[5763448] Fix Resize input type mismatch error

modelopt/onnx/autocast/precisionconverter.py

@@ -68,7 +68,7 @@ class InitializerConsumerTracker:
 OP_TYPES_NOT_SUPPORTED_IN_LOW_PRECISION = ["Upsample", "NonMaxSuppression", "Celu"]
 
 # Mapping of op types to indices of inputs that should not be converted to low precision.
-SKIP_LOW_PRECISION_MAPPING_FP16 = {"Resize": {2}}
+SKIP_LOW_PRECISION_MAPPING_FP16 = {"Resize": {1, 2}}  # {X, roi (opt), scales (opt)}
 SKIP_LOW_PRECISION_MAPPING_BF16 = {"Resize": {1, 2}}
 
 

tests/_test_utils/onnx/lib_test_models.py

@@ -924,3 +924,86 @@ def build_conv_isinf_model(opset_version=13):
     onnx.checker.check_model(model_inferred)
 
     return model_inferred
+
+
+def build_conv_resize_model():
+    # Define your model inputs and outputs
+    input_names = ["input_0"]
+    output_names = ["output_0"]
+    input_shapes = [(1, 288, 32, 32)]
+    output_shapes = [(1, 16, 64, 64)]
+
+    inputs = [
+        helper.make_tensor_value_info(input_name, onnx.TensorProto.FLOAT, input_shape)
+        for input_name, input_shape in zip(input_names, input_shapes)
+    ]
+    outputs = [
+        helper.make_tensor_value_info(output_name, onnx.TensorProto.FLOAT, output_shape)
+        for output_name, output_shape in zip(output_names, output_shapes)
+    ]
+
+    # Create the ONNX graph with the nodes
+    nodes = [
+        helper.make_node(
+            op_type="Conv",
+            inputs=["input_0", "weights_1"],
+            outputs=["conv1_conv/Conv2D:0"],
+            name="conv1_conv/Conv2D",
+            dilations=[1, 1],
+            group=1,
+            kernel_shape=[1, 1],
+            pads=[0, 0, 0, 0],
+            strides=[1, 1],
+        ),
+        helper.make_node(
+            op_type="Resize",
+            inputs=[
+                "conv1_conv/Conv2D:0",
+                "resize_roi_scales",
+                "resize_roi_scales",
+                "resize_sizes",
+            ],
+            outputs=["output_0"],
+            name="resize1_resize/Resize",
+            coordinate_transformation_mode="asymmetric",
+            cubic_coeff_a=-0.75,
+            mode="nearest",
+            nearest_mode="floor",
+        ),
+    ]
+
+    # Create the ONNX initializers
+    initializers = [
+        helper.make_tensor(
+            name="weights_1",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(16, 288, 1, 1),
+            vals=np.random.uniform(low=0.5, high=1.0, size=16 * 288 * 1 * 1),
+        ),
+        helper.make_tensor(
+            name="resize_roi_scales",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(0,),
+            vals=[],
+        ),
+        helper.make_tensor(
+            name="resize_sizes",
+            data_type=onnx.TensorProto.INT64,
+            dims=(4,),
+            vals=[1, 16, 64, 64],
+        ),
+    ]
+
+    # Create the ONNX graph with the nodes and initializers
+    graph = helper.make_graph(nodes, "conv_resize", inputs, outputs, initializer=initializers)
+
+    # Create the ONNX model
+    model = helper.make_model(graph)
+    model.opset_import[0].version = 13
+    model.ir_version = 10
+
+    # Check the ONNX model
+    model_inferred = onnx.shape_inference.infer_shapes(model)
+    onnx.checker.check_model(model_inferred)
+
+    return model_inferred

tests/unit/onnx/autocast/test_autocast.py

@@ -20,7 +20,7 @@
 import onnx
 import onnx_graphsurgeon as gs
 import pytest
-from _test_utils.onnx.lib_test_models import build_conv_isinf_model
+from _test_utils.onnx.lib_test_models import build_conv_isinf_model, build_conv_resize_model
 
 import modelopt.onnx.autocast.utils as utils
 import modelopt.onnx.utils as onnx_utils
@@ -190,6 +190,26 @@ def test_conv_isinf_conversion(tmp_path, opset_version):
     assert assert_input_precision(isinf_nodes, dtype=supported_dtype)
 
 
+def test_conv_resize_conversion(tmp_path):
+    onnx_model = build_conv_resize_model()
+    onnx_path = os.path.join(tmp_path, "conv_resize_model.onnx")
+    onnx.save(onnx_model, onnx_path)
+
+    # Convert the model
+    converted_model = convert_to_mixed_precision(onnx_path=onnx_path)
+
+    # Output model should be produced in the same tmp_path
+    output_onnx_path = onnx_path.replace(".onnx", ".fp16.onnx")
+    onnx.save(converted_model, output_onnx_path)
+
+    # Load the output model and check QDQ node placements
+    graph = gs.import_onnx(converted_model)
+
+    # Check that Conv is converted
+    conv_nodes = [n for n in graph.nodes if "Conv" in n.op]
+    assert assert_input_precision(conv_nodes)
+
+
 @pytest.mark.parametrize("target_opset", [13, 17, 19, 21])
 def test_opset_parameter(temp_model_path, target_opset):
     """Test that the opset parameter correctly sets the output model's opset version."""