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 asynchronous 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 operations

Pointwise operations such as elementwise addition, multiplication, and math functions like sin(), cos(), sigmoid(), etc., can be combined into a single kernel. This fusion helps reduce memory access and kernel launch times. Typically, pointwise operations are memory-bound; PyTorch eager-mode initiates a separate kernel for each operation, which involves loading data from memory, executing the operation (often not the most time-consuming step), and writing the results back to memory.

By using a fused operator, only one kernel is launched for multiple pointwise operations, and data is loaded and stored just once. This efficiency is particularly beneficial for activation functions, optimizers, and custom RNN cells etc.

PyTorch 2 introduces a compile-mode facilitated by TorchInductor, an underlying compiler that automatically fuses kernels. TorchInductor extends its capabilities beyond simple element-wise operations, enabling advanced fusion of eligible pointwise and reduction operations for optimized performance.

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

@torch.compile
def gelu(x):
    return x * 0.5 * (1.0 + torch.erf(x / 1.41421))

Refer to Introduction to torch.compile 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:

• anomaly detection: torch.autograd.detect_anomaly or torch.autograd.set_detect_anomaly(True)

• profiler related: torch.autograd.profiler.emit_nvtx, torch.autograd.profiler.profile

• autograd gradcheck: torch.autograd.gradcheck or torch.autograd.gradgradcheck

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. An important tuning parameter is core pinning which prevent the threads of migrating between multiple CPUs, enhancing data location and minimizing inter core communication.

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:

export LD_PRELOAD=<path>/libiomp5.so:$LD_PRELOAD

Similar to CPU affinity settings in GNU OpenMP, environment variables are provided in libiomp to control CPU affinity settings. KMP_AFFINITY binds OpenMP threads to physical processing units. KMP_BLOCKTIME sets the time, in milliseconds, that a thread should wait, after completing the execution of a parallel region, before sleeping. In most cases, setting KMP_BLOCKTIME to 1 or 0 yields good performances. The following commands show a common settings with Intel OpenMP Runtime Library.

export KMP_AFFINITY=granularity=fine,compact,1,0
export KMP_BLOCKTIME=1

Switch Memory allocator

For deep learning workloads, Jemalloc or TCMalloc can get better performance by reusing memory as much as possible than default malloc function. Jemalloc is a general purpose malloc implementation that emphasizes fragmentation avoidance and scalable concurrency support. TCMalloc also features a couple of optimizations to speed up program executions. One of them is holding memory in caches to speed up access of commonly-used objects. Holding such caches even after deallocation also helps avoid costly system calls if such memory is later re-allocated. Use environment variable LD_PRELOAD to take advantage of one of them.

export LD_PRELOAD=<jemalloc.so/tcmalloc.so>:$LD_PRELOAD

Use oneDNN Graph with TorchScript for inference

oneDNN Graph can significantly boost inference performance. It fuses some compute-intensive operations such as convolution, matmul with their neighbor operations. In PyTorch 2.0, it is supported as a beta feature for Float32 & BFloat16 data-types. oneDNN Graph receives the model’s graph and identifies candidates for operator-fusion with respect to the shape of the example input. A model should be JIT-traced using an example input. Speed-up would then be observed after a couple of warm-up iterations for inputs with the same shape as the example input. The example code-snippets below are for resnet50, but they can very well be extended to use oneDNN Graph with custom models as well.

# Only this extra line of code is required to use oneDNN Graph
torch.jit.enable_onednn_fusion(True)

Using the oneDNN Graph API requires just one extra line of code for inference with Float32. If you are using oneDNN Graph, please avoid calling torch.jit.optimize_for_inference.

# sample input should be of the same shape as expected inputs
sample_input = [torch.rand(32, 3, 224, 224)]
# Using resnet50 from torchvision in this example for illustrative purposes,
# but the line below can indeed be modified to use custom models as well.
model = getattr(torchvision.models, "resnet50")().eval()
# Tracing the model with example input
traced_model = torch.jit.trace(model, sample_input)
# Invoking torch.jit.freeze
traced_model = torch.jit.freeze(traced_model)

Once a model is JIT-traced with a sample input, it can then be used for inference after a couple of warm-up runs.

with torch.no_grad():
    # a couple of warm-up runs
    traced_model(*sample_input)
    traced_model(*sample_input)
    # speedup would be observed after warm-up runs
    traced_model(*sample_input)

While the JIT fuser for oneDNN Graph also supports inference with BFloat16 datatype, performance benefit with oneDNN Graph is only exhibited by machines with AVX512_BF16 instruction set architecture (ISA). The following code snippets serves as an example of using BFloat16 datatype for inference with oneDNN Graph:

# AMP for JIT mode is enabled by default, and is divergent with its eager mode counterpart
torch._C._jit_set_autocast_mode(False)

with torch.no_grad(), torch.cpu.amp.autocast(cache_enabled=False, dtype=torch.bfloat16):
    # Conv-BatchNorm folding for CNN-based Vision Models should be done with ``torch.fx.experimental.optimization.fuse`` when AMP is used
    import torch.fx.experimental.optimization as optimization
    # Please note that optimization.fuse need not be called when AMP is not used
    model = optimization.fuse(model)
    model = torch.jit.trace(model, (example_input))
    model = torch.jit.freeze(model)
    # a couple of warm-up runs
    model(example_input)
    model(example_input)
    # speedup would be observed in subsequent runs.
    model(example_input)

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 communications 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 module, torch-ccl accelerates communication operations. Beside the optimizations made to communication kernels, torch-ccl also features simultaneous computation-communication functionality.

GPU specific optimizations

Enable Tensor cores

Tensor cores are specialized hardware designed to compute matrix-matrix multiplication operations, primarily utilized in deep learning and AI workloads. Tensor cores have specific precision requirements which can be adjusted manually or via the Automatic Mixed Precision API.

In particular, tensor operations take advantage of lower precision workloads. Which can be controlled via torch.set_float32_matmul_precision. The default format is set to ‘highest,’ which utilizes the tensor data type. However, PyTorch offers alternative precision settings: ‘high’ and ‘medium.’ These options prioritize computational speed over numerical precision.”

Use CUDA Graphs

At the time of using a GPU, work first must be launched from the CPU and in some cases the context switch between CPU and GPU can lead to bad resource utilization. CUDA graphs are a way to keep computation within the GPU without paying the extra cost of kernel launches and host synchronization.

# It can be enabled using
torch.compile(m, "reduce-overhead")
# or
torch.compile(m, "max-autotune")

Support for CUDA graph is in development, and its usage can incur in increased device memory consumption and some models might not compile.

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='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

  • Introduction to Mixed Precision Training and AMP: video, slides

  • native PyTorch AMP is available starting from PyTorch 1.6: documentation, examples, tutorial

Preallocate 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 preallocation. 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 preallocates 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:

• torch.nn.DataParallel

• torch.nn.parallel.DistributedDataParallel

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.