softmax mask fuse upper triangle (#33981)

* softmax mask fuse upper triangle

* cover not implemented cpu code

paddle/fluid/operators/softmax_mask_fuse_upper_triangle_op.cc

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#include "paddle/fluid/operators/softmax_mask_fuse_upper_triangle_op.h"
+#include "paddle/fluid/framework/generator.h"
+#include "paddle/fluid/framework/op_registry.h"
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class SoftmaxMaskFuseUpperTriangleOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X",
+                   "SoftmaxMaskFuseUpperTriangle");
+    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out",
+                   "SoftmaxMaskFuseUpperTriangle");
+
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_EQ(
+        x_dims.size(), 4,
+        platform::errors::InvalidArgument("Input x must be in 4D dimension but "
+                                          "received the dimension of X is %d",
+                                          x_dims.size()));
+
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", "Out");
+  }
+};
+
+class SoftmaxMaskFuseUpperTriangleOpMaker
+    : public framework::OpProtoAndCheckerMaker {
+ public:
+  void Make() override {
+    AddInput("X",
+             "The input of softmax_mask_fuse_upper_triangle op, "
+             "which is the result of matmul(QK)/sqrt(dk).");
+    AddOutput("Out", "The result of softmax_mask_fuse_upper_triangle op.");
+
+    AddComment(R"DOC(
+Softmax Mask Fuse Operator.
+product = matmul(QK)/sqrt(dk)
+output = softmax_mask_fuse_upper_triangle(product)
+to get the final output.
+)DOC");
+  }
+};
+
+class SoftmaxMaskFuseUpperTriangleOpGrad
+    : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
+                   framework::GradVarName("Out"),
+                   "SoftmaxMaskFuseUpperTriangleGrad");
+
+    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
+    ctx->SetOutputDim(framework::GradVarName("X"), out_dims);
+    ctx->ShareLoD(framework::GradVarName("Out"), framework::GradVarName("X"));
+  }
+};
+
+template <typename T>
+class SoftmaxMaskFuseUpperTriangleGradOpMaker
+    : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("softmax_mask_fuse_upper_triangle_grad");
+    op->SetInput("Softmax", this->Output("Out"));
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(
+    softmax_mask_fuse_upper_triangle, ops::SoftmaxMaskFuseUpperTriangleOp,
+    ops::SoftmaxMaskFuseUpperTriangleOpMaker,
+    ops::SoftmaxMaskFuseUpperTriangleGradOpMaker<paddle::framework::OpDesc>,
+    ops::SoftmaxMaskFuseUpperTriangleGradOpMaker<paddle::imperative::OpBase>);
+REGISTER_OPERATOR(softmax_mask_fuse_upper_triangle_grad,
+                  ops::SoftmaxMaskFuseUpperTriangleOpGrad);
+REGISTER_OP_CPU_KERNEL(softmax_mask_fuse_upper_triangle,
+                       ops::SoftmaxMaskFuseUpperTriangleCPUKernel<
+                           paddle::platform::CPUDeviceContext, float>,
+                       ops::SoftmaxMaskFuseUpperTriangleCPUKernel<
+                           paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/softmax_mask_fuse_upper_triangle_op.h

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+#include "paddle/fluid/framework/eigen.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+template <typename DeviceContext, typename T>
+class SoftmaxMaskFuseUpperTriangleCPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx.GetPlace()), true,
+                      platform::errors::Unimplemented(
+                          "Softmax mask fuse op only supports GPU now."));
+  }
+};
+}  // namespace operators
+}  // namespace paddle

python/paddle/fluid/tests/unittests/test_softmax_mask_fuse_upper_triangle_op.py

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+import unittest
+import numpy as np
+import paddle.fluid.core as core
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+import paddle.incubate as incubate
+
+paddle.enable_static()
+
+
+def _get_softmax_upper(x, fp16=True):
+    x_lower = np.tril(x)
+    masked_x = np.where(x_lower == 0, -10000.0, x_lower).astype("float32")
+    max_value = np.max(masked_x, axis=-1, keepdims=True)
+    before_exp = masked_x - max_value
+    exp = np.exp(before_exp)
+    exp_sum = np.sum(exp, axis=-1, keepdims=True)
+    rst = exp / exp_sum
+    if fp16:
+        rst = rst.astype("float16")
+    return rst
+
+
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestSoftmaxMaskFuseOp(OpTest):
+    def setUp(self):
+        self.op_type = "softmax_mask_fuse_upper_triangle"
+        x = np.random.random((1, 1, 32, 32)).astype("float16")
+        self.inputs = {'X': x}
+        rst = _get_softmax_upper(x)
+        self.outputs = {'Out': rst}
+
+    def test_check_output(self):
+        self.check_output_with_place(core.CUDAPlace(0))
+
+    def test_check_grad(self):
+        self.check_grad_with_place(core.CUDAPlace(0), ["X"], "Out")
+
+
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestSoftmaxMaskFuseOp1(OpTest):
+    def setUp(self):
+        self.op_type = "softmax_mask_fuse_upper_triangle"
+        x = np.random.random((1, 1, 32, 32))
+        self.inputs = {'X': x}
+        rst = _get_softmax_upper(x)
+        self.outputs = {'Out': rst}
+
+    def test_check_output(self):
+        try:
+            self.check_output_with_place(core.CPUPlace())
+        except NotImplementedError:
+            pass
+
+    def test_check_grad(self):
+        try:
+            self.check_grad_with_place(core.CPUPlace(), ["X"], "Out")
+        except NotImplementedError:
+            pass
+
+
+@unittest.skipIf(not core.is_compiled_with_cuda(),
+                 "core is not compiled with CUDA")
+class TestDropoutBiasFuseOp2(unittest.TestCase):
+    # test the python side API for softmax_mask_fuse op
+    def setUp(self):
+        np.random.seed(123)
+        self.dtypes = ['float16', 'float32']
+
+    def test_static(self):
+        for dtype in self.dtypes:
+            with fluid.program_guard(fluid.Program(), fluid.Program()):
+                input_x = fluid.data(
+                    name="x", shape=[1, 1, 32, 32], dtype=dtype)
+                rst = incubate.softmax_mask_fuse_upper_triangle(input_x)
+
+                x_in_np = np.random.random((1, 1, 32, 32)).astype(dtype)
+                rst_np = _get_softmax_upper(x_in_np, dtype == 'float16')
+
+                exe = fluid.Executor(fluid.CUDAPlace(0))
+                fetches = exe.run(fluid.default_main_program(),
+                                  feed={"x": x_in_np},
+                                  fetch_list=[rst])
+                self.assertTrue(np.allclose(fetches[0], rst_np))
+
+    def test_dygraph(self):
+        for dtype in self.dtypes:
+            with fluid.dygraph.guard(fluid.CUDAPlace(0)):
+                x_in_np = np.random.random((1, 1, 32, 32)).astype(dtype)
+                rst_np = _get_softmax_upper(x_in_np, dtype == 'float16')
+                input_x = fluid.dygraph.to_variable(x_in_np)
+
+                rst = incubate.softmax_mask_fuse_upper_triangle(input_x)
+                self.assertTrue(np.allclose(rst, rst_np))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/incubate/__init__.py

@@ -16,7 +16,8 @@ from .optimizer import LookAhead  # noqa: F401
 from .optimizer import ModelAverage  # noqa: F401
 from .checkpoint import auto_checkpoint  # noqa: F401
 from ..fluid.layer_helper import LayerHelper  # noqa: F401
+from .operators import softmax_mask_fuse_upper_triangle  # noqa: F401
 
 __all__ = [  # noqa
-    'LookAhead', 'ModelAverage'
+    'LookAhead', 'ModelAverage', 'softmax_mask_fuse_upper_triangle'
 ]

python/paddle/incubate/operators/__init__.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .softmax_mask_fuse_upper_triangle import softmax_mask_fuse_upper_triangle  # noqa: F401

python/paddle/incubate/operators/softmax_mask_fuse_upper_triangle.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+
+from paddle.fluid.layer_helper import LayerHelper
+from paddle.fluid.framework import in_dygraph_mode
+from paddle.fluid import core
+
+
+def softmax_mask_fuse_upper_triangle(x):
+    """
+    Fuse softmax mask together without even give a mask.
+    Under GPT model, the mask is always be a upper triangle
+    so we can simply mask the upper triangle part of x to get the mask result
+    :param x: the input x (rst of QK)
+    :return: the result of softmax mask fuse (upper triangle)
+    """
+    if in_dygraph_mode():
+        out = core.ops.softmax_mask_fuse_upper_triangle(x)
+        return out
+
+    helper = LayerHelper('softmax_mask_fuse_upper_triangle', **locals())
+
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
+
+    helper.append_op(
+        type='softmax_mask_fuse_upper_triangle',
+        inputs={'X': [x]},
+        outputs={'Out': [out]})
+    return out

python/setup.py.in

@@ -146,6 +146,7 @@ packages=['paddle',
           'paddle.incubate',
           'paddle.incubate.optimizer',
           'paddle.incubate.checkpoint',
+          'paddle.incubate.operators',
           'paddle.distributed.fleet',
           'paddle.distributed.fleet.base',
           'paddle.distributed.fleet.meta_optimizers',