| Title | pytorch pytorch (in torch.cuda.memory.caching_allocator_delete) torch 2.6.0 invalid memory access |
|---|
| Description | When attempting to manually delete a tensor's memory using `torch.cuda.memory.caching_allocator_delete` after allocating memory on the GPU with `torch.empty`, I encounter an `Aborted` error.
To reproduce:
```
import torch
from torch.cuda.memory import caching_allocator_delete
torch.cuda.empty_cache()
dev_props = torch.cuda.get_device_properties(0)
total_memory = dev_props.total_memory
allocation = int(total_memory * 0.5)
tmp_tensor = torch.empty(allocation, dtype=torch.int8, device='cuda')
mem_ptr = tmp_tensor.data_ptr()
caching_allocator_delete(mem_ptr)
``` |
|---|
| Source | ⚠️ https://github.com/pytorch/pytorch/issues/149821 |
|---|
| User | Default436352 (UID 81891) |
|---|
| Submission | 03/23/2025 02:26 (1 Year ago) |
|---|
| Moderation | 04/02/2025 22:47 (11 days later) |
|---|
| Status | Accepted |
|---|
| VulDB entry | 303041 [PyTorch 2.6.0 CUDACachingAllocator.cpp torch.cuda.memory.caching_allocator_delete memory corruption] |
|---|
| Points | 20 |
|---|