Reproducibility

TensorPlay random number generator

You can use tensorplay.manual_seed() to seed the RNG for all devices (both CPU and CUDA):

import tensorplay
tensorplay.manual_seed(0)

Some TensorPlay operations may use random numbers internally. tensorplay.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 tensorplay.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 tensorplay.manual_seed() to the same value between subsequent calls.

Python

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

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:

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 tensorplay.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 tensorplay.backends.cudnn.benchmark = True.

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

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 tensorplay.use_deterministic_algorithms(True) or tensorplay.backends.cudnn.deterministic = True is set. The latter setting controls only this behavior, unlike tensorplay.use_deterministic_algorithms() which will make other TensorPlay operations behave deterministically, too.

CUDA Scaled Dot Product Attention

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

When tensorplay.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 tensorplay.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 tensorplay.nn.attention.sdpa_kernel() context manager to restrict which backends SDPA may use. For example, to guarantee deterministic behavior regardless of hardware:

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

tensorplay.use_deterministic_algorithms(True)

# Option 1: Use only the Math backend (always deterministic)
with sdpa_kernel(SDPBackend.MATH):
    out = tensorplay.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 = tensorplay.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 tensorplay.nn.RNN() and tensorplay.nn.LSTM() for details and workarounds.

Filling uninitialized memory

Operations like tensorplay.empty() and tensorplay.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, tensorplay.utils.deterministic.fill_uninitialized_memory is set to True by default, which will fill the uninitialized memory with a known value if tensorplay.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.