siyiweigeHEW opened a new issue, #20229:
URL: https://github.com/apache/tvm/issues/20229
### Expected behavior
`torch.Tensor.expand(*sizes)` allows **prepending new leading dimensions**
when `len(sizes) > input.rank`: the input shape is *right-aligned* under
`sizes`, and `-1` in `sizes` keeps the size of the corresponding input
dimension. Both of these are valid and return the documented shape:
```python
torch.zeros(2, 3).expand(4, -1, -1) # -> (4, 2, 3)
torch.zeros(3).expand(2, -1) # -> (2, 3)
torch.zeros(2, 3).expand(4, -1, 3) # -> (4, 2, 3)
```
### Actual behavior
`tvm.relax.frontend.torch.from_exported_program` crashes at model import for
all of the above. The converter (`_expand` in
`python/tvm/relax/frontend/torch/base_fx_graph_translator.py:1752-1761`)
resolves each `-1` by **left-to-right index** into the input shape:
```python
for idx, i in enumerate(sizes):
if isinstance(i, int) and i == -1:
broadcast_shape.append(in_shape[idx]) # line 1758
...
```
It never right-aligns the input shape under `sizes`. When a `-1` sits at a
position past the input's rank (a newly prepended leading dim), `in_shape[idx]`
is out of range and the whole model import fails:
```
(2,3).expand(4,-1,-1) torch -> (4,2,3) TVM IndexError: ShapeExpr index
out of range
(3,).expand(2,-1) torch -> (2,3) TVM IndexError: ShapeExpr index
out of range
```
When the wrong index is still in range, the `-1` is mapped to the **wrong**
input dimension, producing an invalid broadcast target:
```
(2,3).expand(4,-1,3) torch -> (4,2,3)
# TVM maps -1 -> in_shape[1]=3 instead of in_shape[0]=2 ->
broadcast_shape=[4,3,3]
TVM InternalError: broadcast_to expects the input tensor shape is
broadcastable to the target ...
```
All six tested models are plain, runnable PyTorch modules.
### Environment
- OS: Linux
- TVM: v0.24.dev0 (main branch, commit `262c6d2e0`, built 2026-02-11)
- Python: 3.11
- torch: 2.10.0
### Steps to reproduce
```python
"""Repro: torch.Tensor.expand -1 / right-alignment mishandled by TVM relax
torch frontend."""
import torch
import torch.nn as nn
from tvm.relax.frontend.torch import from_exported_program
class M(nn.Module):
def forward(self, x):
return x.expand(4, -1, -1) # sizes longer than input rank
x = torch.randn(2, 3)
print("torch:", tuple(M()(x).shape)) # (4, 2, 3)
exp = torch.export.export(M().eval().cpu(), (x.cpu(),))
print(exp.graph) # %expand =
aten.expand.default(%x, [4, -1, -1])
from_exported_program(exp) # TVM raises IndexError
```
Actual output:
```
torch: (4, 2, 3)
graph():
%x : [num_users=1] = placeholder[target=x]
%expand : [num_users=1] =
call_function[target=torch.ops.aten.expand.default](args = (%x, [4, -1, -1]),
kwargs = {})
return (expand,)
Error converting operator expand, with inputs: [data,
metadata["relax.expr.Constant"][0]]
TVM: IndexError: ShapeExpr index out of range
File "tvm/relax/frontend/torch/base_fx_graph_translator.py", line 1758, in
_expand
broadcast_shape.append(in_shape[idx])
```
More cases (all pass in torch, all fail in TVM):
```python
torch.zeros(3).expand(2, -1) # torch (2,3) TVM IndexError
torch.zeros(2, 3, 4).expand(5, -1, -1, -1) # torch (5,2,3,4) TVM IndexError
torch.zeros(1, 3).expand(4, -1, -1) # torch (4,1,3) TVM IndexError
torch.zeros(2, 3).expand(4, -1, 3) # torch (4,2,3) TVM
InternalError (wrong -1 mapping)
torch.zeros(2, 1).expand(4, -1, -1) # torch (4,2,1) TVM IndexError
```
For comparison, the common cases where `-1` stays within the input rank (no
new leading dims) work correctly (`max|diff|=0` vs torch):
`(1,3).expand(4,-1)`, `(2,3,4).expand(2,-1,4)`, `(2,3).expand(2,3)`.
### Triage
* needs-triage
* bug
* relax
* frontend/torch
--
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]
