[BUG]: Memory leak in fp8 kernel launch

[ZhangZhiPku]

Version

1.0.0

Version

13.1

Which installation method(s) does this occur on?

Pip

Describe the bug.

It seems that launching the FP8 kernel may cause PyTorch memory leaks, where the memory occupied by the input tensor cannot be released properly.

No similar issues have been observed when the input tensor is of other data types.

Test Environment:
GPU: RTX 5090
CUDA Version: 13.1
CuTILE Version: 1.0.0

Minimum reproducible example

try following script:

import cuda.tile as ct
import torch

@ct.kernel
def kernel(A: ct.Array):
    pass

for i in range(100000):
    x = torch.zeros(size=[4025, 5394], device="cuda", dtype=torch.float8_e4m3fn)
    ct.launch(torch.cuda.current_stream(), (1, 1, 1), kernel, (x, ))

Relevant log output

Traceback (most recent call last):
  File "/root/autodl-tmp/super.py", line 9, in <module>
    x = torch.zeros(size=[4025, 5394], device="cuda", dtype=torch.float8_e4m3fn)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 22.00 MiB. GPU 0 has a total capacity of 31.36 GiB of which 17.06 MiB is free. Including non-PyTorch memory, this process has 31.33 GiB memory in use. Of the allocated memory 28.96 GiB is allocated by PyTorch, and 1.81 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

Full env printout

Other/Misc.

No response

Contributing Guidelines

• I agree to follow cuTile Python's contributing guidelines
• I have searched the open bugs and have found no duplicates for this bug report

[haijieg]

@ZhangZhiPku thanks for raising this issue. This turned out to be a bug in PyTorch when they convert tensor to dlpack.

This code will trigger the same OOM error.

import torch

for i in range(100000):
    x = torch.zeros(size=[4025, 5394], device="cuda", dtype=torch.float8_e4m3fn)
    try:
        x.__dlpack__()
    except:
        pass

Because PyTorch doesn't support fp8 tensor conversion through DLPack, in cuTile we take the dlpack path and fallback with a slower conversion path here.

I filed a PyTorch issue: pytorch/pytorch#171820
We will add a workaround to avoid triggering dlpack conversion exception from PyTorch.