tinywisdom opened a new issue, #18355:
URL: https://github.com/apache/tvm/issues/18355
### Expected behavior
`from_exported_program` followed by relax.build segfaults when the exported
PyTorch program performs a batch-wise advanced indexing assignment of the form
`M[:, rows, cols] = x` (two integer index tensors on the last two dims + a
leading batch slice). The crash happens in `tvm::relax::Tuple` construction
inside FFI. Plain reads (no assignment) or functional rewrites do not crash.
### Actual behavior
```
!!!!!!! TVM FFI encountered a Segfault !!!!!!!
...
tvm::relax::Tuple::Tuple(tvm::ffi::Array<tvm::RelaxExpr, void>, tvm::Span)
[clone .cold]
...
Segmentation fault (core dumped)
```
### Environment
- OS: (Ubuntu 22.04.4 LTS (x86_64))
- TVM version: (release v0.21.0)
- Python: (3.10.16)
- LLVM: (17.0.6)
### Steps to reproduce
```python
import torch
import torch.nn as nn
from torch.export import export as torch_export
from tvm.relax.frontend.torch import from_exported_program
import tvm
from tvm import relax
class IndexAssign(nn.Module):
def forward(self, x): # x: [B, 10]
B = x.size(0)
M = torch.zeros(B, 11, 11) # could also set device/dtype to
match x
rows = torch.arange(10)
cols = rows + 1
M[:, rows, cols] = x # batch-wise paired index write
return M
def main():
m = IndexAssign().eval()
inp = torch.randn(2, 10)
ep = torch_export(m, (inp,)) # export succeeds
mod = from_exported_program(ep) # import OK or still OK
target = tvm.target.Target("llvm")
relax.build(mod, target=target) # <-- Segfault here (often)
print("Built successfully (unexpected).")
if __name__ == "__main__":
main()
```
### Triage
* needs-triage
* bug
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
