CUDA semantics¶

torch.cuda is used to set up and run CUDA operations. It keeps track of the currently selected GPU, and all CUDA tensors you allocate will by default be created on that device. The selected device can be changed with a torch.cuda.device context manager.

However, once a tensor is allocated, you can do operations on it irrespective of the selected device, and the results will be always placed on the same device as the tensor.

Cross-GPU operations are not allowed by default, with the exception of copy_() and other methods with copy-like functionality such as to() and cuda(). Unless you enable peer-to-peer memory access, any attempts to launch ops on tensors spread across different devices will raise an error.

Below you can find a small example showcasing this:

cuda = torch.device('cuda')     # Default CUDA device
cuda0 = torch.device('cuda:0')
cuda2 = torch.device('cuda:2')  # GPU 2 (these are 0-indexed)

x = torch.tensor([1., 2.], device=cuda0)
# x.device is device(type='cuda', index=0)
y = torch.tensor([1., 2.]).cuda()
# y.device is device(type='cuda', index=0)

with torch.cuda.device(1):
    # allocates a tensor on GPU 1
    a = torch.tensor([1., 2.], device=cuda)

    # transfers a tensor from CPU to GPU 1
    b = torch.tensor([1., 2.]).cuda()
    # a.device and b.device are device(type='cuda', index=1)

    # You can also use ``Tensor.to`` to transfer a tensor:
    b2 = torch.tensor([1., 2.]).to(device=cuda)
    # b.device and b2.device are device(type='cuda', index=1)

    c = a + b
    # c.device is device(type='cuda', index=1)

    z = x + y
    # z.device is device(type='cuda', index=0)

    # even within a context, you can specify the device
    # (or give a GPU index to the .cuda call)
    d = torch.randn(2, device=cuda2)
    e = torch.randn(2).to(cuda2)
    f = torch.randn(2).cuda(cuda2)
    # d.device, e.device, and f.device are all device(type='cuda', index=2)

TensorFloat-32 (TF32) on Ampere (and later) devices¶

Starting in PyTorch 1.7, there is a new flag called allow_tf32. This flag defaults to True in PyTorch 1.7 to PyTorch 1.11, and False in PyTorch 1.12 and later. This flag controls whether PyTorch is allowed to use the TensorFloat32 (TF32) tensor cores, available on NVIDIA GPUs since Ampere, internally to compute matmul (matrix multiplies and batched matrix multiplies) and convolutions.

TF32 tensor cores are designed to achieve better performance on matmul and convolutions on torch.float32 tensors by rounding input data to have 10 bits of mantissa, and accumulating results with FP32 precision, maintaining FP32 dynamic range.

matmuls and convolutions are controlled separately, and their corresponding flags can be accessed at:

# The flag below controls whether to allow TF32 on matmul. This flag defaults to False
# in PyTorch 1.12 and later.
torch.backends.cuda.matmul.allow_tf32 = True

# The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
torch.backends.cudnn.allow_tf32 = True

The precision of matmuls can also be set more broadly (limited not just to CUDA) via set_float_32_matmul_precision(). Note that besides matmuls and convolutions themselves, functions and nn modules that internally uses matmuls or convolutions are also affected. These include nn.Linear, nn.Conv*, cdist, tensordot, affine grid and grid sample, adaptive log softmax, GRU and LSTM.

To get an idea of the precision and speed, see the example code and benchmark data (on A100) below:

a_full = torch.randn(10240, 10240, dtype=torch.double, device='cuda')
b_full = torch.randn(10240, 10240, dtype=torch.double, device='cuda')
ab_full = a_full @ b_full
mean = ab_full.abs().mean()  # 80.7277

a = a_full.float()
b = b_full.float()

# Do matmul at TF32 mode.
torch.backends.cuda.matmul.allow_tf32 = True
ab_tf32 = a @ b  # takes 0.016s on GA100
error = (ab_tf32 - ab_full).abs().max()  # 0.1747
relative_error = error / mean  # 0.0022

# Do matmul with TF32 disabled.
torch.backends.cuda.matmul.allow_tf32 = False
ab_fp32 = a @ b  # takes 0.11s on GA100
error = (ab_fp32 - ab_full).abs().max()  # 0.0031
relative_error = error / mean  # 0.000039

From the above example, we can see that with TF32 enabled, the speed is ~7x faster on A100, and that relative error compared to double precision is approximately 2 orders of magnitude larger. Note that the exact ratio of TF32 to single precision speed depends on the hardware generation, as properties such as the ratio of memory bandwidth to compute as well as the ratio of TF32 to FP32 matmul throughput may vary from generation to generation or model to model. If full FP32 precision is needed, users can disable TF32 by:

torch.backends.cuda.matmul.allow_tf32 = False
torch.backends.cudnn.allow_tf32 = False

To toggle the TF32 flags off in C++, you can do

at::globalContext().setAllowTF32CuBLAS(false);
at::globalContext().setAllowTF32CuDNN(false);

For more information about TF32, see:

Reduced Precision Reduction in FP16 GEMMs¶

fp16 GEMMs are potentially done with some intermediate reduced precision reductions (e.g., in fp16 rather than fp32). These selective reductions in precision can allow for higher performance on certain workloads (particularly those with a large k dimension) and GPU architectures at the cost of numerical precision and potential for overflow.

Some example benchmark data on V100:

[--------------------------- bench_gemm_transformer --------------------------]
      [  m ,  k  ,  n  ]    |  allow_fp16_reduc=True  |  allow_fp16_reduc=False
1 threads: --------------------------------------------------------------------
      [4096, 4048, 4096]    |           1634.6        |           1639.8
      [4096, 4056, 4096]    |           1670.8        |           1661.9
      [4096, 4080, 4096]    |           1664.2        |           1658.3
      [4096, 4096, 4096]    |           1639.4        |           1651.0
      [4096, 4104, 4096]    |           1677.4        |           1674.9
      [4096, 4128, 4096]    |           1655.7        |           1646.0
      [4096, 4144, 4096]    |           1796.8        |           2519.6
      [4096, 5096, 4096]    |           2094.6        |           3190.0
      [4096, 5104, 4096]    |           2144.0        |           2663.5
      [4096, 5112, 4096]    |           2149.1        |           2766.9
      [4096, 5120, 4096]    |           2142.8        |           2631.0
      [4096, 9728, 4096]    |           3875.1        |           5779.8
      [4096, 16384, 4096]   |           6182.9        |           9656.5
(times in microseconds).

If full precision reductions are needed, users can disable reduced precision reductions in fp16 GEMMs with:

torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False

To toggle the reduced precision reduction flags in C++, one can do

at::globalContext().setAllowFP16ReductionCuBLAS(false);

Reduced Precision Reduction in BF16 GEMMs¶

A similar flag (as above) exists for BFloat16 GEMMs. Note that this switch is set to True by default for BF16, if you observe numerical instability in your workload, you may wish to set it to False.

If reduced precision reductions are not desired, users can disable reduced precision reductions in bf16 GEMMs with:

torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False

To toggle the reduced precision reduction flags in C++, one can do

at::globalContext().setAllowBF16ReductionCuBLAS(true);

Asynchronous execution¶

By default, GPU operations are asynchronous. When you call a function that uses the GPU, the operations are enqueued to the particular device, but not necessarily executed until later. This allows us to execute more computations in parallel, including operations on CPU or other GPUs.

In general, the effect of asynchronous computation is invisible to the caller, because (1) each device executes operations in the order they are queued, and (2) PyTorch automatically performs necessary synchronization when copying data between CPU and GPU or between two GPUs. Hence, computation will proceed as if every operation was executed synchronously.

You can force synchronous computation by setting environment variable CUDA_LAUNCH_BLOCKING=1. This can be handy when an error occurs on the GPU. (With asynchronous execution, such an error isn’t reported until after the operation is actually executed, so the stack trace does not show where it was requested.)

A consequence of the asynchronous computation is that time measurements without synchronizations are not accurate. To get precise measurements, one should either call torch.cuda.synchronize() before measuring, or use torch.cuda.Event to record times as following:

start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()

# Run some things here

end_event.record()
torch.cuda.synchronize()  # Wait for the events to be recorded!
elapsed_time_ms = start_event.elapsed_time(end_event)

As an exception, several functions such as to() and copy_() admit an explicit non_blocking argument, which lets the caller bypass synchronization when it is unnecessary. Another exception is CUDA streams, explained below.

CUDA streams¶

A CUDA stream is a linear sequence of execution that belongs to a specific device. You normally do not need to create one explicitly: by default, each device uses its own “default” stream.

Operations inside each stream are serialized in the order they are created, but operations from different streams can execute concurrently in any relative order, unless explicit synchronization functions (such as synchronize() or wait_stream()) are used. For example, the following code is incorrect:

cuda = torch.device('cuda')
s = torch.cuda.Stream()  # Create a new stream.
A = torch.empty((100, 100), device=cuda).normal_(0.0, 1.0)
with torch.cuda.stream(s):
    # sum() may start execution before normal_() finishes!
    B = torch.sum(A)

When the “current stream” is the default stream, PyTorch automatically performs necessary synchronization when data is moved around, as explained above. However, when using non-default streams, it is the user’s responsibility to ensure proper synchronization. The fixed version of this example is:

cuda = torch.device('cuda')
s = torch.cuda.Stream()  # Create a new stream.
A = torch.empty((100, 100), device=cuda).normal_(0.0, 1.0)
s.wait_stream(torch.cuda.default_stream(cuda))  # NEW!
with torch.cuda.stream(s):
    B = torch.sum(A)
A.record_stream(s)  # NEW!

There are two new additions. The torch.cuda.Stream.wait_stream() call ensures that the normal_() execution has finished before we start running sum(A) on a side stream. The torch.Tensor.record_stream() (see for more details) ensures that we do not deallocate A before sum(A) has completed. You can also manually wait on the stream at some later point in time with torch.cuda.default_stream(cuda).wait_stream(s) (note that it is pointless to wait immediately, since that will prevent the stream execution from running in parallel with other work on the default stream.) See the documentation for torch.Tensor.record_stream() on more details on when to use one or another.

Note that this synchronization is necessary even when there is no read dependency, e.g., as seen in this example:

cuda = torch.device('cuda')
s = torch.cuda.Stream()  # Create a new stream.
A = torch.empty((100, 100), device=cuda)
s.wait_stream(torch.cuda.default_stream(cuda))  # STILL REQUIRED!
with torch.cuda.stream(s):
    A.normal_(0.0, 1.0)
    A.record_stream(s)

Despite the computation on s not reading the contents of A and no other uses of A, it is still necessary to synchronize, because A may correspond to memory reallocated by the CUDA caching allocator, with pending operations from the old (deallocated) memory.

Stream semantics of backward passes¶

Each backward CUDA op runs on the same stream that was used for its corresponding forward op. If your forward pass runs independent ops in parallel on different streams, this helps the backward pass exploit that same parallelism.

The stream semantics of a backward call with respect to surrounding ops are the same as for any other call. The backward pass inserts internal syncs to ensure this even when backward ops run on multiple streams as described in the previous paragraph. More concretely, when calling autograd.backward, autograd.grad, or tensor.backward, and optionally supplying CUDA tensor(s) as the initial gradient(s) (e.g., autograd.backward(..., grad_tensors=initial_grads), autograd.grad(..., grad_outputs=initial_grads), or tensor.backward(..., gradient=initial_grad)), the acts of

optionally populating initial gradient(s),
invoking the backward pass, and
using the gradients

have the same stream-semantics relationship as any group of ops:

s = torch.cuda.Stream()

# Safe, grads are used in the same stream context as backward()
with torch.cuda.stream(s):
    loss.backward()
    use grads

# Unsafe
with torch.cuda.stream(s):
    loss.backward()
use grads

# Safe, with synchronization
with torch.cuda.stream(s):
    loss.backward()
torch.cuda.current_stream().wait_stream(s)
use grads

# Safe, populating initial grad and invoking backward are in the same stream context
with torch.cuda.stream(s):
    loss.backward(gradient=torch.ones_like(loss))

# Unsafe, populating initial_grad and invoking backward are in different stream contexts,
# without synchronization
initial_grad = torch.ones_like(loss)
with torch.cuda.stream(s):
    loss.backward(gradient=initial_grad)

# Safe, with synchronization
initial_grad = torch.ones_like(loss)
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
    initial_grad.record_stream(s)
    loss.backward(gradient=initial_grad)

BC note: Using grads on the default stream¶

In prior versions of PyTorch (1.9 and earlier), the autograd engine always synced the default stream with all backward ops, so the following pattern:

with torch.cuda.stream(s):
    loss.backward()
use grads

was safe as long as use grads happened on the default stream. In present PyTorch, that pattern is no longer safe. If backward() and use grads are in different stream contexts, you must sync the streams:

with torch.cuda.stream(s):
    loss.backward()
torch.cuda.current_stream().wait_stream(s)
use grads

even if use grads is on the default stream.

Memory management¶

PyTorch uses a caching memory allocator to speed up memory allocations. This allows fast memory deallocation without device synchronizations. However, the unused memory managed by the allocator will still show as if used in nvidia-smi. You can use memory_allocated() and max_memory_allocated() to monitor memory occupied by tensors, and use memory_reserved() and max_memory_reserved() to monitor the total amount of memory managed by the caching allocator. Calling empty_cache() releases all unused cached memory from PyTorch so that those can be used by other GPU applications. However, the occupied GPU memory by tensors will not be freed so it can not increase the amount of GPU memory available for PyTorch.

To better understand how CUDA memory is being used over time, Understanding CUDA Memory Usage describes tools for capturing and visualizing traces of memory use.

For more advanced users, we offer more comprehensive memory benchmarking via memory_stats(). We also offer the capability to capture a complete snapshot of the memory allocator state via memory_snapshot(), which can help you understand the underlying allocation patterns produced by your code.

Optimizing memory usage with `PYTORCH_CUDA_ALLOC_CONF`¶

Use of a caching allocator can interfere with memory checking tools such as cuda-memcheck. To debug memory errors using cuda-memcheck, set PYTORCH_NO_CUDA_MEMORY_CACHING=1 in your environment to disable caching.

The behavior of the caching allocator can be controlled via the environment variable PYTORCH_CUDA_ALLOC_CONF. The format is PYTORCH_CUDA_ALLOC_CONF=<option>:<value>,<option2>:<value2>... Available options:

backend allows selecting the underlying allocator implementation. Currently, valid options are native, which uses PyTorch’s native implementation, and cudaMallocAsync, which uses CUDA’s built-in asynchronous allocator. cudaMallocAsync requires CUDA 11.4 or newer. The default is native. backend applies to all devices used by the process, and can’t be specified on a per-device basis.
max_split_size_mb prevents the native allocator from splitting blocks larger than this size (in MB). This can reduce fragmentation and may allow some borderline workloads to complete without running out of memory. Performance cost can range from ‘zero’ to ‘substantial’ depending on allocation patterns. Default value is unlimited, i.e. all blocks can be split. The memory_stats() and memory_summary() methods are useful for tuning. This option should be used as a last resort for a workload that is aborting due to ‘out of memory’ and showing a large amount of inactive split blocks. max_split_size_mb is only meaningful with backend:native. With backend:cudaMallocAsync, max_split_size_mb is ignored.
roundup_power2_divisions helps with rounding the requested allocation size to nearest power-2 division and making better use of the blocks. In the native CUDACachingAllocator, the sizes are rounded up in multiple of blocks size of 512, so this works fine for smaller sizes. However, this can be inefficient for large near-by allocations as each will go to different size of blocks and re-use of those blocks are minimized. This might create lots of unused blocks and will waste GPU memory capacity. This option enables the rounding of allocation size to nearest power-2 division. For example, if we need to round-up size of 1200 and if number of divisions is 4, the size 1200 lies between 1024 and 2048 and if we do 4 divisions between them, the values are 1024, 1280, 1536, and 1792. So, allocation size of 1200 will be rounded to 1280 as the nearest ceiling of power-2 division. Specify a single value to apply for all allocation sizes or specify an array of key value pairs to set power-2 division individually for each power of two interval. For example to set 1 division for all allocations under 256MB, 2 division for allocations between 256MB and 512MB, 4 divisions for allocations between 512MB and 1GB and 8 divisions for any larger allocations, set the knob value to: [256:1,512:2,1024:4,>:8]. roundup_power2_divisions is only meaningful with backend:native. With backend:cudaMallocAsync, roundup_power2_divisions is ignored.
max_non_split_rounding_mb will allow non-split blocks for better reuse, eg,
a 1024MB cached block can be re-used for a 512MB allocation request. In the default case, we only allow up to 20MB of rounding of non-split blocks, so a 512MB block can only be served with between 512-532 MB size block. If we set the value of this option to 1024, it will alow 512-1536 MB size blocks to be used for a 512MB block which increases reuse of larger blocks. This will also help in reducing the stalls in avoiding expensive cudaMalloc calls.
garbage_collection_threshold helps actively reclaiming unused GPU memory to avoid triggering expensive sync-and-reclaim-all operation (release_cached_blocks), which can be unfavorable to latency-critical GPU applications (e.g., servers). Upon setting this threshold (e.g., 0.8), the allocator will start reclaiming GPU memory blocks if the GPU memory capacity usage exceeds the threshold (i.e., 80% of the total memory allocated to the GPU application). The algorithm prefers to free old & unused blocks first to avoid freeing blocks that are actively being reused. The threshold value should be between greater than 0.0 and less than 1.0. garbage_collection_threshold is only meaningful with backend:native. With backend:cudaMallocAsync, garbage_collection_threshold is ignored.
expandable_segments (experimental, default: False) If set to True, this setting instructs the allocator to create CUDA allocations that can later be expanded to better handle cases where a job changing allocation sizes frequently, such as having a changing batch size. Normally for large (>2MB) allocations, the allocator calls cudaMalloc to get allocations that are the same size as what the user requests. In the future, parts of these allocations can be reused for other requests if they are free. This works well when the program makes many requests of exactly the same size or of sizes that even multiples of that size. Many deep learning models follow this behavior. However, one common exception is when the batch size changes slightly from one iteration to the next, e.g. in batched inference. When the program runs initially with batch size N, it will make allocations appropriate for that size. If in the future, it runs at size N - 1, the existing allocations will still be big enough. However, if it runs at size N + 1, then it will have to make new allocations that are slightly larger. Not all the tensors are the same size. Some might be (N + 1)*A and others (N + 1)*A*B where A and B are some non-batch dimensions in the model. Because the allocator reuses existing allocations when they are big enough, some number of (N + 1)*A allocations will actually fit in the already existing N*B*A segments, though not perfectly. As the model runs it will partially fill up all of these segments leaving unusable free slices of memory at the end of these segments. The allocator at some point will need to cudaMalloc a new (N + 1)*A*B segment. If there is not enough memory, there is now no way to recover the slices of memory that are free at the end of existing segments. With models 50+ layers deep, this pattern might repeat 50+ times creating many slivers.

expandable_segments allows the allocator to create a segment initially and then expand its size later when more memory is needed. Instead of making one segment per allocation, it tries to make one segment (per stream) that grows as necessary. Now when the N + 1 case runs, the allocations will tile nicely into the one large segment until it fills up. Then more memory is requested and appended to the end of the segment. This process does not create as many slivers of unusable memory, so it is more likely to succeed at finding this memory.

pinned_use_cuda_host_register option is a boolean flag that determines whether to use the CUDA API’s cudaHostRegister function for allocating pinned memory instead of the default cudaHostAlloc. When set to True, the memory is allocated using regular malloc and then pages are mapped to the memory before calling cudaHostRegister. This pre-mapping of pages helps reduce the lock time during the execution of cudaHostRegister.

pinned_num_register_threads option is only valid when pinned_use_cuda_host_register is set to True. By default, one thread is used to map the pages. This option allows using more threads to parallelize the page mapping operations to reduce the overall allocation time of pinned memory. A good value for this option is 8 based on benchmarking results.

pinned_use_background_threads option is a boolean flag to enable background thread for processing events. This avoids any slow path associated with querying/processing of events in the fast allocation path. This feature is disabled by default.

Note

Some stats reported by the CUDA memory management API are specific to backend:native, and are not meaningful with backend:cudaMallocAsync. See each function’s docstring for details.

Using custom memory allocators for CUDA¶

It is possible to define allocators as simple functions in C/C++ and compile them as a shared library, the code below shows a basic allocator that just traces all the memory operations.

#include <sys/types.h>
#include <cuda_runtime_api.h>
#include <iostream>
// Compile with g++ alloc.cc -o alloc.so -I/usr/local/cuda/include -shared -fPIC
extern "C" {
void* my_malloc(ssize_t size, int device, cudaStream_t stream) {
   void *ptr;
   cudaMalloc(&ptr, size);
   std::cout<<"alloc "<<ptr<<size<<std::endl;
   return ptr;
}

void my_free(void* ptr, ssize_t size, int device, cudaStream_t stream) {
   std::cout<<"free "<<ptr<< " "<<stream<<std::endl;
   cudaFree(ptr);
}
}

This can be used in python through the torch.cuda.memory.CUDAPluggableAllocator. The user is responsible for supplying the path to the .so file and the name of the alloc/free functions that match the signatures specified above.

import torch

# Load the allocator
new_alloc = torch.cuda.memory.CUDAPluggableAllocator(
    'alloc.so', 'my_malloc', 'my_free')
# Swap the current allocator
torch.cuda.memory.change_current_allocator(new_alloc)
# This will allocate memory in the device using the new allocator
b = torch.zeros(10, device='cuda')

import torch

# Do an initial memory allocator
b = torch.zeros(10, device='cuda')
# Load the allocator
new_alloc = torch.cuda.memory.CUDAPluggableAllocator(
    'alloc.so', 'my_malloc', 'my_free')
# This will error since the current allocator was already instantiated
torch.cuda.memory.change_current_allocator(new_alloc)

cuBLAS workspaces¶

For each combination of cuBLAS handle and CUDA stream, a cuBLAS workspace will be allocated if that handle and stream combination executes a cuBLAS kernel that requires a workspace. In order to avoid repeatedly allocating workspaces, these workspaces are not deallocated unless torch._C._cuda_clearCublasWorkspaces() is called. The workspace size per allocation can be specified via the environment variable CUBLAS_WORKSPACE_CONFIG with the format :[SIZE]:[COUNT]. As an example, the default workspace size per allocation is CUBLAS_WORKSPACE_CONFIG=:4096:2:16:8 which specifies a total size of 2 * 4096 + 8 * 16 KiB. To force cuBLAS to avoid using workspaces, set CUBLAS_WORKSPACE_CONFIG=:0:0.

cuFFT plan cache¶

For each CUDA device, an LRU cache of cuFFT plans is used to speed up repeatedly running FFT methods (e.g., torch.fft.fft()) on CUDA tensors of same geometry with same configuration. Because some cuFFT plans may allocate GPU memory, these caches have a maximum capacity.

You may control and query the properties of the cache of current device with the following APIs:

torch.backends.cuda.cufft_plan_cache.max_size gives the capacity of the cache (default is 4096 on CUDA 10 and newer, and 1023 on older CUDA versions). Setting this value directly modifies the capacity.
torch.backends.cuda.cufft_plan_cache.size gives the number of plans currently residing in the cache.
torch.backends.cuda.cufft_plan_cache.clear() clears the cache.

To control and query plan caches of a non-default device, you can index the torch.backends.cuda.cufft_plan_cache object with either a torch.device object or a device index, and access one of the above attributes. E.g., to set the capacity of the cache for device 1, one can write torch.backends.cuda.cufft_plan_cache[1].max_size = 10.

Just-in-Time Compilation¶

PyTorch just-in-time compiles some operations, like torch.special.zeta, when performed on CUDA tensors. This compilation can be time consuming (up to a few seconds depending on your hardware and software) and may occur multiple times for a single operator since many PyTorch operators actually select from a variety of kernels, each of which must be compiled once, depending on their input. This compilation occurs once per process, or just once if a kernel cache is used.

By default, PyTorch creates a kernel cache in $XDG_CACHE_HOME/torch/kernels if XDG_CACHE_HOME is defined and $HOME/.cache/torch/kernels if it’s not (except on Windows, where the kernel cache is not yet supported). The caching behavior can be directly controlled with two environment variables. If USE_PYTORCH_KERNEL_CACHE is set to 0 then no cache will be used, and if PYTORCH_KERNEL_CACHE_PATH is set then that path will be used as a kernel cache instead of the default location.

Best practices¶

Device-agnostic code¶

Due to the structure of PyTorch, you may need to explicitly write device-agnostic (CPU or GPU) code; an example may be creating a new tensor as the initial hidden state of a recurrent neural network.

The first step is to determine whether the GPU should be used or not. A common pattern is to use Python’s argparse module to read in user arguments, and have a flag that can be used to disable CUDA, in combination with is_available(). In the following, args.device results in a torch.device object that can be used to move tensors to CPU or CUDA.

import argparse
import torch

parser = argparse.ArgumentParser(description='PyTorch Example')
parser.add_argument('--disable-cuda', action='store_true',
                    help='Disable CUDA')
args = parser.parse_args()
args.device = None
if not args.disable_cuda and torch.cuda.is_available():
    args.device = torch.device('cuda')
else:
    args.device = torch.device('cpu')

Note

When assessing the availability of CUDA in a given environment (is_available()), PyTorch’s default behavior is to call the CUDA Runtime API method cudaGetDeviceCount. Because this call in turn initializes the CUDA Driver API (via cuInit) if it is not already initialized, subsequent forks of a process that has run is_available() will fail with a CUDA initialization error.

One can set PYTORCH_NVML_BASED_CUDA_CHECK=1 in your environment before importing PyTorch modules that execute is_available() (or before executing it directly) in order to direct is_available() to attempt an NVML-based assessment (nvmlDeviceGetCount_v2). If the NVML-based assessment is successful (i.e. NVML discovery/initialization does not fail), is_available() calls will not poison subsequent forks.

If NVML discovery/initialization fails, is_available() will fallback to the standard CUDA Runtime API assessment and the aforementioned fork constraint will apply.

Note that the above NVML-based CUDA availability assessment provides a weaker guarantee than the default CUDA Runtime API approach (which requires CUDA initialization to succeed). In some circumstances, the NVML-based check may succeed while later CUDA initialization fails.

Now that we have args.device, we can use it to create a Tensor on the desired device.

x = torch.empty((8, 42), device=args.device)
net = Network().to(device=args.device)

This can be used in a number of cases to produce device agnostic code. Below is an example when using a dataloader:

cuda0 = torch.device('cuda:0')  # CUDA GPU 0
for i, x in enumerate(train_loader):
    x = x.to(cuda0)

When working with multiple GPUs on a system, you can use the CUDA_VISIBLE_DEVICES environment flag to manage which GPUs are available to PyTorch. As mentioned above, to manually control which GPU a tensor is created on, the best practice is to use a torch.cuda.device context manager.

print("Outside device is 0")  # On device 0 (default in most scenarios)
with torch.cuda.device(1):
    print("Inside device is 1")  # On device 1
print("Outside device is still 0")  # On device 0

If you have a tensor and would like to create a new tensor of the same type on the same device, then you can use a torch.Tensor.new_* method (see torch.Tensor). Whilst the previously mentioned torch.* factory functions (Creation Ops) depend on the current GPU context and the attributes arguments you pass in, torch.Tensor.new_* methods preserve the device and other attributes of the tensor.

This is the recommended practice when creating modules in which new tensors need to be created internally during the forward pass.

cuda = torch.device('cuda')
x_cpu = torch.empty(2)
x_gpu = torch.empty(2, device=cuda)
x_cpu_long = torch.empty(2, dtype=torch.int64)

y_cpu = x_cpu.new_full([3, 2], fill_value=0.3)
print(y_cpu)

    tensor([[ 0.3000,  0.3000],
            [ 0.3000,  0.3000],
            [ 0.3000,  0.3000]])

y_gpu = x_gpu.new_full([3, 2], fill_value=-5)
print(y_gpu)

    tensor([[-5.0000, -5.0000],
            [-5.0000, -5.0000],
            [-5.0000, -5.0000]], device='cuda:0')

y_cpu_long = x_cpu_long.new_tensor([[1, 2, 3]])
print(y_cpu_long)

    tensor([[ 1,  2,  3]])

If you want to create a tensor of the same type and size of another tensor, and fill it with either ones or zeros, ones_like() or zeros_like() are provided as convenient helper functions (which also preserve torch.device and torch.dtype of a Tensor).

x_cpu = torch.empty(2, 3)
x_gpu = torch.empty(2, 3)

y_cpu = torch.ones_like(x_cpu)
y_gpu = torch.zeros_like(x_gpu)

Use pinned memory buffers¶

Warning

This is an advanced tip. If you overuse pinned memory, it can cause serious problems when running low on RAM, and you should be aware that pinning is often an expensive operation.

Host to GPU copies are much faster when they originate from pinned (page-locked) memory. CPU tensors and storages expose a pin_memory() method, that returns a copy of the object, with data put in a pinned region.

Also, once you pin a tensor or storage, you can use asynchronous GPU copies. Just pass an additional non_blocking=True argument to a to() or a cuda() call. This can be used to overlap data transfers with computation.

You can make the DataLoader return batches placed in pinned memory by passing pin_memory=True to its constructor.

Use nn.parallel.DistributedDataParallel instead of multiprocessing or nn.DataParallel¶

Most use cases involving batched inputs and multiple GPUs should default to using DistributedDataParallel to utilize more than one GPU.

There are significant caveats to using CUDA models with multiprocessing; unless care is taken to meet the data handling requirements exactly, it is likely that your program will have incorrect or undefined behavior.

It is recommended to use DistributedDataParallel, instead of DataParallel to do multi-GPU training, even if there is only a single node.

The difference between DistributedDataParallel and DataParallel is: DistributedDataParallel uses multiprocessing where a process is created for each GPU, while DataParallel uses multithreading. By using multiprocessing, each GPU has its dedicated process, this avoids the performance overhead caused by GIL of Python interpreter.

If you use DistributedDataParallel, you could use torch.distributed.launch utility to launch your program, see Third-party backends.

CUDA Graphs¶

A CUDA graph is a record of the work (mostly kernels and their arguments) that a CUDA stream and its dependent streams perform. For general principles and details on the underlying CUDA API, see Getting Started with CUDA Graphs and the Graphs section of the CUDA C Programming Guide.

PyTorch supports the construction of CUDA graphs using stream capture, which puts a CUDA stream in capture mode. CUDA work issued to a capturing stream doesn’t actually run on the GPU. Instead, the work is recorded in a graph.

After capture, the graph can be launched to run the GPU work as many times as needed. Each replay runs the same kernels with the same arguments. For pointer arguments this means the same memory addresses are used. By filling input memory with new data (e.g., from a new batch) before each replay, you can rerun the same work on new data.

Why CUDA Graphs?¶

Replaying a graph sacrifices the dynamic flexibility of typical eager execution in exchange for greatly reduced CPU overhead. A graph’s arguments and kernels are fixed, so a graph replay skips all layers of argument setup and kernel dispatch, including Python, C++, and CUDA driver overheads. Under the hood, a replay submits the entire graph’s work to the GPU with a single call to cudaGraphLaunch. Kernels in a replay also execute slightly faster on the GPU, but eliding CPU overhead is the main benefit.

You should try CUDA graphs if all or part of your network is graph-safe (usually this means static shapes and static control flow, but see the other constraints) and you suspect its runtime is at least somewhat CPU-limited.

PyTorch API¶

Warning

This API is in beta and may change in future releases.

PyTorch exposes graphs via a raw torch.cuda.CUDAGraph class and two convenience wrappers, torch.cuda.graph and torch.cuda.make_graphed_callables.

torch.cuda.graph is a simple, versatile context manager that captures CUDA work in its context. Before capture, warm up the workload to be captured by running a few eager iterations. Warmup must occur on a side stream. Because the graph reads from and writes to the same memory addresses in every replay, you must maintain long-lived references to tensors that hold input and output data during capture. To run the graph on new input data, copy new data to the capture’s input tensor(s), replay the graph, then read the new output from the capture’s output tensor(s). Example:

g = torch.cuda.CUDAGraph()

# Placeholder input used for capture
static_input = torch.empty((5,), device="cuda")

# Warmup before capture
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
    for _ in range(3):
        static_output = static_input * 2
torch.cuda.current_stream().wait_stream(s)

# Captures the graph
# To allow capture, automatically sets a side stream as the current stream in the context
with torch.cuda.graph(g):
    static_output = static_input * 2

# Fills the graph's input memory with new data to compute on
static_input.copy_(torch.full((5,), 3, device="cuda"))
g.replay()
# static_output holds the results
print(static_output)  # full of 3 * 2 = 6

# Fills the graph's input memory with more data to compute on
static_input.copy_(torch.full((5,), 4, device="cuda"))
g.replay()
print(static_output)  # full of 4 * 2 = 8

See Whole-network capture, Usage with torch.cuda.amp, and Usage with multiple streams for realistic and advanced patterns.

make_graphed_callables is more sophisticated. make_graphed_callables accepts Python functions and torch.nn.Modules. For each passed function or Module, it creates separate graphs of the forward-pass and backward-pass work. See Partial-network capture.

Constraints¶

A set of ops is capturable if it doesn’t violate any of the following constraints.

Constraints apply to all work in a torch.cuda.graph context and all work in the forward and backward passes of any callable you pass to torch.cuda.make_graphed_callables().

Violating any of these will likely cause a runtime error:

Capture must occur on a non-default stream. (This is only a concern if you use the raw CUDAGraph.capture_begin and CUDAGraph.capture_end calls. graph and make_graphed_callables() set a side stream for you.)
Ops that synchronize the CPU with the GPU (e.g., .item() calls) are prohibited.
CUDA RNG operations are permitted, and when using multiple torch.Generator instances within a graph, they must be registered using CUDAGraph.register_generator_state before graph capture. Avoid using Generator.get_state and Generator.set_state during capture; instead, utilize Generator.graphsafe_set_state and Generator.graphsafe_get_state for managing generator states safely within the graph context. This ensures proper RNG operation and generator management within CUDA graphs.

Violating any of these will likely cause silent numerical errors or undefined behavior:

Within a process, only one capture may be underway at a time.
No non-captured CUDA work may run in this process (on any thread) while capture is underway.
CPU work is not captured. If the captured ops include CPU work, that work will be elided during replay.
Every replay reads from and writes to the same (virtual) memory addresses.
Dynamic control flow (based on CPU or GPU data) is prohibited.
Dynamic shapes are prohibited. The graph assumes every tensor in the captured op sequence has the same size and layout in every replay.
Using multiple streams in a capture is allowed, but there are restrictions.

Non-constraints¶

Once captured, the graph may be replayed on any stream.

Whole-network capture¶

If your entire network is capturable, you can capture and replay an entire iteration:

N, D_in, H, D_out = 640, 4096, 2048, 1024
model = torch.nn.Sequential(torch.nn.Linear(D_in, H),
                            torch.nn.Dropout(p=0.2),
                            torch.nn.Linear(H, D_out),
                            torch.nn.Dropout(p=0.1)).cuda()
loss_fn = torch.nn.MSELoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

# Placeholders used for capture
static_input = torch.randn(N, D_in, device='cuda')
static_target = torch.randn(N, D_out, device='cuda')

# warmup
# Uses static_input and static_target here for convenience,
# but in a real setting, because the warmup includes optimizer.step()
# you must use a few batches of real data.
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
    for i in range(3):
        optimizer.zero_grad(set_to_none=True)
        y_pred = model(static_input)
        loss = loss_fn(y_pred, static_target)
        loss.backward()
        optimizer.step()
torch.cuda.current_stream().wait_stream(s)

# capture
g = torch.cuda.CUDAGraph()
# Sets grads to None before capture, so backward() will create
# .grad attributes with allocations from the graph's private pool
optimizer.zero_grad(set_to_none=True)
with torch.cuda.graph(g):
    static_y_pred = model(static_input)
    static_loss = loss_fn(static_y_pred, static_target)
    static_loss.backward()
    optimizer.step()

real_inputs = [torch.rand_like(static_input) for _ in range(10)]
real_targets = [torch.rand_like(static_target) for _ in range(10)]

for data, target in zip(real_inputs, real_targets):
    # Fills the graph's input memory with new data to compute on
    static_input.copy_(data)
    static_target.copy_(target)
    # replay() includes forward, backward, and step.
    # You don't even need to call optimizer.zero_grad() between iterations
    # because the captured backward refills static .grad tensors in place.
    g.replay()
    # Params have been updated. static_y_pred, static_loss, and .grad
    # attributes hold values from computing on this iteration's data.

Partial-network capture¶

If some of your network is unsafe to capture (e.g., due to dynamic control flow, dynamic shapes, CPU syncs, or essential CPU-side logic), you can run the unsafe part(s) eagerly and use torch.cuda.make_graphed_callables() to graph only the capture-safe part(s).

By default, callables returned by make_graphed_callables() are autograd-aware, and can be used in the training loop as direct replacements for the functions or nn.Modules you passed.

make_graphed_callables() internally creates CUDAGraph objects, runs warmup iterations, and maintains static inputs and outputs as needed. Therefore (unlike with torch.cuda.graph) you don’t need to handle those manually.

In the following example, data-dependent dynamic control flow means the network isn’t capturable end-to-end, but make_graphed_callables() lets us capture and run graph-safe sections as graphs regardless:

N, D_in, H, D_out = 640, 4096, 2048, 1024

module1 = torch.nn.Linear(D_in, H).cuda()
module2 = torch.nn.Linear(H, D_out).cuda()
module3 = torch.nn.Linear(H, D_out).cuda()

loss_fn = torch.nn.MSELoss()
optimizer = torch.optim.SGD(chain(module1.parameters(),
                                  module2.parameters(),
                                  module3.parameters()),
                            lr=0.1)

# Sample inputs used for capture
# requires_grad state of sample inputs must match
# requires_grad state of real inputs each callable will see.
x = torch.randn(N, D_in, device='cuda')
h = torch.randn(N, H, device='cuda', requires_grad=True)

module1 = torch.cuda.make_graphed_callables(module1, (x,))
module2 = torch.cuda.make_graphed_callables(module2, (h,))
module3 = torch.cuda.make_graphed_callables(module3, (h,))

real_inputs = [torch.rand_like(x) for _ in range(10)]
real_targets = [torch.randn(N, D_out, device="cuda") for _ in range(10)]

for data, target in zip(real_inputs, real_targets):
    optimizer.zero_grad(set_to_none=True)

    tmp = module1(data)  # forward ops run as a graph

    if tmp.sum().item() > 0:
        tmp = module2(tmp)  # forward ops run as a graph
    else:
        tmp = module3(tmp)  # forward ops run as a graph

    loss = loss_fn(tmp, target)
    # module2's or module3's (whichever was chosen) backward ops,
    # as well as module1's backward ops, run as graphs
    loss.backward()
    optimizer.step()

Usage with torch.cuda.amp¶

For typical optimizers, GradScaler.step syncs the CPU with the GPU, which is prohibited during capture. To avoid errors, either use partial-network capture, or (if forward, loss, and backward are capture-safe) capture forward, loss, and backward but not the optimizer step:

# warmup
# In a real setting, use a few batches of real data.
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
    for i in range(3):
        optimizer.zero_grad(set_to_none=True)
        with torch.cuda.amp.autocast():
            y_pred = model(static_input)
            loss = loss_fn(y_pred, static_target)
        scaler.scale(loss).backward()
        scaler.step(optimizer)
        scaler.update()
torch.cuda.current_stream().wait_stream(s)

# capture
g = torch.cuda.CUDAGraph()
optimizer.zero_grad(set_to_none=True)
with torch.cuda.graph(g):
    with torch.cuda.amp.autocast():
        static_y_pred = model(static_input)
        static_loss = loss_fn(static_y_pred, static_target)
    scaler.scale(static_loss).backward()
    # don't capture scaler.step(optimizer) or scaler.update()

real_inputs = [torch.rand_like(static_input) for _ in range(10)]
real_targets = [torch.rand_like(static_target) for _ in range(10)]

for data, target in zip(real_inputs, real_targets):
    static_input.copy_(data)
    static_target.copy_(target)
    g.replay()
    # Runs scaler.step and scaler.update eagerly
    scaler.step(optimizer)
    scaler.update()

Usage with multiple streams¶

Capture mode automatically propagates to any streams that sync with a capturing stream. Within capture, you may expose parallelism by issuing calls to different streams, but the overall stream dependency DAG must branch out from the initial capturing stream after capture begins and rejoin the initial stream before capture ends:

with torch.cuda.graph(g):
    # at context manager entrance, torch.cuda.current_stream()
    # is the initial capturing stream

    # INCORRECT (does not branch out from or rejoin initial stream)
    with torch.cuda.stream(s):
        cuda_work()

    # CORRECT:
    # branches out from initial stream
    s.wait_stream(torch.cuda.current_stream())
    with torch.cuda.stream(s):
        cuda_work()
    # rejoins initial stream before capture ends
    torch.cuda.current_stream().wait_stream(s)

Note

To avoid confusion for power users looking at replays in nsight systems or nvprof: Unlike eager execution, the graph interprets a nontrivial stream DAG in capture as a hint, not a command. During replay, the graph may reorganize independent ops onto different streams or enqueue them in a different order (while respecting your original DAG’s overall dependencies).

Usage with DistributedDataParallel¶

NCCL < 2.9.6¶

NCCL versions earlier than 2.9.6 don’t allow collectives to be captured. You must use partial-network capture, which defers allreduces to happen outside graphed sections of backward.

Call make_graphed_callables() on graphable network sections before wrapping the network with DDP.

NCCL >= 2.9.6¶

NCCL versions 2.9.6 or later allow collectives in the graph. Approaches that capture an entire backward pass are a viable option, but need three setup steps.

Disable DDP’s internal async error handling:

os.environ["NCCL_ASYNC_ERROR_HANDLING"] = "0"
torch.distributed.init_process_group(...)

Before full-backward capture, DDP must be constructed in a side-stream context:

with torch.cuda.stream(s):
    model = DistributedDataParallel(model)

Your warmup must run at least 11 DDP-enabled eager iterations before capture.

Graph memory management¶

A captured graph acts on the same virtual addresses every time it replays. If PyTorch frees the memory, a later replay can hit an illegal memory access. If PyTorch reassigns the memory to new tensors, the replay can corrupt the values seen by those tensors. Therefore, the virtual addresses used by the graph must be reserved for the graph across replays. The PyTorch caching allocator achieves this by detecting when capture is underway and satisfying the capture’s allocations from a graph-private memory pool. The private pool stays alive until its CUDAGraph object and all tensors created during capture go out of scope.

Private pools are maintained automatically. By default, the allocator creates a separate private pool for each capture. If you capture multiple graphs, this conservative approach ensures graph replays never corrupt each other’s values, but sometimes needlessly wastes memory.

CUDA semantics¶

TensorFloat-32 (TF32) on Ampere (and later) devices¶

Reduced Precision Reduction in FP16 GEMMs¶

Reduced Precision Reduction in BF16 GEMMs¶

Asynchronous execution¶

CUDA streams¶

Stream semantics of backward passes¶

BC note: Using grads on the default stream¶

Memory management¶

Optimizing memory usage with `PYTORCH_CUDA_ALLOC_CONF`¶

Using custom memory allocators for CUDA¶

cuBLAS workspaces¶

cuFFT plan cache¶

Just-in-Time Compilation¶

Best practices¶

Device-agnostic code¶

Use pinned memory buffers¶

Use nn.parallel.DistributedDataParallel instead of multiprocessing or nn.DataParallel¶

CUDA Graphs¶

Why CUDA Graphs?¶

PyTorch API¶

Constraints¶

Non-constraints¶

Whole-network capture¶

Partial-network capture¶

Usage with torch.cuda.amp¶

Usage with multiple streams¶

Usage with DistributedDataParallel¶

NCCL < 2.9.6¶

NCCL >= 2.9.6¶

Graph memory management¶

Docs

Tutorials

Resources

CUDA semantics¶

TensorFloat-32 (TF32) on Ampere (and later) devices¶

Reduced Precision Reduction in FP16 GEMMs¶

Reduced Precision Reduction in BF16 GEMMs¶

Asynchronous execution¶

CUDA streams¶

Stream semantics of backward passes¶

BC note: Using grads on the default stream¶

Memory management¶

Optimizing memory usage with PYTORCH_CUDA_ALLOC_CONF¶

Using custom memory allocators for CUDA¶

cuBLAS workspaces¶

cuFFT plan cache¶

Just-in-Time Compilation¶

Best practices¶

Device-agnostic code¶

Use pinned memory buffers¶

Use nn.parallel.DistributedDataParallel instead of multiprocessing or nn.DataParallel¶

CUDA Graphs¶

Why CUDA Graphs?¶

PyTorch API¶

Constraints¶

Non-constraints¶

Whole-network capture¶

Partial-network capture¶

Usage with torch.cuda.amp¶

Usage with multiple streams¶

Usage with DistributedDataParallel¶

NCCL < 2.9.6¶

NCCL >= 2.9.6¶

Graph memory management¶

Sharing memory across captures¶

Docs

Tutorials

Resources

Optimizing memory usage with `PYTORCH_CUDA_ALLOC_CONF`¶