Updates Onnx Version to 1.16

build_tools/python_deploy/build_linux_packages.sh

@@ -46,7 +46,7 @@ TM_RELEASE_DOCKER_IMAGE="${TM_RELEASE_DOCKER_IMAGE:-quay.io/pypa/manylinux2014_$
 # ./build_tools/docker/Dockerfile
 TM_CI_DOCKER_IMAGE="${TM_CI_DOCKER_IMAGE:-powderluv/torch-mlir-ci:latest}"
 # Version of Python to use in Release builds. Ignored in CIs.
-TM_PYTHON_VERSIONS="${TM_PYTHON_VERSIONS:-cp38-cp38 cp310-cp310 cp311-cp311}"
+TM_PYTHON_VERSIONS="${TM_PYTHON_VERSIONS:-cp38-cp38 cp310-cp310 cp311-cp311 cp312-cp312}"
 # Location to store Release wheels
 TM_OUTPUT_DIR="${TM_OUTPUT_DIR:-${this_dir}/wheelhouse}"
 # What "packages to build"

projects/pt1/e2e_testing/xfail_sets.py

@@ -2510,8 +2510,6 @@
     "SplitDimStaticModule_basic",
     "SqrtIntConstantModule_basic",
     "SqrtIntModule_basic",
-    "StdCorrectionEmptyDimModule_basic",
-    "StdDimEmptyDimModule_basic",
     "SubFloatModule_basic",
     "SubIntModule_basic",
     "TanhBackward_basic",
@@ -2565,8 +2563,6 @@
     "UpSampleNearest2dDynamicFactor_basic",
     "UpSampleNearest2dStaticFactor_basic",
     "UpSampleNearest2d_basic",
-    "VarCorrectionEmptyDimModule_basic",
-    "VarDimEmptyDimModule_basic",
     "ViewCollapseDynamicWithAtenSizeIntModule_basic",
     "ViewCollapseModule_basic",
     "ViewDynamicExpandCollapseModule_basic",
@@ -2683,17 +2679,23 @@
 
 
 ONNX_CRASHING_SET = {
+    # unknown
+    "ElementwisePreluModule_basic",
     "FakeQuantizePerTensorAffineModule_basic",
     "FakeQuantizePerTensorAffineDynamicShapeModule_basic",
-    "ElementwisePreluModule_basic",
     "ViewDynamicExpandCollapseWithParallelUnknownDimModule_basic",
+    # The following test sporadically stopped producing correct numerics for the golden value in the CI.
+    # For now, we are removing the test until this issue has been debugged.
+    "QuantizedMLP_basic",
     "ScatterReduceFloatProdModuleIncludeSelf",
     "ScatterReduceFloatSumModuleIncludeSelf",
     "ScatterReduceIntProdModuleIncludeSelf",
     "ScatterReduceIntSumModuleIncludeSelf",
-    # The following test sporadically stopped producing correct numerics for the golden value in the CI.
-    # For now, we are removing the test until this issue has been debugged.
-    "QuantizedMLP_basic",
+    # these empty dim modules generate IR with an empty float constant, which gets compared to int dims
+    "StdDimEmptyDimModule_basic",
+    "StdCorrectionEmptyDimModule_basic",
+    "VarCorrectionEmptyDimModule_basic",
+    "VarDimEmptyDimModule_basic",
     # Runtime crash: mismatched size for broadcast
     "MaxPool2dWithIndicesAllNegativeValuesModule_basic",
     "MaxPool2dWithIndicesNonDefaultPaddingModule_basic",

python/torch_mlir/extras/onnx_importer.py

@@ -652,6 +652,8 @@ def tensor_proto_to_attr(self, tp: onnx.TensorProto) -> Attribute:
     onnx.TensorProto.DataType.FLOAT8E5M2: lambda: Float8E5M2Type.get(),
     onnx.TensorProto.DataType.FLOAT8E5M2FNUZ: lambda: Float8E5M2FNUZType.get(),
     onnx.TensorProto.DataType.STRING: lambda: "!torch.str",
+    onnx.TensorProto.DataType.UINT4: lambda: IntegerType.get_unsigned(4),
+    onnx.TensorProto.DataType.INT4: lambda: IntegerType.get_signed(4),
     # Ommitted: STRING,
 }
 
@@ -688,6 +690,9 @@ def tensor_proto_to_attr(self, tp: onnx.TensorProto) -> Attribute:
         ),
         signless=False,
     ),
+    onnx.TensorProto.DataType.UINT8: lambda tp: DenseElementsAttr.get(
+        np.asarray(tp.int32_data, dtype=np.uint8).reshape(tp.dims), signless=False
+    ),
     onnx.TensorProto.DataType.INT8: lambda tp: DenseElementsAttr.get(
         np.asarray(tp.int32_data, dtype=np.int8).reshape(tp.dims), signless=False
     ),

python/torch_mlir/tools/import_onnx/__main__.py

@@ -79,7 +79,7 @@ def load_onnx_model(args: argparse.Namespace) -> onnx.ModelProto:
         raw_model = onnx.load(args.input_file)
     else:
         raw_model = onnx.load(args.input_file, load_external_data=False)
-        onnx.load_external_data_for_model(raw_model, args.data_dir)
+        onnx.load_external_data_for_model(raw_model, str(args.data_dir))
 
     # Run the checker to test whether the file is above the threshold for
     # in-memory shape inference.  If not, go ahead and do the shape inference.

test-requirements.txt

@@ -1,5 +1,5 @@
 pillow
 dill
 multiprocess
-onnx==1.15.0
+onnx==1.16.0
 mpmath==1.3.0

test/python/onnx_importer/import_smoke_test.py

@@ -79,7 +79,9 @@ def tearDownClass(cls):
     def load_onnx_model(self, file_path: Path) -> onnx.ModelProto:
         raw_model = onnx.load(file_path)
         try:
-            inferred_model = onnx.shape_inference.infer_shapes(raw_model)
+            inferred_model = onnx.shape_inference.infer_shapes(
+                raw_model, data_prop=True
+            )
         except onnx.onnx_cpp2py_export.shape_inference.InferenceError as e:
             print("WARNING: Shape inference failure (skipping test):", e)
             self.skipTest(reason="shape inference failure")