Add Decompostion for Aten_SafeSoftmaxOp (#3708)

zjgarvey · vivekkhandelwal1 · web-flow · commit d61986cfcf30 · 2024-09-12T16:58:10.000-05:00
Co-authored-by: Vivek Khandelwal &lt;vivekkhandelwal1424@gmail.com&gt;
diff --git a/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td b/include/torch-mlir/Dialect/Torch/IR/GeneratedTorchOps.td
@@ -8370,6 +8370,31 @@ def Torch_Aten_SoftmaxOp : Torch_Op<"aten._softmax", [
   }];
 }
 
+def Torch_Aten_SafeSoftmaxOp : Torch_Op<"aten._safe_softmax", [
+    AllowsTypeRefinement,
+    HasValueSemantics,
+    ReadOnly
+  ]> {
+  let summary = "Generated op for `aten::_safe_softmax : (Tensor, int, int?) -> (Tensor)`";
+  let arguments = (ins
+    AnyTorchTensorType:$self,
+    Torch_IntType:$dim,
+    AnyTorchOptionalIntType:$dtype
+  );
+  let results = (outs
+    AnyTorchOptionalTensorType:$result
+  );
+  let hasCustomAssemblyFormat = 1;
+  let extraClassDefinition = [{
+    ParseResult Aten_SafeSoftmaxOp::parse(OpAsmParser &parser, OperationState &result) {
+      return parseDefaultTorchOp(parser, result, 3, 1);
+    }
+    void Aten_SafeSoftmaxOp::print(OpAsmPrinter &printer) {
+      printDefaultTorchOp(printer, *this, 3, 1);
+    }
+  }];
+}
+
 def Torch_AtenMeanOp : Torch_Op<"aten.mean", [
     AllowsTypeRefinement,
     HasValueSemantics,
diff --git a/lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp b/lib/Dialect/Torch/Transforms/AbstractInterpLibrary.cpp
@@ -6772,6 +6772,10 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
 "  }\n"
+"  func.func @\"__torch_mlir_shape_fn.aten._safe_softmax\"(%arg0: !torch.list<int>, %arg1: !torch.int, %arg2: !torch.optional<int>) -> !torch.list<int> {\n"
+"    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
+"    return %0 : !torch.list<int>\n"
+"  }\n"
 "  func.func @\"__torch_mlir_shape_fn.aten.softmax.int\"(%arg0: !torch.list<int>, %arg1: !torch.int, %arg2: !torch.optional<int>) -> !torch.list<int> {\n"
 "    %0 = call @__torch__.torch.jit._shape_functions.unary(%arg0) : (!torch.list<int>) -> !torch.list<int>\n"
 "    return %0 : !torch.list<int>\n"
@@ -15367,6 +15371,18 @@ StringRef mlir::torch::Torch::getAbstractInterpLibrary() {
 "    }\n"
 "    return %1 : !torch.int\n"
 "  }\n"
+"  func.func @\"__torch_mlir_dtype_fn.aten._safe_softmax\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.int, %arg2: !torch.optional<int>) -> !torch.int {\n"
+"    %none = torch.constant.none\n"
+"    %0 = torch.aten.__isnot__ %arg2, %none : !torch.optional<int>, !torch.none -> !torch.bool\n"
+"    %1 = torch.prim.If %0 -> (!torch.int) {\n"
+"      %2 = torch.prim.unchecked_cast %arg2 : !torch.optional<int> -> !torch.int\n"
+"      torch.prim.If.yield %2 : !torch.int\n"
+"    } else {\n"
+"      %2:2 = torch.prim.TupleUnpack %arg0 : !torch.tuple<int, int> -> !torch.int, !torch.int\n"
+"      torch.prim.If.yield %2#1 : !torch.int\n"
+"    }\n"
+"    return %1 : !torch.int\n"
+"  }\n"
 "  func.func @\"__torch_mlir_dtype_fn.aten._log_softmax\"(%arg0: !torch.tuple<int, int>, %arg1: !torch.int, %arg2: !torch.bool) -> !torch.int {\n"
 "    %int6 = torch.constant.int 6\n"
 "    %none = torch.constant.none\n"
diff --git a/lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp b/lib/Dialect/Torch/Transforms/DecomposeComplexOps.cpp
@@ -2148,6 +2148,62 @@ class DecomposeAten_SoftmaxOp : public OpRewritePattern<Aten_SoftmaxOp> {
 };
 } // namespace
 
+// Ref:
+// https://github.com/pytorch/pytorch/blob/5314ae2660a778b87987030182f787bb6cb092c0/aten/src/ATen/native/transformers/attention.cpp#L663-L673
+namespace {
+class DecomposeAten_SafeSoftmaxOp
+    : public OpRewritePattern<Aten_SafeSoftmaxOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(Aten_SafeSoftmaxOp op,
+                                PatternRewriter &rewriter) const override {
+    BaseTensorType resultTensorType = cast<BaseTensorType>(op.getType());
+    if (!resultTensorType.hasDtype() || !resultTensorType.hasSizes()) {
+      return rewriter.notifyMatchFailure(
+          op, "expected result type to have sizes and dtype");
+    }
+    SmallVector<int64_t> sizes(resultTensorType.getSizes());
+
+    int64_t dimInt;
+    if (!matchPattern(op.getDim(), m_TorchConstantInt(&dimInt)))
+      return rewriter.notifyMatchFailure(op, "Unsupported: non-constant dim");
+
+    dimInt = toPositiveDim(dimInt, sizes.size());
+    if (!isValidDim(dimInt, sizes.size()))
+      return rewriter.notifyMatchFailure(op, "dim int is not valid");
+
+    Location loc = op.getLoc();
+    Value softmax = rewriter.create<AtenSoftmaxIntOp>(
+        loc, op.getType(), op.getSelf(), op.getDim(), op.getDtype());
+
+    Type resultTensorDtype = resultTensorType.getDtype();
+
+    Value negInfinity = getConstantWithGivenDtypeAndValue(
+        rewriter, loc, -std::numeric_limits<double>::infinity(),
+        resultTensorDtype);
+
+    auto boolDtype = rewriter.getI1Type();
+    auto boolTensorType =
+        resultTensorType.getWithSizesAndDtype(sizes, boolDtype);
+    Value masked = rewriter.create<AtenEqScalarOp>(loc, boolTensorType,
+                                                   op.getSelf(), negInfinity);
+
+    sizes[dimInt] = 1;
+    auto maskedRowsType =
+        resultTensorType.getWithSizesAndDtype(sizes, boolDtype);
+    Value cstTrue =
+        rewriter.create<Torch::ConstantBoolOp>(loc, rewriter.getBoolAttr(true));
+    Value maskedRows = rewriter.create<AtenAllDimOp>(
+        loc, maskedRowsType, masked, op.getDim(), cstTrue);
+    Value cstZero = getConstantWithGivenDtypeAndValue(rewriter, loc, 0.0,
+                                                      resultTensorDtype);
+    rewriter.replaceOpWithNewOp<AtenWhereScalarSelfOp>(
+        op, resultTensorType, maskedRows, cstZero, softmax);
+    return success();
+  }
+};
+} // namespace
+
 // Aten_SoftmaxBackwardDataOp(gradOutput, output, dim) =>
 //    newGrad = gradOutput * output
 //    result = newGrad - output * sum(newGrad, dim))
@@ -9608,6 +9664,7 @@ class DecomposeComplexOpsPass
         patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenSoftmaxIntOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAten_SoftmaxOp>(patterns);
+    addPatternIfTargetOpIsIllegal<DecomposeAten_SafeSoftmaxOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAten_LogSoftmaxOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenLogSoftmaxIntOp>(patterns);
     addPatternIfTargetOpIsIllegal<DecomposeAtenLogSigmoidOp>(patterns);
diff --git a/lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp b/lib/Dialect/Torch/Transforms/LowerToBackendContract.cpp
@@ -371,6 +371,7 @@ static void markDecomposedOpsAsIllegal(MLIRContext *context,
                                        llvm::StringSet<> backendLegalOpsSet) {
   target.addIllegalOp<AtenSoftmaxIntOp>();
   target.addIllegalOp<Aten_SoftmaxOp>();
+  target.addIllegalOp<Aten_SafeSoftmaxOp>();
   target.addIllegalOp<Aten_LogSoftmaxOp>();
   target.addIllegalOp<AtenLogSoftmaxIntOp>();
   target.addIllegalOp<AtenLogSigmoidOp>();
diff --git a/projects/pt1/e2e_testing/xfail_sets.py b/projects/pt1/e2e_testing/xfail_sets.py
@@ -504,14 +504,6 @@
     "ViewCollapseDynamicWithAtenSizeIntModule_basic",
     "ViewSizeFromOtherTensor_basic",
     "WeightNormInterfaceModule_basic",
-    # REMOVE WHEN ENABLE_GQA IS ADDED
-    "ScaledDotProductAttentionBoolMaskModule_basic",
-    "ScaledDotProductAttentionDifferentCausalModule_basic",
-    "ScaledDotProductAttentionDifferentModule_basic",
-    "ScaledDotProductAttentionMaskModule_basic",
-    "ScaledDotProductAttentionSameCausalModule_basic",
-    "ScaledDotProductAttentionSameDynamicModule_basic",
-    "ScaledDotProductAttentionSameModule_basic",
 }
 
 FX_IMPORTER_CRASHING_SET = LINALG_CRASHING_SET | {
@@ -826,6 +818,9 @@
     "ReplicationPad2dModule_top0",
     "RsubInt0d_NumToTensor_Module_basic",
     "ScalarImplicitFloatModule_basic",
+    # need aten.all.dim lowering to stablehlo
+    "SafeSoftmaxModule_basic",
+    "SafeSoftmaxNonNoneDtypeModule_basic",
     # REMOVE WHEN ENABLE_GQA IS ADDED
     "ScaledDotProductAttentionBoolMaskModule_basic",
     "ScaledDotProductAttentionDifferentCausalModule_basic",
@@ -2770,6 +2765,8 @@
     "ReshapeAliasExpandModule_basic",
     "ReshapeExpandModule_basic",
     "Rot90DynamicDimsModule_basic",
+    "SafeSoftmaxModule_basic",
+    "SafeSoftmaxNonNoneDtypeModule_basic",
     "ScalarConstantTupleModule_basic",
     "ScalarImplicitFloatModule_basic",
     "ScalarImplicitIntModule_basic",
diff --git a/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py b/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/abstract_interp_lib_gen.py
@@ -348,6 +348,9 @@ def aten〇glu〡shape(self: List[int], dim: int = -1) -> List[int]:
 def aten〇_softmax〡shape(self: List[int], dim: int, half_to_float: bool) -> List[int]:
     return upstream_shape_functions.unary(self)
 
+def aten〇_safe_softmax〡shape(self: List[int], dim: int, dtype: Optional[int] = None) -> List[int]:
+    return upstream_shape_functions.unary(self)
+
 def aten〇softmax〇int〡shape(self: List[int], dim: int, dtype: Optional[int] = None) -> List[int]:
     return upstream_shape_functions.unary(self)
 
@@ -5426,6 +5429,12 @@ def aten〇_softmax〡dtype(self_rank_dtype: Tuple[int, int], dim: int, half_to_
         return torch.float32
     return self_dtype
 
+def aten〇_safe_softmax〡dtype(self_rank_dtype: Tuple[int, int], dim: int, dtype: Optional[int] = None) -> int:
+    if dtype is not None:
+        return dtype
+    self_rank, self_dtype = self_rank_dtype
+    return self_dtype
+
 @check_dtype_function(
     # _check_tensors_with_the_same_dtype(num_of_tensors=1, dim=0, half_to_float=False) +
     _check_tensors_with_the_same_dtype(
diff --git a/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py b/projects/pt1/python/torch_mlir/jit_ir_importer/build_tools/torch_ods_gen.py
@@ -692,6 +692,7 @@ def emit_with_mutating_variants(key, **kwargs):
     emit("aten::__lshift__.Scalar : (Tensor, Scalar) -> (Tensor)")
     emit("aten::__rshift__.Scalar : (Tensor, Scalar) -> (Tensor)")
     emit("aten::_softmax : (Tensor, int, bool) -> (Tensor)")
+    emit("aten::_safe_softmax : (Tensor, int, int?) -> (Tensor)")
     emit("aten::mean : (Tensor, int?) -> (Tensor)")
     emit("aten::std : (Tensor, bool) -> (Tensor)")
     emit("aten::std.dim : (Tensor, int[]?, bool, bool) -> (Tensor)")
diff --git a/projects/pt1/python/torch_mlir_e2e_test/test_suite/basic.py b/projects/pt1/python/torch_mlir_e2e_test/test_suite/basic.py
@@ -1907,6 +1907,52 @@ def _LogSoftmaxModuleStable_basic(module, tu: TestUtils):
 # ==============================================================================
 
 
+class SafeSoftmaxModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, tensor):
+        return torch.ops.aten._safe_softmax(tensor, dim=0)
+
+
+@register_test_case(module_factory=lambda: SafeSoftmaxModule())
+def SafeSoftmaxModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 2, 4))
+
+
+# ==============================================================================
+
+
+class SafeSoftmaxNonNoneDtypeModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    @export
+    @annotate_args(
+        [
+            None,
+            ([-1, -1, -1], torch.float32, True),
+        ]
+    )
+    def forward(self, tensor):
+        return torch.ops.aten._safe_softmax(tensor, dim=2, dtype=torch.float64)
+
+
+@register_test_case(module_factory=lambda: SafeSoftmaxNonNoneDtypeModule())
+def SafeSoftmaxNonNoneDtypeModule_basic(module, tu: TestUtils):
+    module.forward(tu.rand(3, 2, 4))
+
+
+# ==============================================================================
+
+
 class SoftplusModule(torch.nn.Module):
     def __init__(self):
         super().__init__()
