[Semi Auto] Entropy SPMD Rule (PaddlePaddle#55394)

* base rule

* add sharidng merge

* add sharidng axis merge

* define unified data class for inferencing dist_attr

* test wrap DistTensorSpec in dygraph mode

* matmul main logic done

* shape int64

* common cc

* define unified data class for inferencing dist_attr

* test wrap DistTensorSpec in dygraph mode

* define python api and wrap function in static mode for DistTensorSpec

* revise syntax

* map bugfix

* broadcast func

* compile 1

* add unitest

* add registry

* update unitest

* bugfix

* bugfix

* add pybind

* bugfix

* bugfix macro gloabl name space

* bugfix macro gloabl name space

* pybind

* pybind test

* pybind bugfixed1

* pybind bugfixed2

* pybind unitest

* merge dev

* merge dev

* merge dev

* fixed cmake conflict

* fixed cmake conflict

* rename get method

* revise inferforward output type

* revise comment

* replicated rule

* replicated rule 2

* revert bug deps

* add rule

* add unitest

* add rule

* add unitest

* move ut of auto_parallel

* fix ut

* bugfix

* bugfix

* bugfix

* bugfix

* bugfix

* bugfix

* bugfix

* resolute input sharding conflict maybe

* fixed comment

* add rule

* add unitest

* fixed typoes

---------

Co-authored-by: Yichen Zhang <zhangyichen03@baidu.com>
Co-authored-by: zhiqiu <chenqiuliang@baidu.com>

paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.cc

@@ -0,0 +1,184 @@
+/* Copyright (c) 2023 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/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h"
+
+namespace paddle {
+namespace distributed {
+namespace auto_parallel {
+
+using phi::distributed::auto_parallel::str_join;
+
+std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
+CrossEntropyWithSoftmaxSPMDRule::InferForward(
+    const std::vector<DistTensorSpec>& input_specs,
+    const paddle::framework::AttributeMap& attrs) {
+  // step0: verify input args based on cross_entropy_with_softmax logic
+  auto input_specs_size = input_specs.size();
+  PADDLE_ENFORCE_EQ(
+      input_specs_size,
+      2,
+      phi::errors::InvalidArgument("The size of InputSpec of cross entropy "
+                                   "with softmax should be 2, but got [%d].",
+                                   input_specs_size));
+
+  auto x_shape = input_specs[0].shape();
+  int x_ndim = x_shape.size();
+  auto x_dist_attr_src = input_specs[0].dist_attr();
+  std::vector<int64_t> x_dims_mapping_src = x_dist_attr_src.dims_mapping();
+
+  auto label_shape = input_specs[1].shape();
+  auto label_dist_attr_src = input_specs[1].dist_attr();
+  std::vector<int64_t> label_dims_mapping_src =
+      label_dist_attr_src.dims_mapping();
+
+  int axis = ExtractAttr<int>("axis", attrs);
+  int ignore_index = ExtractAttr<int>("ignore_index", attrs);
+  bool numeric_stable_mode = ExtractAttr<bool>("numeric_stable_mode", attrs);
+  bool use_softmax = ExtractAttr<bool>("use_softmax", attrs);
+  bool soft_label = ExtractAttr<bool>("soft_label", attrs);
+
+  VLOG(6) << "CrossEntropyWithSoftmaxSPMDRule InferForward Inputs: "
+          << "X shape: [" << str_join(x_shape) << "], x_dims_mapping_src: ["
+          << str_join(x_dims_mapping_src) << "]; Label shape: ["
+          << str_join(label_shape) << "], Label dims mapping: ["
+          << str_join(label_dims_mapping_src) << "]; axis: "
+          << "[" << axis << "], ignore_index: [" << ignore_index
+          << "], numeric_stable_mode: [" << numeric_stable_mode
+          << "], use_softmax: [" << use_softmax << "], soft_label: ["
+          << soft_label << "].";
+
+  // normalize axis
+  if (axis < 0) {
+    axis = x_ndim + axis;
+  }
+
+  // trying to shard the normal axis of softmax, BUT
+  // c_softmax_with_entropy kernel not support:
+  // 1. soft label
+  // 2. axis != -1
+  // support above two features in future.
+  if (x_dims_mapping_src[axis] > -1) {
+    PADDLE_ENFORCE_EQ(
+        soft_label,
+        false,
+        phi::errors::InvalidArgument(
+            "Trying to shard the softmax_normalize axis of the input tensor, "
+            "but the soft_label is set as True, which is not supported yet!"));
+
+    PADDLE_ENFORCE_EQ(
+        axis,
+        x_ndim - 1,
+        phi::errors::InvalidArgument(
+            "Trying to shard the softmax_normalize axis of the input tensor, "
+            "but the softmax_normalize axis is not the last axis, which is not "
+            "supported yet! The softmax_normalize is [%d].",
+            axis));
+
+    PADDLE_ENFORCE_EQ(use_softmax,
+                      true,
+                      phi::errors::InvalidArgument(
+                          "Trying to shard the softmax_normalize axis of the "
+                          "input tensor, use_softmax must be set to True !"));
+  }
+
+  // step1: build Einsum Notation
+  std::string alphabet =
+      "abcdefghijlmnopqrstuvwxyz";  // k for softmax_normalize axis
+  std::string broadcast_axes =
+      GetBroadcastAxes(x_ndim - 1, x_ndim - 1, alphabet);
+  std::string x_axes = broadcast_axes;
+  x_axes.insert(axis, "k");
+  std::string label_axes;
+  if (soft_label) {
+    label_axes = x_axes;
+  } else {
+    label_axes = broadcast_axes;
+    label_axes.insert(axis, "1");
+  }
+  std::string loss_axes = broadcast_axes;
+  loss_axes.insert(axis, "1");
+  // optional output
+  std::string softmax_out_axes;
+  if (use_softmax) {
+    softmax_out_axes = x_axes;
+  } else {
+    softmax_out_axes = "";
+  }
+
+  // step2: Sharding Propogation
+  std::vector<std::pair<std::string, std::vector<int64_t>>> axes_sharding_info;
+  axes_sharding_info =
+      GetAxesDimsMappingPair({x_axes, label_axes}, input_specs);
+  std::unordered_map<std::string, int64_t> axis_to_dim_map =
+      ShardingMergeForTensors(axes_sharding_info);
+
+  // step3: Infer dst Dims Mapping.
+  TensorDistAttr loss_dist_attr_dst =
+      CopyTensorDistAttrForOutput(label_dist_attr_src);
+  loss_dist_attr_dst.set_dims_mapping(
+      GetDimsMappingForAxes(loss_axes, axis_to_dim_map));
+  TensorDistAttr softmax_out_dist_attr_dst =
+      CopyTensorDistAttrForOutput(x_dist_attr_src);
+  softmax_out_dist_attr_dst.set_dims_mapping(
+      GetDimsMappingForAxes(softmax_out_axes, axis_to_dim_map));
+
+  TensorDistAttr x_dist_attr_dst = CopyTensorDistAttrForOutput(x_dist_attr_src);
+  x_dist_attr_dst.set_dims_mapping(
+      GetDimsMappingForAxes(x_axes, axis_to_dim_map));
+  TensorDistAttr label_dist_attr_dst =
+      CopyTensorDistAttrForOutput(label_dist_attr_src);
+  label_dist_attr_dst.set_dims_mapping(
+      GetDimsMappingForAxes(label_axes, axis_to_dim_map));
+
+  VLOG(4) << "CrossEntropyWithSoftmaxSPMDRule InferForward Inputs: "
+          << "Einsum notation: [" << x_axes << "," << label_axes << " --> "
+          << softmax_out_axes << "," << loss_axes << "]. " << std::endl
+          << "X shape: [" << str_join(x_shape) << "], x_dims_mapping_src: ["
+          << str_join(x_dims_mapping_src) << "], x_dims_mapping_dst: ["
+          << str_join(x_dist_attr_dst.dims_mapping()) << "]; Label shape: ["
+          << str_join(label_shape) << "], label_dims_mapping_src: ["
+          << str_join(label_dims_mapping_src) << "], label_dims_mapping_dst: ["
+          << str_join(label_dist_attr_dst.dims_mapping())
+          << "]; loss_dims_mapping: ["
+          << str_join(loss_dist_attr_dst.dims_mapping())
+          << "], softmax_out_dims_mapping_src: ["
+          << str_join(softmax_out_dist_attr_dst.dims_mapping()) << "]; axis: "
+          << "[" << axis << "], ignore_index: [" << ignore_index
+          << "], numeric_stable_mode: ["
+          << (numeric_stable_mode ? "true" : "false") << "], use_softmax: ["
+          << (use_softmax ? "true" : "false") << "], soft_label: ["
+          << (soft_label ? "true" : "false") << "].";
+
+  // todo if softmax_normalize axis is sharded, notify downstream phi api to
+  // select c_softmax_with_entropy_kernel.
+
+  // according to the phi api implemetation, the softmax_out tensor will alway
+  // be genereated not matter the value of use_softmax.
+  return {{x_dist_attr_dst, label_dist_attr_dst},
+          {softmax_out_dist_attr_dst, loss_dist_attr_dst}};
+}
+
+std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
+CrossEntropyWithSoftmaxSPMDRule::InferBackward(
+    const std::vector<DistTensorSpec>& input_specs,
+    const paddle::framework::AttributeMap& attrs) {
+  PADDLE_THROW(phi::errors::Unimplemented(
+      "InferBackward of CrossEntropyWithSoftmaxSPMDRule is NOT implemented "
+      "yet."));
+}
+
+}  // namespace auto_parallel
+}  // namespace distributed
+}  // namespace paddle

paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h

@@ -0,0 +1,35 @@
+/* Copyright (c) 2023 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/distributed/auto_parallel/spmd_rules/common.h"
+
+namespace paddle {
+namespace distributed {
+namespace auto_parallel {
+
+class CrossEntropyWithSoftmaxSPMDRule : public SPMDRuleBase {
+ public:
+  std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
+  InferForward(const std::vector<DistTensorSpec>& input_specs,
+               const paddle::framework::AttributeMap& attrs) override;
+
+  std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
+  InferBackward(const std::vector<DistTensorSpec>& output_specs,
+                const paddle::framework::AttributeMap& attrs) override;
+};
+}  // namespace auto_parallel
+}  // namespace distributed
+}  // namespace paddle

paddle/fluid/distributed/auto_parallel/spmd_rules/rules.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/common.h"
+#include "paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/elementwise_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/embedding_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/layer_norm_spmd_rule.h"
@@ -145,6 +146,10 @@ REGISTER_SPMD_RULE(lookup_table_v2, EmbeddingSPMDRule);
 REGISTER_SPMD_RULE(softmax, SoftmaxSPMDRule);
 REGISTER_SPMD_RULE(log_softmax, SoftmaxSPMDRule);
 
+// cross_entropy_with_softmax
+REGISTER_SPMD_RULE(cross_entropy_with_softmax, CrossEntropyWithSoftmaxSPMDRule);
+REGISTER_SPMD_RULE(softmax_with_cross_entropy, CrossEntropyWithSoftmaxSPMDRule);
+
 }  // namespace auto_parallel
 }  // namespace distributed
 }  // namespace paddle

test/auto_parallel/spmd_rules/test_cross_entropy_with_softmax_rule.py

@@ -0,0 +1,152 @@
+# Copyright (c) 2023 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.
+
+import unittest
+
+from paddle.distributed.auto_parallel.static.completion import get_spmd_rule
+from paddle.distributed.auto_parallel.static.dist_attribute import (
+    DistTensorSpec,
+    TensorDistAttr,
+)
+from paddle.distributed.fleet import auto
+
+
+class TestCrossEntropyWithSoftmaxSPMDRule(unittest.TestCase):
+    def setUp(self):
+        self.rule1 = get_spmd_rule("cross_entropy_with_softmax")
+
+        x_shape = [8, 1024, 50304]  # [batch_size, max_seq_len, vocab_size]
+        label_shape = [8, 1024, 1]
+
+        process_mesh = auto.ProcessMesh(mesh=[[0, 1, 2, 3], [4, 5, 6, 7]])
+        x_tensor_dist_attr = TensorDistAttr()
+        x_tensor_dist_attr.process_mesh = process_mesh
+        self.x_dist_tensor_spec = DistTensorSpec(x_shape, x_tensor_dist_attr)
+        label_tensor_dist_attr = TensorDistAttr()
+        label_tensor_dist_attr.process_mesh = process_mesh
+        self.lable_dist_tensor_spec = DistTensorSpec(
+            label_shape, label_tensor_dist_attr
+        )
+
+        self.attrs = {
+            'ignore_index': -1,
+            'axis': -1,
+            'numeric_stable_mode': True,
+            'use_softmax': True,
+            'soft_label': False,
+        }
+
+    def test_cross_entropy_with_softmax_infer_forward(self):
+        # GPT DP case
+        self.x_dist_tensor_spec.set_dims_mapping([1, -1, -1])
+        self.lable_dist_tensor_spec.set_dims_mapping([-1, 0, -1])
+
+        result_dist_attrs = self.rule1.infer_forward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec], self.attrs
+        )
+        self.assertEqual(len(result_dist_attrs), 2)
+        infered_input_dist_attrs = result_dist_attrs[0]
+        infered_output_dist_attrs = result_dist_attrs[1]
+
+        self.assertEqual(len(infered_input_dist_attrs), 2)
+        self.assertEqual(len(infered_output_dist_attrs), 2)
+
+        self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [1, 0, -1])
+        self.assertEqual(infered_input_dist_attrs[1].dims_mapping, [1, 0, -1])
+
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [1, 0, -1]
+        )  # loss
+        self.assertEqual(
+            infered_output_dist_attrs[0].dims_mapping, [1, 0, -1]
+        )  # softmax output
+
+        # GPT MP case, shard normalized axis
+        self.x_dist_tensor_spec.set_dims_mapping([-1, -1, 0])
+        self.lable_dist_tensor_spec.set_dims_mapping([-1, -1, -1])
+
+        result_dist_attrs = self.rule1.infer_forward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec], self.attrs
+        )
+        infered_input_dist_attrs = result_dist_attrs[0]
+        infered_output_dist_attrs = result_dist_attrs[1]
+
+        self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [-1, -1, 0])
+        self.assertEqual(infered_input_dist_attrs[1].dims_mapping, [-1, -1, -1])
+
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [-1, -1, -1]
+        )  # loss
+        self.assertEqual(
+            infered_output_dist_attrs[0].dims_mapping, [-1, -1, 0]
+        )  # softmax output
+
+        # GPT MP-DP case
+        self.x_dist_tensor_spec.set_dims_mapping([-1, -1, 0])
+        self.lable_dist_tensor_spec.set_dims_mapping([1, -1, -1])
+
+        result_dist_attrs = self.rule1.infer_forward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec], self.attrs
+        )
+        infered_input_dist_attrs = result_dist_attrs[0]
+        infered_output_dist_attrs = result_dist_attrs[1]
+
+        self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [1, -1, 0])
+        self.assertEqual(infered_input_dist_attrs[1].dims_mapping, [1, -1, -1])
+
+        self.assertEqual(infered_output_dist_attrs[1].dims_mapping, [1, -1, -1])
+        self.assertEqual(infered_output_dist_attrs[0].dims_mapping, [1, -1, 0])
+
+        # Soft Label Error
+        self.attrs['soft_label'] = True
+        self.x_dist_tensor_spec.set_dims_mapping([-1, -1, 0])
+        self.lable_dist_tensor_spec.set_dims_mapping([1, -1, -1])
+        with self.assertRaises(ValueError):
+            result_dist_attrs = self.rule1.infer_forward(
+                [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+                self.attrs,
+            )
+        self.attrs['soft_label'] = False
+
+        # Normalized axis
+        self.attrs['axis'] = 1
+        self.x_dist_tensor_spec.set_dims_mapping([1, -1, 0])
+        self.lable_dist_tensor_spec.set_dims_mapping([-1, -1, -1])
+        result_dist_attrs = self.rule1.infer_forward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec], self.attrs
+        )
+        infered_input_dist_attrs = result_dist_attrs[0]
+        infered_output_dist_attrs = result_dist_attrs[1]
+
+        self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [1, -1, 0])
+        self.assertEqual(infered_input_dist_attrs[1].dims_mapping, [1, -1, 0])
+
+        self.assertEqual(infered_output_dist_attrs[1].dims_mapping, [1, -1, 0])
+        self.assertEqual(infered_output_dist_attrs[0].dims_mapping, [1, -1, 0])
+        self.attrs['axis'] = -1
+
+        # Soft Normalized axis Error
+        self.attrs['axis'] = 1
+        self.x_dist_tensor_spec.set_dims_mapping([-1, 0, -1])
+        self.lable_dist_tensor_spec.set_dims_mapping([1, -1, -1])
+        with self.assertRaises(ValueError):
+            result_dist_attrs = self.rule1.infer_forward(
+                [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+                self.attrs,
+            )
+        self.attrs['axis'] = -1
+
+
+if __name__ == "__main__":
+    unittest.main()