[AutoBump] Merge with 3f46348e (Sep 18) (54)

lib/Conversion/TorchOnnxToTorch/DefaultDomainAtoF.cpp

@@ -2802,14 +2802,14 @@ void mlir::torch::onnx_c::populateDefaultDomainAtoF(
         // Get fill_value if it is present.
         // Assumption : resultDType and value attr type match.
         auto attr = binder.op->getAttr("torch.onnx.value");
-        auto resultDType = resultType.getDtype();
 
         // Extract the fill value and dtype
         // ONNX requires value attr to be a tensor
+        Value splatvalue;
+        // if no value attr is provided, default is 0.0 float value
         if (!attr) {
-          attr =
-              DenseElementsAttr::get(resultType.toBuiltinTensor(),
-                                     rewriter.getFloatAttr(resultDType, 0.0));
+          splatvalue = rewriter.create<Torch::ConstantFloatOp>(
+              binder.getLoc(), rewriter.getF64FloatAttr(0.0));
         }
 
         // If its a dense resource attr we need to convert to a dense type:
@@ -2830,19 +2830,18 @@ void mlir::torch::onnx_c::populateDefaultDomainAtoF(
         }
 
         Attribute splattr;
-        if (isa<SplatElementsAttr>(attr)) {
+        if (attr && isa<SplatElementsAttr>(attr)) {
           auto denseAttr = cast<DenseElementsAttr>(attr);
           splattr = denseAttr.getSplatValue<Attribute>();
         }
 
-        if (!isa<FloatAttr, IntegerAttr>(splattr)) {
+        if (splattr && !isa<FloatAttr, IntegerAttr>(splattr)) {
           return rewriter.notifyMatchFailure(
               binder.op,
               "`value` attr tensor only supports types int and float for now.");
         }
 
-        Value splatvalue;
-        if (auto intattr = dyn_cast<IntegerAttr>(splattr)) {
+        if (auto intattr = dyn_cast_or_null<IntegerAttr>(splattr)) {
           IntegerType intty = cast<IntegerType>(intattr.getType());
           int64_t value;
           if (intty.isUnsignedInteger()) {
@@ -2856,7 +2855,7 @@ void mlir::torch::onnx_c::populateDefaultDomainAtoF(
               rewriter.create<Torch::ConstantIntOp>(binder.getLoc(), value);
         }
 
-        if (auto fpattr = dyn_cast<FloatAttr>(splattr))
+        if (auto fpattr = dyn_cast_or_null<FloatAttr>(splattr))
           splatvalue = rewriter.create<Torch::ConstantFloatOp>(
               binder.getLoc(),
               rewriter.getF64FloatAttr(fpattr.getValueAsDouble()));

projects/pt1/e2e_testing/xfail_sets.py

@@ -536,12 +536,26 @@
     "TensorToBool_basic",
     "TensorToFloatZeroRank_basic",
     "TensorToFloat_basic",
+    "TensorsSplitTensorLastSmallerModule_basic",
+    "TensorsSplitTensorModule_basic",
+    "TensorsSplitTensorNegativeDimModule_basic",
     "ThresholdBackward2dMixedModule_basic",
     "UnsafeViewCollapseDynamicWithAtenSizeIntModule_basic",
     "UpSampleNearest2dDynamicFactor_basic",
     "ViewCollapseDynamicWithAtenSizeIntModule_basic",
     "ViewSizeFromOtherTensor_basic",
     "WeightNormInterfaceModule_basic",
+    # Error: `aten.as_strided` op is not supported
+    "ChunkListUnpackDynamic_Module_basic",
+    "ChunkListUnpackUnevenDynamic_Module_basic",
+    "ChunkListUnpackUneven_Module_basic",
+    "ChunkListUnpack_Module_basic",
+    "SplitTensorGetItem_Module_basic",
+    "SplitTensorLastSmallerModule_basic",
+    "SplitTensorListUnpackModule_basic",
+    "SplitTensorNegativeDimModule_basic",
+    "SplitWithSizesListUnpackModule_basic",
+    "SplitWithSizes_Module_basic",
 }
 
 FX_IMPORTER_CRASHING_SET = LINALG_CRASHING_SET | {
@@ -936,6 +950,17 @@
     "UpSampleNearest2dBackward_basic",
     "ViewCollapseDynamicWithAtenSizeIntModule_basic",
     "ViewSizeFromOtherTensor_basic",
+    # Error: `aten.as_strided` op is not supported
+    "ChunkListUnpackDynamic_Module_basic",
+    "ChunkListUnpackUnevenDynamic_Module_basic",
+    "ChunkListUnpackUneven_Module_basic",
+    "ChunkListUnpack_Module_basic",
+    "SplitTensorGetItem_Module_basic",
+    "SplitTensorLastSmallerModule_basic",
+    "SplitTensorListUnpackModule_basic",
+    "SplitTensorNegativeDimModule_basic",
+    "SplitWithSizesListUnpackModule_basic",
+    "SplitWithSizes_Module_basic",
 }
 
 FX_IMPORTER_STABLEHLO_CRASHING_SET = {
@@ -2503,6 +2528,20 @@
         "ScaledDotProductAttentionSameModule_basic",
     }
 
+if torch_version_for_comparison() > version.parse("2.6.0.dev"):
+    MAKE_FX_TOSA_PASS_SET = MAKE_FX_TOSA_PASS_SET - {
+        "ChunkListUnpackUneven_Module_basic",
+        "ChunkListUnpack_Module_basic",
+        "SplitTensorGetItem_Module_basic",
+        "SplitTensorLastSmallerModule_basic",
+        "SplitTensorListUnpackModule_basic",
+        "SplitTensorNegativeDimModule_basic",
+        "SplitWithSizesListUnpackModule_basic",
+        "TensorsSplitTensorLastSmallerModule_basic",
+        "TensorsSplitTensorModule_basic",
+        "TensorsSplitTensorNegativeDimModule_basic",
+    }
+
 LTC_CRASHING_SET = {
     # TODO: update test to move all inputs to the lazy device. Otherwise test fails with:
     # Check failed: lazy_tensor Input tensor is not a lazy tensor: CPUBoolType.

pytorch-hash.txt

@@ -1 +1 @@
-995ec16c7adf111348db617fa59e22e7ef9d7a3c
+79d8db50043ace9938cbbf4230b3515894452271

pytorch-requirements.txt

@@ -4,4 +4,4 @@
 # release page, and we use this page as an additional source for the wheels.
 -f https://xilinx.github.io/torch-mlir/package-index/
 --pre
-torch==2.5.0.dev20240909
+torch==2.6.0.dev20240916

test/Conversion/TorchOnnxToTorch/simple_ops_a_to_f.mlir

@@ -2016,6 +2016,24 @@ func.func @test_flatten_1d_axis_1(%arg0: !torch.vtensor<[2],f32>) -> !torch.vten
 
 // -----
 
+// CHECK-LABEL: func.func @test_constant_of_shape_arg_input
+func.func @test_constant_of_shape_arg_input(%arg0: !torch.vtensor<[2], si64>) -> !torch.vtensor<[?,?], f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
+  // CHECK: %[[INT0:.*]] = torch.constant.int 0
+  // CHECK: %[[INT0_0:.*]] = torch.constant.int 0
+  // CHECK: %[[EXTRACT_0:.*]] = torch.aten.select.int %arg0, %[[INT0]], %[[INT0_0]] : !torch.vtensor<[2],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
+  // CHECK: %[[ELE_0:.*]] = torch.aten.item %[[EXTRACT_0]] : !torch.vtensor<[],si64> -> !torch.int
+  // CHECK: %[[INT1:.*]] = torch.constant.int 1
+  // CHECK: %[[EXTRACT_1:.*]] = torch.aten.select.int %arg0, %[[INT0]], %[[INT1]] : !torch.vtensor<[2],si64>, !torch.int, !torch.int -> !torch.vtensor<[],si64>
+  // CHECK: %[[ELE_1:.*]] = torch.aten.item %[[EXTRACT_1]] : !torch.vtensor<[],si64> -> !torch.int
+  // CHECK: %[[DIM_LIST:.*]] = torch.prim.ListConstruct %[[ELE_0]], %[[ELE_1]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[NONE:.*]] = torch.constant.none
+  // CHECK: %[[FILL_VAL:.*]] = torch.constant.float 0.000000e+00
+  // CHECK: %[[ATEN_FULL:.*]] = torch.aten.full %[[DIM_LIST]], %[[FILL_VAL]], %[[NONE]], %[[NONE]], %[[NONE]], %[[NONE]] : !torch.list<int>, !torch.float, !torch.none, !torch.none, !torch.none, !torch.none -> !torch.vtensor<[?,?],f32>
+  %0 = "torch.operator"(%arg0) <{name = "onnx.ConstantOfShape"}> : (!torch.vtensor<[2], si64>) -> !torch.vtensor<[?,?], f32>
+  return %0 : !torch.vtensor<[?,?], f32>
+}
+// -----
+
 // CHECK-LABEL: func.func @test_constant_of_shape_dense_float_default
 func.func @test_constant_of_shape_dense_float_default() -> !torch.vtensor<[2,3,4], f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
   // CHECK: %[[SHAPE_CST:.*]] = torch.vtensor.literal(dense<[2, 3, 4]> : tensor<3xsi64>) : !torch.vtensor<[3],si64>

torchvision-requirements.txt

@@ -4,4 +4,4 @@
 # release page, and we use this page as an additional source for the wheels.
 -f https://xilinx.github.io/torch-mlir/package-index/
 --pre
-torchvision==0.20.0.dev20240909
+torchvision==0.20.0.dev20240916