• cuda_tensor.nonzero()
• python control flow which depends on results of operations performed on cuda tensors e.g. if (cuda_tensor != 0).all()

### Create tensors directly on the target device¶

Instead of calling torch.rand(size).cuda() to generate a random tensor, produce the output directly on the target device: torch.rand(size, device=torch.device('cuda')).

This is applicable to all functions which create new tensors and accept device argument: torch.rand(), torch.zeros(), torch.full() and similar.

### Use mixed precision and AMP¶

Mixed precision leverages Tensor Cores and offers up to 3x overall speedup on Volta and newer GPU architectures. To use Tensor Cores AMP should be enabled and matrix/tensor dimensions should satisfy requirements for calling kernels that use Tensor Cores.

To use Tensor Cores:

• set sizes to multiples of 8 (to map onto dimensions of Tensor Cores)
• see Deep Learning Performance Documentation for more details and guidelines specific to layer type
• if layer size is derived from other parameters rather than fixed, it can still be explicitly padded e.g. vocabulary size in NLP models
• enable AMP

### Pre-allocate memory in case of variable input length¶

Models for speech recognition or for NLP are often trained on input tensors with variable sequence length. Variable length can be problematic for PyTorch caching allocator and can lead to reduced performance or to unexpected out-of-memory errors. If a batch with a short sequence length is followed by an another batch with longer sequence length, then PyTorch is forced to release intermediate buffers from previous iteration and to re-allocate new buffers. This process is time consuming and causes fragmentation in the caching allocator which may result in out-of-memory errors.

A typical solution is to implement pre-allocation. It consists of the following steps:

1. generate a (usually random) batch of inputs with maximum sequence length (either corresponding to max length in the training dataset or to some predefined threshold)
2. execute a forward and a backward pass with the generated batch, do not execute an optimizer or a learning rate scheduler, this step pre-allocates buffers of maximum size, which can be reused in subsequent training iterations
3. zero out gradients
4. proceed to regular training

## Distributed optimizations¶

### Use efficient data-parallel backend¶

PyTorch has two ways to implement data-parallel training:

DistributedDataParallel offers much better performance and scaling to multiple-GPUs. For more information refer to the relevant section of CUDA Best Practices from PyTorch documentation.

### Skip unnecessary all-reduce if training with DistributedDataParallel and gradient accumulation¶

By default torch.nn.parallel.DistributedDataParallel executes gradient all-reduce after every backward pass to compute the average gradient over all workers participating in the training. If training uses gradient accumulation over N steps, then all-reduce is not necessary after every training step, it’s only required to perform all-reduce after the last call to backward, just before the execution of the optimizer.

DistributedDataParallel provides no_sync() context manager which disables gradient all-reduce for particular iteration. no_sync() should be applied to first N-1 iterations of gradient accumulation, the last iteration should follow the default execution and perform the required gradient all-reduce.

### Match the order of layers in constructors and during the execution if using DistributedDataParallel(find_unused_parameters=True)¶

torch.nn.parallel.DistributedDataParallel with find_unused_parameters=True uses the order of layers and parameters from model constructors to build buckets for DistributedDataParallel gradient all-reduce. DistributedDataParallel overlaps all-reduce with the backward pass. All-reduce for a particular bucket is asynchronously triggered only when all gradients for parameters in a given bucket are available.

To maximize the amount of overlap, the order in model constructors should roughly match the order during the execution. If the order doesn’t match, then all-reduce for the entire bucket waits for the gradient which is the last to arrive, this may reduce the overlap between backward pass and all-reduce, all-reduce may end up being exposed, which slows down the training.

DistributedDataParallel with find_unused_parameters=False (which is the default setting) relies on automatic bucket formation based on order of operations encountered during the backward pass. With find_unused_parameters=False it’s not necessary to reorder layers or parameters to achieve optimal performance.

### Load-balance workload in a distributed setting¶

Load imbalance typically may happen for models processing sequential data (speech recognition, translation, language models etc.). If one device receives a batch of data with sequence length longer than sequence lengths for the remaining devices, then all devices wait for the worker which finishes last. Backward pass functions as an implicit synchronization point in a distributed setting with DistributedDataParallel backend.

There are multiple ways to solve the load balancing problem. The core idea is to distribute workload over all workers as uniformly as possible within each global batch. For example Transformer solves imbalance by forming batches with approximately constant number of tokens (and variable number of sequences in a batch), other models solve imbalance by bucketing samples with similar sequence length or even by sorting dataset by sequence length.

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery