bad_alloc when using Conv2d after torchvision.io.read_video

🐛 Describe the bug

Short:

Attempting a forward pass on Conv2d after loading an unrelated video with torchvision.io.read_video throws an std::bad_alloc runtime exception

Code:

import torch
import torchvision

if __name__ == '__main__':
    print(f"Pytorch Version: {torch.__version__}")
    print(f"Torchvision Version: {torchvision.__version__}")

    mp4 = torchvision.io.read_video("testpov_video.mp4", pts_unit='sec')
    print(f"Video shape: {mp4[0].shape}\tdevice: {mp4[0].device}\tdtype:{mp4[0].dtype}")

    print(f"Cuda available: {torch.cuda.is_available()}")
    print(f"Cuda device: {torch.cuda.get_device_name()}")

    test_conv = torch.nn.Conv2d(3, 8, kernel_size=1).to("cuda")
    test_in = torch.rand((1, 3, 32, 32)).to("cuda")
    test_conv(test_in)

Output

Pytorch Version: 2.3.0+cu121
Torchvision Version: 0.18.0+cu121
Video shape: torch.Size([240, 256, 256, 3])	device: cpu	dtype:torch.uint8
Cuda available: True
Cuda device: NVIDIA RTX A4000
terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc

Process finished with exit code 134 (interrupted by signal 6: SIGABRT)

Long/Additional information:

• My system is Ubuntu 22.04.4 in a virtual machine with ~96GB ram allocated. There are 2 gpus, an A4000 and an A5000 and this error occurs when attempting to use either of them.
• The error does not occur when using the CPU
• It does not occur without loading the mp4 video
• I have tested with other mp4 files and the error still occurs
• I attempted pip install --force-reinstall torch torchvision torchaudio and the problem persisted
• I have not tested with other torchvision.io functions

Versions

Collecting environment information...
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-35-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.5.40
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA RTX A5000
GPU 1: NVIDIA RTX A4000

Nvidia driver version: 535.171.04
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 40 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 20
On-line CPU(s) list: 0-19
Vendor ID: AuthenticAMD
Model name: Common KVM processor
CPU family: 15
Model: 6
Thread(s) per core: 1
Core(s) per socket: 20
Socket(s): 1
Stepping: 1
BogoMIPS: 5599.99
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx lm rep_good nopl cpuid extd_apicid tsc_known_freq pni cx16 x2apic hypervisor cmp_legacy 3dnowprefetch vmmcall
Hypervisor vendor: KVM
Virtualisation type: full
L1d cache: 1.3 MiB (20 instances)
L1i cache: 1.3 MiB (20 instances)
L2 cache: 10 MiB (20 instances)
L3 cache: 320 MiB (20 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-19
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Not affected
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] pytorch-fast-transformers==0.4.0
[pip3] pytorch-lightning==2.2.4
[pip3] torch==2.3.0
[pip3] torchaudio==2.3.0
[pip3] torchdata==0.7.1
[pip3] torchinfo==1.8.0
[pip3] torchmetrics==1.3.2
[pip3] torchvision==0.18.0
[pip3] triton==2.3.0
[conda] No relevant packages

Start by running the supplied Python reproducer and compare GPU Conv2d behavior with and without the preceding torchvision.io.read_video call. Trace the interaction between video loading and the first CUDA convolution; done means the reproducer no longer terminates with std::bad_alloc on the reported GPU setup, with the existing CPU behavior preserved.