wuyii8941 opened a new issue, #19575:
URL: https://github.com/apache/tvm/issues/19575
## Description
The `ReorderPermuteDimsAfterConcat` pass crashes with `IndexError: Index -1
out of bounds 4` when the concat operation uses a negative axis value (e.g.,
`axis=-1`). The pass reads the axis value directly without normalizing negative
indices.
Using a positive axis value (e.g., `axis=3` instead of `axis=-1`) works
correctly.
## Reproducer
```python
import tvm
from tvm import relax
import tvm.relax.op as R
bb = relax.BlockBuilder()
x = relax.Var('x', relax.TensorStructInfo((1, 4, 8, 8), 'float32'))
y = relax.Var('y', relax.TensorStructInfo((1, 4, 8, 8), 'float32'))
with bb.function('main', [x, y]):
with bb.dataflow():
xt = bb.emit(R.permute_dims(x, [0, 2, 3, 1]))
yt = bb.emit(R.permute_dims(y, [0, 2, 3, 1]))
out = bb.emit_output(R.concat([xt, yt], axis=-1)) # negative axis
bb.emit_func_output(out)
mod = bb.finalize()
# This crashes:
pipeline = tvm.ir.transform.Sequential([
relax.transform.ReorderPermuteDimsAfterConcat(),
relax.transform.LegalizeOps()
])
mod_l = pipeline(mod) # IndexError: Index -1 out of bounds 4
```
## Error
```
IndexError: Index -1 out of bounds 4
```
at `tvm/3rdparty/tvm-ffi/include/tvm/ffi/container/seq_base.h` line 86.
## Expected behavior
The pass should normalize negative axis values (e.g., `axis=-1` →
`axis=ndim-1`) before using them as array indices.
## Workaround
Use positive axis values in concat operations.
## Environment
- TVM version: 0.24.dev0 (commit 0b0afd8dd, 2026-04-24)
--
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]
