tinywisdom opened a new issue, #19553:
URL: https://github.com/apache/tvm/issues/19553
### Summary
A minimal PyTorch model using `torch.logical_not` on a float tensor can be
exported to Relax, but fails during `tvm.compile(..., target="llvm")`.
PyTorch supports `torch.logical_not` on non-bool tensors. For example, a
float input produces a bool output:
```python
x = torch.tensor([[0.0, 1.0]], dtype=torch.float32)
torch.logical_not(x)
# tensor([[ True, False]])
```
However, after `torch.export` and `from_exported_program`, TVM emits a Relax
program where `R.logical_not` is applied directly to a `float32` tensor, and
the output is also annotated as `float32`.
This later fails in `relax.transform.LegalizeOps`, because TOPI/TIR
logical-not expects a boolean argument.
### Actual behavior
`tvm.compile` fails during Relax legalization:
```
tvm.error.InternalError: Check failed: (arg.dtype().is_bool()) is false:
Expected boolean argument for ! operator (logical NOT), but received x[ax0,
ax1] of type float32
```
The relevant stack trace reaches:
```
relax.transform.LegalizeOps
-> tvm.topi.logical_not
-> tvm::logical_not
-> type_check_boolean_args
-> Expected boolean argument for ! operator (logical NOT)
```
### Environment
TVM: 0.23.0
LLVM: 17.0.6
Python: 3.10.16 (from stack paths)
NumPy: 2.2.6
### Steps to reproduce
```python
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import sys
import platform
import traceback
import torch
import tvm
class MyModel(torch.nn.Module):
def forward(self, x):
return torch.logical_not(x)
def main():
print("=" * 80)
print("Environment")
print("=" * 80)
print("python:", sys.version.replace("\n", " "))
print("platform:", platform.platform())
print("torch:", torch.__version__)
print("tvm:", getattr(tvm, "__version__", "<unknown>"))
print("tvm path:", getattr(tvm, "__file__", "<unknown>"))
model = MyModel().eval()
x = torch.tensor([[0.0, 1.0]], dtype=torch.float32)
with torch.no_grad():
eager = model(x)
print("=" * 80)
print("PyTorch eager")
print("=" * 80)
print("input:", x)
print("input dtype:", x.dtype, "shape:", tuple(x.shape))
print("eager output:", eager)
print("eager dtype:", eager.dtype, "shape:", tuple(eager.shape))
ep = torch.export.export(model, (x,))
from tvm.relax.frontend.torch import from_exported_program
ir_mod = from_exported_program(ep)
print("=" * 80)
print("Exported Relax IR")
print("=" * 80)
print(ir_mod.script(show_meta=True))
print("=" * 80)
print("tvm.compile with LLVM")
print("=" * 80)
ex = tvm.compile(
ir_mod,
target=tvm.target.Target("llvm"),
relax_pipeline="default",
tir_pipeline="default",
)
print("compile: OK")
print(ex)
if __name__ == "__main__":
try:
main()
except Exception:
print("compile: FAILED")
traceback.print_exc()
```
### 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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
