TensorFloat-32(TF32) on Ampere devices¶

Starting in PyTorch 1.7, there is a new flag called allow_tf32 which defaults to true. This flag controls whether PyTorch is allowed to use the TensorFloat32 (TF32) tensor cores, available on new 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 truncating 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 True.
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

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 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.
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, relative error compared to double precision is approximately 2 orders of magnitude larger. If the full FP32 precision is needed, users can disable TF32 by:

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

For more information about TF32, see:

• TensorFloat-32

• CUDA 11

• Ampere architecture

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 = 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)

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.

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.

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.

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()) 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.

Best practices¶

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')

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

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

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

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]])

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)