wuyii8941 opened a new issue, #19571:
URL: https://github.com/apache/tvm/issues/19571
## Expected behavior
`from_onnx()` should successfully convert ONNX models containing `TopK`
operators.
## Actual behavior
Any ONNX model with a `TopK` operator crashes with:
```
TypeError: only 0-dimensional arrays can be converted to Python scalars
```
This is 100% reproducible — no TopK model can be converted.
## Reproduction
```python
import numpy as np
import onnx
from onnx import helper, TensorProto, numpy_helper
import tvm
from tvm.relax.frontend.onnx import from_onnx
X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [5, 10])
K_init = numpy_helper.from_array(np.array([3], dtype=np.int64), "K")
Y_val = helper.make_tensor_value_info("Values", TensorProto.FLOAT, None)
Y_idx = helper.make_tensor_value_info("Indices", TensorProto.INT64, None)
node = helper.make_node("TopK", ["X", "K"], ["Values", "Indices"], axis=1)
graph = helper.make_graph([node], "test", [X], [Y_val, Y_idx],
initializer=[K_init])
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 18)])
model = onnx.shape_inference.infer_shapes(model)
mod = from_onnx(model) # TypeError: only 0-dimensional arrays can be
converted to Python scalars
```
## Root cause
The TopK converter receives `K` as a 1-element tensor constant (shape
`[1]`), but tries to extract it as a Python scalar directly. The 1-D tensor
needs `.item()` or indexing `[0]` to extract the scalar value.
Tested on 4 different configurations — all crash:
- `shape=(5, 10), k=3, axis=1` → CRASH
- `shape=(5, 10), k=3, axis=0` → CRASH
- `shape=(3, 4, 5), k=2, axis=2` → CRASH
- `shape=(100,), k=5, axis=0` → CRASH
## Environment
- TVM: 0.24.dev0, commit 0b0afd8dd (2026-04-24)
- Python: 3.11
- OS: Linux
--
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]
