(reproducibility)=

Reproducibility

Completely reproducible results are not guaranteed across PyTorch releases, individual commits, or different platforms. Furthermore, results may not be reproducible between CPU and GPU executions, even when using identical seeds.

However, there are some steps you can take to limit the number of sources of nondeterministic behavior for a specific platform, device, and PyTorch release. First, you can control sources of randomness that can cause multiple executions of your application to behave differently. Second, you can configure PyTorch to avoid using nondeterministic algorithms for some operations, so that multiple calls to those operations, given the same inputs, will produce the same result.

:::{warning} Deterministic operations are often slower than nondeterministic operations, so single-run performance may decrease for your model. However, determinism may save time in development by facilitating experimentation, debugging, and regression testing. :::

Controlling sources of randomness

PyTorch random number generator

You can use {meth}torch.manual_seed() to seed the RNG for all devices (both CPU and CUDA):

python

import torch
torch.manual_seed(0)

Some PyTorch operations may use random numbers internally. {meth}torch.svd_lowrank() does this, for instance. Consequently, calling it multiple times back-to-back with the same input arguments may give different results. However, as long as {meth}torch.manual_seed() is set to a constant at the beginning of an application and all other sources of nondeterminism have been eliminated, the same series of random numbers will be generated each time the application is run in the same environment.

It is also possible to obtain identical results from an operation that uses random numbers by setting {meth}torch.manual_seed() to the same value between subsequent calls.

Python

For custom operators, you might need to set python seed as well:

python

import random
random.seed(0)

Random number generators in other libraries

If you or any of the libraries you are using rely on NumPy, you can seed the global NumPy RNG with:

python

import numpy as np
np.random.seed(0)

However, some applications and libraries may use NumPy Random Generator objects, not the global RNG (https://numpy.org/doc/stable/reference/random/generator.html), and those will need to be seeded consistently as well.

If you are using any other libraries that use random number generators, refer to the documentation for those libraries to see how to set consistent seeds for them.

CUDA convolution benchmarking

The cuDNN library, used by CUDA convolution operations, can be a source of nondeterminism across multiple executions of an application. When a cuDNN convolution is called with a new set of size parameters, an optional feature can run multiple convolution algorithms, benchmarking them to find the fastest one. Then, the fastest algorithm will be used consistently during the rest of the process for the corresponding set of size parameters. Due to benchmarking noise and different hardware, the benchmark may select different algorithms on subsequent runs, even on the same machine.

Disabling the benchmarking feature with torch.backends.cudnn.benchmark = False causes cuDNN to deterministically select an algorithm, possibly at the cost of reduced performance.

However, if you do not need reproducibility across multiple executions of your application, then performance might improve if the benchmarking feature is enabled with torch.backends.cudnn.benchmark = True.

Note that this setting is different from the torch.backends.cudnn.deterministic setting discussed below.

Avoiding nondeterministic algorithms

{meth}torch.use_deterministic_algorithms lets you configure PyTorch to use deterministic algorithms instead of nondeterministic ones where available, and to throw an error if an operation is known to be nondeterministic (and without a deterministic alternative).

Please check the documentation for {meth}torch.use_deterministic_algorithms() for a full list of affected operations. If an operation does not act correctly according to the documentation, or if you need a deterministic implementation of an operation that does not have one, please submit an issue: https://github.com/pytorch/pytorch/issues?q=label:%22module:%20determinism%22

For example, running the nondeterministic CUDA implementation of {meth}torch.Tensor.index_add_ will throw an error:

python

>>> import torch
>>> torch.use_deterministic_algorithms(True)
>>> torch.randn(2, 2).cuda().index_add_(0, torch.tensor([0, 1]), torch.randn(2, 2))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: index_add_cuda_ does not have a deterministic implementation, but you set
'torch.use_deterministic_algorithms(True)'. ...

When {meth}torch.bmm is called with sparse-dense CUDA tensors it typically uses a nondeterministic algorithm, but when the deterministic flag is turned on, its alternate deterministic implementation will be used:

python

>>> import torch
>>> torch.use_deterministic_algorithms(True)
>>> torch.bmm(torch.randn(2, 2, 2).to_sparse().cuda(), torch.randn(2, 2, 2).cuda())
tensor([[[ 1.1900, -2.3409],
         [ 0.4796,  0.8003]],
        [[ 0.1509,  1.8027],
         [ 0.0333, -1.1444]]], device='cuda:0')

CUDA convolution determinism

While disabling CUDA convolution benchmarking (discussed above) ensures that CUDA selects the same algorithm each time an application is run, that algorithm itself may be nondeterministic, unless either torch.use_deterministic_algorithms(True) or torch.backends.cudnn.deterministic = True is set. The latter setting controls only this behavior, unlike {meth}torch.use_deterministic_algorithms which will make other PyTorch operations behave deterministically, too.

CUDA Scaled Dot Product Attention

{func}torch.nn.functional.scaled_dot_product_attention (SDPA) dispatches to multiple backends at runtime. Each backend has different determinism characteristics, summarized in the table below:

{list-table}

:header-rows: 1
:widths: 25 15 15 45

* - Backend
  - Forward
  - Backward
  - Notes
* - `SDPBackend.MATH`
  - Deterministic
  - Deterministic
  - Uses standard PyTorch operators (`matmul`, `softmax`). Deterministic
    when {meth}`torch.use_deterministic_algorithms` is enabled.
* - `SDPBackend.FLASH_ATTENTION`
  - Deterministic
  - Non-deterministic
  - The backward pass uses non-deterministic atomic operations by default.
    Setting `torch.use_deterministic_algorithms(True, warn_only=False)`
    enables a deterministic backward implementation.
* - `SDPBackend.EFFICIENT_ATTENTION`
  - Deterministic
  - Non-deterministic
  - The backward pass may split work across keys (`num_splits_key > 1`)
    for performance, which is non-deterministic. Setting
    `torch.use_deterministic_algorithms(True, warn_only=False)` forces
    `num_splits_key = 1`, making the backward pass deterministic.
* - `SDPBackend.CUDNN_ATTENTION`
  - Deterministic
  - Non-deterministic
  - cuDNN provides an opt-in deterministic backward for `float16`
    starting in
    [cuDNN 9.18](https://docs.nvidia.com/deeplearning/cudnn/backend/latest/release-notes.html#cudnn-9-18-0)
    (via
    [set_deterministic_algorithm](https://docs.nvidia.com/deeplearning/cudnn/frontend/latest/operations/Attention.html#sdpa-fp16-bf16-backward)),
    but this has not yet been integrated into PyTorch. This backend is
    **disabled** when
    `torch.use_deterministic_algorithms(True, warn_only=False)` is set,
    regardless of whether inputs require gradients or the call is
    inside a `torch.no_grad()` context.

When torch.use_deterministic_algorithms(True, warn_only=True) is set, the fused backends (Flash, Efficient, and cuDNN) emit a one-time warning but still use their default non-deterministic code paths. To actually enforce determinism, pass warn_only=False.

Low-precision dtypes and numerical reproducibility

Bitwise matching numerics across different SDPA backends are not guaranteed, even for the same inputs and dtype. Each backend performs floating-point accumulation in a different order, and because floating-point addition is not associative, the results will differ between backends.

The Math backend by default accumulates in float32 even when given float16 or bfloat16 inputs. The function {func}torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp can enable reduced-precision accumulation for higher performance at the cost of different numerical results.

Selecting a specific backend

Use the {func}torch.nn.attention.sdpa_kernel context manager to restrict which backends SDPA may use. For example, to guarantee deterministic behavior regardless of hardware:

python

import torch
from torch.nn.attention import sdpa_kernel, SDPBackend

torch.use_deterministic_algorithms(True)

# Option 1: Use only the Math backend (always deterministic)
with sdpa_kernel(SDPBackend.MATH):
    out = torch.nn.functional.scaled_dot_product_attention(q, k, v)

# Option 2: Allow Flash and Efficient (deterministic with the flag above)
with sdpa_kernel([SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION]):
    out = torch.nn.functional.scaled_dot_product_attention(q, k, v)

CUDA RNN and LSTM

In some versions of CUDA, RNNs and LSTM networks may have non-deterministic behavior. See {meth}torch.nn.RNN and {meth}torch.nn.LSTM for details and workarounds.

Filling uninitialized memory

Operations like {meth}torch.empty and {meth}torch.Tensor.resize_ can return tensors with uninitialized memory that contain undefined values. Using such a tensor as an input to another operation is invalid if determinism is required, because the output will be nondeterministic. But there is nothing to actually prevent such invalid code from being run. So for safety, {attr}torch.utils.deterministic.fill_uninitialized_memory is set to True by default, which will fill the uninitialized memory with a known value if torch.use_deterministic_algorithms(True) is set. This will prevent the possibility of this kind of nondeterministic behavior.

However, filling uninitialized memory is detrimental to performance. So if your program is valid and does not use uninitialized memory as the input to an operation, then this setting can be turned off for better performance.

DataLoader

DataLoader will reseed workers following the {ref}data-loading-randomness algorithm. Use {meth}worker_init_fn and generator to preserve reproducibility:

python

def seed_worker(worker_id):
    worker_seed = torch.initial_seed() % 2**32
    numpy.random.seed(worker_seed)
    random.seed(worker_seed)

g = torch.Generator()
g.manual_seed(0)

DataLoader(
    train_dataset,
    batch_size=batch_size,
    num_workers=num_workers,
    worker_init_fn=seed_worker,
    generator=g,
)