wuyii8941 opened a new issue, #19692:
URL: https://github.com/apache/tvm/issues/19692
## Expected behavior
ONNX `CumSum` has two boolean attributes — `reverse` and `exclusive`. With
`exclusive=1`, the output excludes the current element:
> Y[..., i] = sum(X[..., 0..i-1]) (exclusive=1, reverse=0)
ONNX Runtime supports both attributes; this is the standard PyTorch `cumsum`
semantics shifted by one.
## Actual behavior
The TVM frontend hits a bare assertion that fails the entire `from_onnx`
call:
```
AssertionError: Exclusive option not yet supported.
```
This is not even raised as a structured `NotImplementedError` — it's an
`assert` with no operator/opset context, making downstream error reporting
confusing.
## Reproduction
```python
import numpy as np
import onnx
from onnx import helper, TensorProto, numpy_helper
import onnxruntime as ort
from tvm.relax.frontend.onnx import from_onnx
X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 4])
Y = onnx.ValueInfoProto(); Y.name = "Y"
axis = numpy_helper.from_array(np.array(1, dtype=np.int64), "axis")
node = helper.make_node("CumSum", ["X", "axis"], ["Y"], exclusive=1,
reverse=0)
g = helper.make_graph([node], "g", [X], [Y], initializer=[axis])
m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 14)])
arr = np.array([[1., 2., 3., 4.]], dtype=np.float32)
print("ORT:", ort.InferenceSession(m.SerializeToString()).run(None, {"X":
arr})[0])
# ORT: [[0. 1. 3. 6.]]
inf = onnx.shape_inference.infer_shapes(m)
mod = from_onnx(inf) # AssertionError: Exclusive option not yet supported.
```
## Root cause
`python/tvm/relax/frontend/onnx/onnx_frontend.py`, `CumSum._impl_v14`:
```python
@classmethod
def _impl_v14(cls, bb, inputs, attr, params):
data = inputs[0]
axis_input = get_constant(inputs[1], params)
assert not attr.get("exclusive", False), "Exclusive option not yet
supported."
...
```
`topi.cumsum` (and `relax.op.cumsum`) only implement the inclusive form, but
the converter just asserts instead of (a) raising a structured error or (b)
emitting an `exclusive` decomposition.
## Suggested fix
Two-stage approach:
1. **Short-term**: replace the assertion with a clear `NotImplementedError`
that mentions the op and opset:
```python
if attr.get("exclusive", False):
raise NotImplementedError(
"CumSum with exclusive=1 is not yet supported in the Relax ONNX
frontend"
)
```
2. **Better**: decompose `exclusive` to inclusive `cumsum` followed by a
one-element shift, since this is straightforward:
```python
y = relax.op.cumsum(data, axis=axis)
if exclusive:
# shift right by one along `axis`, fill with identity (0)
y = relax.op.subtract(y, data)
```
`reverse=1` already lacks a dedicated branch and silently falls through to
`relax.op.cumsum(...)` — that should be audited as well.
## Impact
Any model using ONNX `CumSum` with `exclusive=1` cannot be imported. Common
in attention masking utilities and statistical models.
## Environment
- TVM: latest `main` (commit b172d5ea3)
- Python: 3.11
- ONNX Runtime: 1.24.4
--
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]
