add cross_entropy backward rule (#56507)

pkuzyc · web-flow · commit 1642e84bb550 · 2023-09-08T14:01:11.000+08:00
diff --git a/paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.cc b/paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.cc
@@ -34,7 +34,7 @@ CrossEntropyWithSoftmaxSPMDRule::InferForward(
                                    input_specs_size));
 
   auto x_shape = input_specs[0].shape();
-  int x_ndim = static_cast<int>(x_shape.size());
+  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();
 
@@ -173,10 +173,116 @@ CrossEntropyWithSoftmaxSPMDRule::InferForward(
 std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
 CrossEntropyWithSoftmaxSPMDRule::InferBackward(
     const std::vector<DistTensorSpec>& input_specs,
+    const std::vector<DistTensorSpec>& output_specs,
     const paddle::framework::AttributeMap& attrs) {
-  PADDLE_THROW(phi::errors::Unimplemented(
-      "InferBackward of CrossEntropyWithSoftmaxSPMDRule is NOT implemented "
-      "yet."));
+  // step0: verify input args based on cross_entropy_with_softmax logic
+  int64_t ninputs = input_specs.size();
+  int64_t noutputs = output_specs.size();
+  PADDLE_ENFORCE_EQ(
+      ninputs,
+      2,
+      phi::errors::InvalidArgument("The size of InputSpec of cross entropy "
+                                   "with softmax should be 2, but got [%d].",
+                                   ninputs));
+  PADDLE_ENFORCE_EQ(
+      noutputs,
+      2,
+      phi::errors::InvalidArgument("The size of OutputSpec of cross entropy "
+                                   "with softmax should be 2, but got [%d].",
+                                   noutputs));
+  VerifySpecs(output_specs, "cross_entropy_with_softmax_backward");
+
+  // step1: build Einsum Notation
+  std::vector<int64_t> x_shape = input_specs[0].shape();
+  int64_t x_ndim = x_shape.size();
+  std::vector<int64_t> label_shape = input_specs[1].shape();
+
+  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);
+
+  // normalize axis
+  if (axis < 0) {
+    axis = x_ndim + axis;
+  }
+
+  std::string alphabet =
+      "abcdefghijlmnopqrstuvwxyz";  // k for softmax_normalize axis
+  std::string x_axes = GetBroadcastAxes(x_ndim, x_ndim, alphabet);
+  x_axes[axis] = 'k';
+  std::string label_axes = x_axes;
+  if (!soft_label) {
+    label_axes[axis] = '1';
+  }
+  std::string loss_axes = x_axes;
+  loss_axes[axis] = '1';
+  // optional output
+  std::string softmax_out_axes;
+  if (use_softmax) {
+    softmax_out_axes = x_axes;
+  } else {
+    softmax_out_axes = "";
+  }
+
+  // step2: Sharding Propogation
+  // step2.1 merge output dims mappings
+  std::vector<std::pair<std::string, std::vector<int64_t>>> axes_sharding_info;
+  axes_sharding_info =
+      GetAxesDimsMappingPair({softmax_out_axes, loss_axes}, output_specs);
+  std::unordered_map<std::string, int64_t> axis_to_dim_map =
+      ShardingMergeForTensors(axes_sharding_info);
+
+  // step2.2 infer inputs' dims mappings from merged dims mapping
+  std::vector<TensorDistAttr> input_dist_attrs;
+  input_dist_attrs.emplace_back(input_specs[0].dist_attr());
+  input_dist_attrs.emplace_back(input_specs[1].dist_attr());
+  // infer and set input X's dims mapping
+  input_dist_attrs[0].set_dims_mapping(
+      GetDimsMappingForAxes(x_axes, axis_to_dim_map));
+  // infer and set input label's dims mapping
+  input_dist_attrs[1].set_dims_mapping(
+      GetDimsMappingForAxes(label_axes, axis_to_dim_map));
+
+  // step2.3 update outputs' dims mappings with merged dims mapping
+  std::vector<TensorDistAttr> output_dist_attrs;
+  output_dist_attrs.emplace_back(output_specs[0].dist_attr());  // softmax_out
+  output_dist_attrs.emplace_back(output_specs[1].dist_attr());  // loss
+  output_dist_attrs[0].set_dims_mapping(
+      GetDimsMappingForAxes(softmax_out_axes, axis_to_dim_map));
+  output_dist_attrs[1].set_dims_mapping(
+      GetDimsMappingForAxes(loss_axes, axis_to_dim_map));
+
+  // step3: Handle partial state (TODO)
+
+  VLOG(4) << "CrossEntropyWithSoftmaxSPMDRule InferBackward: "
+          << "axis: " << axis << ", ignore_index: " << ignore_index
+          << ", numeric_stable_mode: "
+          << (numeric_stable_mode ? "true" : "false")
+          << ", use_softmax: " << use_softmax
+          << ", soft_label: " << (soft_label ? "true" : "false");
+  VLOG(4) << "Einsum notation: [" << x_axes << "," << label_axes << " --> "
+          << softmax_out_axes << "," << loss_axes << "]. (inputs --> outputs)";
+  for (int64_t i = 0; i < noutputs; i++) {
+    VLOG(4) << "Output" << std::to_string(i) << ": "
+            << "shape: [" << str_join(output_specs[i].shape())
+            << "], src_dims_mapping: ["
+            << str_join(output_specs[i].dims_mapping())
+            << "], dst_dims_mapping: ["
+            << str_join(output_dist_attrs[i].dims_mapping()) << "]";
+  }
+  for (int64_t i = 0; i < ninputs; i++) {
+    VLOG(4) << "Input" << std::to_string(i) << ": "
+            << "shape: [" << str_join(input_specs[i].shape())
+            << "], infered_dims_mapping: ["
+            << str_join(input_dist_attrs[i].dims_mapping()) << "]";
+  }
+  VLOG(4) << std::endl;
+
+  // according to the phi api implemetation, the softmax_out tensor will alway
+  // be genereated not matter the value of use_softmax.
+  return {input_dist_attrs, output_dist_attrs};
 }
 
 }  // namespace auto_parallel
diff --git a/paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h b/paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h
@@ -27,7 +27,8 @@ class CrossEntropyWithSoftmaxSPMDRule : public SPMDRuleBase {
                const paddle::framework::AttributeMap& attrs) override;
 
   std::pair<std::vector<TensorDistAttr>, std::vector<TensorDistAttr>>
-  InferBackward(const std::vector<DistTensorSpec>& output_specs,
+  InferBackward(const std::vector<DistTensorSpec>& input_specs,
+                const std::vector<DistTensorSpec>& output_specs,
                 const paddle::framework::AttributeMap& attrs) override;
 };
 }  // namespace auto_parallel
diff --git a/test/auto_parallel/spmd_rules/test_cross_entropy_with_softmax_rule.py b/test/auto_parallel/spmd_rules/test_cross_entropy_with_softmax_rule.py
@@ -39,6 +39,9 @@ def setUp(self):
             label_shape, label_tensor_dist_attr
         )
 
+        self.loss_spec = DistTensorSpec(self.lable_dist_tensor_spec)
+        self.softmax_out_spec = DistTensorSpec(self.x_dist_tensor_spec)
+
         self.attrs = {
             'ignore_index': -1,
             'axis': -1,
@@ -147,6 +150,122 @@ def test_cross_entropy_with_softmax_infer_forward(self):
             )
         self.attrs['axis'] = -1
 
+    def test_cross_entropy_with_softmax_infer_backward(self):
+        # GPT DP case
+        # [1, 0, -1], [1, 0, -1] (outputs) -->
+        # [1, 0, -1], [1, 0, -1], (inputs)
+        # [1, 0, -1], [1, 0, -1] (outputs)
+        self.attrs['axis'] = -1
+        self.attrs['use_softmax'] = True
+        self.attrs['soft_label'] = False
+        self.softmax_out_spec.set_dims_mapping([1, 0, -1])
+        self.loss_spec.set_dims_mapping([1, 0, -1])
+
+        result_dist_attrs = self.rule1.infer_backward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+            [self.softmax_out_spec, self.loss_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[0].dims_mapping, [1, 0, -1]
+        )  # softmax output
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [1, 0, -1]
+        )  # loss
+
+        # GPT MP case, shard normalized axis
+        # [-1, -1, 0], [-1, -1, -1] (outputs) -->
+        # [-1, -1, 0], [-1, -1, -1], (inputs)
+        # [-1, -1, 0], [-1, -1, -1] (outputs)
+        self.attrs['axis'] = -1
+        self.attrs['use_softmax'] = True
+        self.attrs['soft_label'] = False
+        self.softmax_out_spec.set_dims_mapping([-1, -1, 0])
+        self.loss_spec.set_dims_mapping([-1, -1, -1])
+
+        result_dist_attrs = self.rule1.infer_backward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+            [self.softmax_out_spec, self.loss_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[0].dims_mapping, [-1, -1, 0]
+        )  # softmax output
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [-1, -1, -1]
+        )  # loss
+
+        # GPT MP-DP case
+        # [-1, -1, 0], [1, -1, -1] (outputs) -->
+        # [1, -1, 0], [1, -1, -1], (inputs)
+        # [1, -1, 0], [1, -1, -1] (outputs)
+        self.attrs['axis'] = -1
+        self.attrs['use_softmax'] = True
+        self.attrs['soft_label'] = False
+        self.softmax_out_spec.set_dims_mapping([-1, -1, 0])
+        self.loss_spec.set_dims_mapping([1, -1, -1])
+
+        result_dist_attrs = self.rule1.infer_backward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+            [self.softmax_out_spec, self.loss_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[0].dims_mapping, [1, -1, 0]
+        )  # softmax output
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [1, -1, -1]
+        )  # loss
+
+        # Soft Label, normalized axis = 1
+        # [1, -1, 0], [1, -1, -1] (outputs) -->
+        # [1, -1, 0], [1, -1, 0], (inputs)
+        # [1, -1, 0], [1, -1, 0] (outputs)
+        self.attrs['axis'] = 1
+        self.attrs['use_softmax'] = True
+        self.attrs['soft_label'] = True
+        self.softmax_out_spec.set_dims_mapping([1, -1, 0])
+        self.loss_spec.set_dims_mapping([1, -1, -1])
+        result_dist_attrs = self.rule1.infer_backward(
+            [self.x_dist_tensor_spec, self.lable_dist_tensor_spec],
+            [self.softmax_out_spec, self.loss_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[0].dims_mapping, [1, -1, 0]
+        )  # softmax output
+        self.assertEqual(
+            infered_output_dist_attrs[1].dims_mapping, [1, -1, 0]
+        )  # loss
+
 
 if __name__ == "__main__":
     unittest.main()
