[5680954][ONNX][Autocast] Fix 0-dim scalar constant issue (#691)

## What does this PR do?

**Type of change:** Bug fix

**Overview:** Ops with 0-dim scalar constants were being forced to have
shape of 1 instead of 0. This PR fixes that issue.

## Usage

```python
$ python -m modelopt.onnx.autocast --onnx_path=$MODEL_NAME.onnx
```

## Testing
Added unittest.

## Before your PR is "*Ready for review*"
<!-- If you haven't finished some of the above items you can still open
`Draft` PR. -->

- **Make sure you read and follow [Contributor
guidelines](https://github.com/NVIDIA/Model-Optimizer/blob/main/CONTRIBUTING.md)**
and your commits are signed.
- **Is this change backward compatible?**: Yes
- **Did you write any new necessary tests?**: Yes
- **Did you add or update any necessary documentation?**: No
- **Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?**:
No

Signed-off-by: gcunhase <4861122+gcunhase@users.noreply.github.com>

Author: gcunhase

modelopt/onnx/autocast/precisionconverter.py

@@ -1336,10 +1336,6 @@ def _convert_constant_values(self, const_node, cast_node: onnx.NodeProto) -> Non
         else:
             casted_data = original_data.astype(cast_dtype)
 
-        # Workaround for 0-dimensional tensors (scalars)
-        if casted_data.ndim == 0:
-            casted_data = casted_data.reshape(1)
-
         # Create a new constant node with casted data
         if cast_to_type == onnx.TensorProto.BFLOAT16:
             # Create TensorProto manually for bfloat16

tests/_test_utils/onnx/lib_test_models.py

@@ -75,7 +75,7 @@ def forward(self, x):
 class SimpleMLP(nn.Module):
     """Simple toy model."""
 
-    def __init__(self, fi=16, f1=18, f2=20, fo=22):
+    def __init__(self, fi=16, f1=18, f2=20, fo=22, bias_add=False):
         super().__init__()
         self.net = nn.Sequential(
             nn.Linear(fi, f1, bias=False),
@@ -84,10 +84,13 @@ def __init__(self, fi=16, f1=18, f2=20, fo=22):
             nn.ReLU(),
             nn.Linear(f2, fo, bias=False),
         )
+        self.bias_add = bias_add
 
     def forward(self, x):
         for mod in self.net:
             x = mod(x)
+        if self.bias_add:
+            x += 1e-4
         return x
 
 

tests/unit/onnx/test_autocast_quantize.py

@@ -36,20 +36,21 @@ def assert_nodes_are_quantized(nodes):
 
 
 @pytest.mark.parametrize("keep_io_types", [True, False])
-def test_autocast_quantize_int8(tmp_path, keep_io_types):
-    model_torch = SimpleMLP()
+@pytest.mark.parametrize("bias_add", [True, False])
+def test_autocast_quantize_int8(tmp_path, keep_io_types, bias_add):
+    model_torch = SimpleMLP(bias_add=bias_add)
     input_tensor = torch.randn(2, 16, 16)
     low_precision_type = "fp16"
 
-    onnx_path = os.path.join(tmp_path, "model.onnx")
+    onnx_path = os.path.join(tmp_path, f"model{'_biasAdd' if bias_add else ''}.onnx")
     export_as_onnx(model_torch, input_tensor, onnx_filename=onnx_path)
 
     # Convert model to low precision
     converted_model = convert_to_mixed_precision(
         onnx_path, keep_io_types=keep_io_types, low_precision_type=low_precision_type
     )
     converted_model_path = onnx_path.replace(
-        ".onnx", f".{low_precision_type}.{'keepIOTypes' if keep_io_types else ''}.onnx"
+        ".onnx", f".{low_precision_type}{'_keepIOTypes' if keep_io_types else ''}.onnx"
     )
     onnx.save(converted_model, converted_model_path)