This is an automated email from the ASF dual-hosted git repository.
tlopex pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 4f5a17a4ae [BugFix][Relax] Fix ONNX Clip converter for opset 11-12
(#19375)
4f5a17a4ae is described below
commit 4f5a17a4ae4f04061bb7ed4327922bc831bb0d3a
Author: Soowon Jeong <[email protected]>
AuthorDate: Fri Apr 10 03:42:33 2026 +0900
[BugFix][Relax] Fix ONNX Clip converter for opset 11-12 (#19375)
## Description
ONNX changed the Clip operator from attribute-based min/max (opset 1-10)
to input-based min/max (opset 11+). The Relax ONNX frontend only had
`_impl_v1` (attributes) and `_impl_v13` (inputs), so opset 11-12 models
were dispatched to `_impl_v1` which ignores the input-based min/max and
falls back to `-inf`/`inf` defaults, making Clip a no-op.
This caused **silent numerical divergence** in any opset 11-12 model
using Clip/ReLU6 (e.g. MobileNetV2-12 from the ONNX Model Zoo).
### Root cause
`OnnxOpConverter.get_converter()` selects the largest `_impl_v*` version
<= the model opset. With only `v1` and `v13`, opset 11-12 mapped to
`v1`, which reads min/max from attributes — but opset 11+ passes them as
inputs.
### Fix
Add `_impl_v11` that delegates to `_impl_v13`.
### Results (MobileNetV2, opset 12)
| Metric | Before | After |
|--------|:---:|:---:|
| max abs diff vs ORT | 1.72e+06 | **8.58e-06** |
| cosine similarity | 0.222 | **1.000** |
| Top-5 match | No | **Yes** |
## Testing
```bash
pytest tests/python/relax/test_frontend_onnx.py -k "clip" -v
```
All 6 existing Clip tests pass (opset 6 and 13+). The fix only affects
opset 11-12 dispatch.
---
python/tvm/relax/frontend/onnx/onnx_frontend.py | 21 +++++++++++++--------
1 file changed, 13 insertions(+), 8 deletions(-)
diff --git a/python/tvm/relax/frontend/onnx/onnx_frontend.py
b/python/tvm/relax/frontend/onnx/onnx_frontend.py
index 64fbf94076..940022de41 100644
--- a/python/tvm/relax/frontend/onnx/onnx_frontend.py
+++ b/python/tvm/relax/frontend/onnx/onnx_frontend.py
@@ -1087,6 +1087,11 @@ class Clip(OnnxOpConverter):
results = bb.emit_te(topi.minimum, results, max)
return results
+ @classmethod
+ def _impl_v11(cls, bb, inputs, attr, params):
+ # Opset 11 changed Clip from attribute-based min/max to input-based.
+ return cls._impl_v13(bb, inputs, attr, params)
+
@classmethod
def _impl_v13(cls, bb, inputs, attr, params):
results = inputs[0]
@@ -3769,13 +3774,13 @@ def _argreduce_select_last_index(bb, data, axis,
keepdims, op):
offset = relax.const(int(axis_size) - 1, "int64")
else:
# dynamic: get axis size at runtime and subtract 1
- shape_tensor = bb.normalize(relax.op.shape_to_tensor(
- bb.normalize(relax.op.shape_of(data))
- ))
- offset = bb.normalize(relax.op.subtract(
- bb.normalize(relax.op.take(shape_tensor, relax.const(axis,
"int64"), axis=0)),
- relax.const(1, "int64"),
- ))
+ shape_tensor =
bb.normalize(relax.op.shape_to_tensor(bb.normalize(relax.op.shape_of(data))))
+ offset = bb.normalize(
+ relax.op.subtract(
+ bb.normalize(relax.op.take(shape_tensor, relax.const(axis,
"int64"), axis=0)),
+ relax.const(1, "int64"),
+ )
+ )
return relax.op.subtract(offset, flipped_idx)
@@ -4353,7 +4358,7 @@ class SplitToSequence(OnnxOpConverter):
for i in range(n)
]
return relax.Tuple(squeezed)
-
+
return output
