Check padding density by input of embedding module

orttraining/orttraining/core/optimizer/compute_optimizer/padding_elimination.cc

@@ -19,6 +19,7 @@ namespace {
 // TODO(pengwa): remove this once customized PythonOp shape inference is supported.
 constexpr const char* kInspectActivationFuncName = "onnxruntime.training.utils.hooks._subscriber_manager._InspectActivation";
 constexpr const char* kIncrementStepFuncName = "onnxruntime.training.utils.hooks._subscriber_manager._IncrementStep";
+constexpr const char* kFlagPaddingEliminationFuncName = "onnxruntime.training.ortmodule._graph_execution_manager._FlagPaddingElimination";
 
 void PushAllOutputNode(Graph& graph, std::queue<Node*>& q, Node* node, std::unordered_set<Node*>& visited) {
   for (auto iter = node->OutputNodesBegin(); iter != node->OutputNodesEnd(); ++iter) {
@@ -311,7 +312,7 @@ void IterateSubgraphFromNode(Graph& graph,
         candidate_outputs.insert(cur);
         continue;
       }
-      auto func_name = static_cast<std::string>(cur->GetAttributes().at("name").s());
+      auto func_name = static_cast<std::string>(cur->GetAttributes().at("func_name").s());
       if (func_name == kInspectActivationFuncName || func_name == kIncrementStepFuncName) {
         subgraph.insert(cur->MutableOutputDefs()[1]);
         PushAllOutputNode(graph, to_visit, cur, visited);
@@ -353,11 +354,6 @@ void IterateSubgraphFromNode(Graph& graph,
 Status PaddingElimination::ApplyImpl(Graph& graph, bool& modified, int graph_level, const logging::Logger& logger) const {
   LOG_DEBUG_INFO(logger, "Enter PaddingElimination");
 
-  if (sparse_embedding_input_names_.size() == 0) {
-    LOG_DEBUG_INFO(logger, "Exit PaddingElimination, no sparse embedding input names.");
-    return Status::OK();
-  }
-
   GraphViewer graph_viewer(graph);
   const auto& node_topology_list = graph_viewer.GetNodesInTopologicalOrder();
   Node* embedding_node = nullptr;
@@ -386,13 +382,28 @@ Status PaddingElimination::ApplyImpl(Graph& graph, bool& modified, int graph_lev
         node.InputDefs()[2]->Exists() &&
         graph_utils::IsConstantInitializer(graph, node.InputDefs()[2]->Name()) &&
         node.InputDefs()[1]->Exists() &&
-        graph_utils::IsGraphInput(graph, node.InputDefs()[1]) &&
         node.InputDefs()[1]->Shape() &&
         node.InputDefs()[1]->Shape()->dim_size() >= 2) {
-      if (std::find(sparse_embedding_input_names_.begin(), sparse_embedding_input_names_.end(),
-                    node.InputDefs()[1]->Name()) == sparse_embedding_input_names_.end()) {
-        LOG_DEBUG_INFO(logger, "Skip node " + node.Name() + "(" + node.OpType() +
-                                   ") due to embedding input is not in the sparse embedding input list.");
+      const auto outputNodeCount = std::distance(node.OutputEdgesBegin(), node.OutputEdgesEnd());
+      if (outputNodeCount != 1) {
+          continue;
+      }
+      auto embedding_output_node = graph.GetNode(node.OutputNodesBegin()->Index());
+      if (embedding_output_node == nullptr ||
+          !graph_utils::IsSupportedOptypeVersionAndDomain(*embedding_output_node, "PythonOp", {1}, kMSDomain) ||
+          static_cast<std::string>(embedding_output_node->GetAttributes().at("func_name").s()) != kFlagPaddingEliminationFuncName) {
+        LOG_DEBUG_INFO(logger, "not find PythonOp of flagPaddingElimination after embedding node");
+        continue;
+      }
+      if (graph_utils::CanRemoveNode(graph, *embedding_output_node, logger)) {
+        if (graph_utils::RemoveNode(graph, *embedding_output_node)) {
+          modified = true;
+        } else {
+          LOG_DEBUG_INFO(logger, "Failed to remove node " + embedding_output_node->Name() + "(" + embedding_output_node->OpType() + ")");
+          continue;
+        }
+      } else {
+        LOG_DEBUG_INFO(logger, "Can not remove node " + embedding_output_node->Name() + "(" + embedding_output_node->OpType() + ")");
         continue;
       }
       const ONNX_NAMESPACE::TensorProto* padding_initializer =
@@ -479,7 +490,6 @@ Status PaddingElimination::ApplyImpl(Graph& graph, bool& modified, int graph_lev
   // to flattern the shape of [batch_size, seqlen, ...] to [valid_token_count, ...]
   InsertFlattenPatternForInput(graph, *embedding_node, 1, squeeze_out_arg, logger);
   handled_input_count++;
-  modified = true;
   for (auto& node : candidate_inputs) {
     for (uint32_t i = 0; i < node->InputDefs().size(); ++i) {
       if (subgraph.find(node->MutableInputDefs()[i]) == subgraph.end()) {

orttraining/orttraining/core/optimizer/compute_optimizer/padding_elimination.h

@@ -127,15 +127,10 @@ namespace onnxruntime {
  */
 class PaddingElimination : public GraphTransformer {
  public:
-  PaddingElimination(const InlinedHashSet<std::string_view>& compatible_execution_providers = {},
-                     const std::vector<std::string>& sparse_embedding_input_names = {}) noexcept
-      : GraphTransformer("PaddingElimination", compatible_execution_providers),
-        sparse_embedding_input_names_{sparse_embedding_input_names} {}
+  PaddingElimination(const InlinedHashSet<std::string_view>& compatible_execution_providers = {}) noexcept
+      : GraphTransformer("PaddingElimination", compatible_execution_providers) {}
 
   Status ApplyImpl(Graph& graph, bool& modified, int graph_level, const logging::Logger& logger) const override;
-
- private:
-  std::vector<std::string> sparse_embedding_input_names_;
 };
 
 }  // namespace onnxruntime

orttraining/orttraining/core/optimizer/graph_transformer_config.h

@@ -25,9 +25,6 @@ struct TrainingGraphTransformerConfiguration : public GraphTransformerConfigurat
   // Enable compute optimizer.
   bool enable_compute_optimizer{false};
 
-  // Enable embedding sparsity compute optimization for the input names in the below list.
-  std::vector<std::string> sparse_embedding_input_names;
-
   // Enable label sparsity compute optimization for the input names in the below list.
   std::vector<std::string> sparse_label_input_names;
 

orttraining/orttraining/core/optimizer/graph_transformer_utils.cc

@@ -195,8 +195,7 @@ std::vector<std::unique_ptr<GraphTransformer>> GeneratePreTrainingTransformers(
 #if defined(USE_CUDA) || defined(USE_ROCM)
         // Put this under CUDA/ROCM guard as it depends on PadAndUnflatten CUDA/ROCM kernel.
         // Once we have a CPU kernel for PadAndUnflatten, we can remove the guard.
-        transformers.emplace_back(std::make_unique<PaddingElimination>(compatible_eps,
-                                                                       config.sparse_embedding_input_names));
+        transformers.emplace_back(std::make_unique<PaddingElimination>(compatible_eps));
         transformers.emplace_back(std::make_unique<Conv1dReplacement>(compatible_eps));
 #endif
       }

orttraining/orttraining/python/orttraining_pybind_state.cc

@@ -488,7 +488,6 @@ void addObjectMethodsForTraining(py::module& m) {
       .def_readwrite("transformer_layer_recompute", &TrainingGraphTransformerConfiguration::transformer_layer_recompute)
       .def_readwrite("number_recompute_layers", &TrainingGraphTransformerConfiguration::number_recompute_layers)
       .def_readwrite("enable_compute_optimizer", &TrainingGraphTransformerConfiguration::enable_compute_optimizer)
-      .def_readwrite("sparse_embedding_input_names", &TrainingGraphTransformerConfiguration::sparse_embedding_input_names)
       .def_readwrite("sparse_label_input_names", &TrainingGraphTransformerConfiguration::sparse_label_input_names)
       .def_readwrite("optimized_pre_grad_filepath", &TrainingGraphTransformerConfiguration::optimized_pre_grad_filepath)
       .def_readwrite("propagate_cast_ops_config", &TrainingGraphTransformerConfiguration::GraphTransformerConfiguration::propagate_cast_ops_config);

orttraining/orttraining/python/training/ortmodule/_graph_execution_manager.py

@@ -10,7 +10,7 @@
 import os
 from abc import ABC, abstractmethod  # noqa: F401
 from hashlib import md5 as hash_fn
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional, Tuple, Union
 
 import onnx
 import torch
@@ -48,6 +48,28 @@ def __init__(self, state, output_info: List[Tuple[torch.Size, torch.device, torc
         self.state = state
         self.output_info = output_info
 
+class _FlagPaddingElimination(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, input):
+        return input
+
+    @staticmethod
+    def backward(ctx, grad_output: torch.Tensor):
+        return grad_output
+
+    @staticmethod
+    def infer_shape(
+        node: onnx.NodeProto,
+        tensor_input_shapes: List[Optional[List[Union[int, str]]]],
+        tensor_input_dtypes: List[torch.onnx.TensorProtoDataType],
+    ) -> Tuple[List[Optional[List[Union[int, str]]]], List[torch.onnx.TensorProtoDataType]]:
+        return tensor_input_shapes, tensor_input_dtypes
+
+    @staticmethod
+    def alias_input(node_proto_str: str):
+        fw_alias_map = [0]
+        bw_alias_map = [0]
+        return fw_alias_map, bw_alias_map
 
 class GraphExecutionManager(GraphExecutionInterface):
     def __init__(
@@ -91,6 +113,7 @@ def __init__(
         # Inspector for runtime information, for example input data, memory usage, etc.
         self._runtime_inspector = RuntimeInspector(self._logger, self._original_module)
         self._runtime_inspector.memory_ob.enable_memory_stats_by_step(self._runtime_options.print_memory_stat_by_step)
+        self._embedding_module_to_padding_density_map = {}
 
         # Tracker for ORTModule model export, session creation overhead.
         self.time_tracker = _logger.TimeTracker()
@@ -622,6 +645,38 @@ def __setstate__(self, state):
 
         _utils.reinitialize_graph_execution_manager(self)
 
+    def _check_embedding_sparsity(self):
+        if not self._runtime_options.enable_embedding_sparse_optimizer or self._device.type != "cuda":
+            return
+        def embedding_hook(module, args, output):
+            ebd_input = args[0]
+            if ebd_input is None or not isinstance(ebd_input, torch.Tensor):
+                self._logger.warning("Embedding input is not a tensor.")
+                return None
+
+            valid_token = torch.count_nonzero(ebd_input - module.padding_idx)
+            total_token = ebd_input.numel()
+            embed_density = float(valid_token) / float(total_token) * 100
+            if embed_density < 90:
+                self._logger.info("Embedding sparsity-based optimization is ON for density: %.0f%%", embed_density)
+                if module not in self._embedding_module_to_padding_density_map:
+                    self._logger.warning("Found Embedding module not in the map. %s", module)
+                    return None
+                if module in self._embedding_module_to_padding_density_map and self._embedding_module_to_padding_density_map[module][1] != -1:
+                    self._logger.warning(
+                        "Found duplicate Embedding module. %s",
+                        self._embedding_module_to_padding_density_map[module][0]
+                    )
+                self._embedding_module_to_padding_density_map[module][1] = embed_density
+                return _FlagPaddingElimination.apply(output)
+            return None
+
+        for name, sub_module in self._flattened_module.named_modules():
+            if isinstance(sub_module, torch.nn.modules.sparse.Embedding):
+                if sub_module.padding_idx >= 0:
+                    self._embedding_module_to_padding_density_map[sub_module] = [name, -1]
+                    sub_module.register_forward_hook(embedding_hook)
+
     @_logger.TrackTime(_logger.ORTModuleInitPhase.DETECTION)
     def _enable_conditional_optimizations(
         self, graph_transformer_config: C.TrainingGraphTransformerConfiguration, inputs: Tuple, kwargs: Dict
@@ -680,16 +735,10 @@ def _enable_conditional_optimizations(
                         [f"{k}:{v:.0f}%" for k, v in label_sparsity_results.items()]
                     )
 
-                if self._runtime_options.enable_embedding_sparse_optimizer and len(embed_sparsity_results) > 0:
-                    if detected_device.type == "cuda":
-                        # Embedding sparsity optimization is only supported on CUDA devices.
-                        graph_transformer_config.sparse_embedding_input_names = list(embed_sparsity_results.keys())
-                        self._logger.info("Embedding sparsity-based optimization is ON for %s", embed_sparsity_results)
-                        self._runtime_options.embed_sparsity_ratio = ",".join(
-                            [f"{k}:{v:.0f}%" for k, v in embed_sparsity_results.items()]
-                        )
-                    else:
-                        self._logger.info("Embedding sparsity-based optimization is not supported on non-CUDA devices.")
+                if self._embedding_module_to_padding_density_map:
+                    self._runtime_options.embed_sparsity_ratio = ",".join(
+                        [f"{v[0]}:{v[1]:.0f}%" for v in self._embedding_module_to_padding_density_map.values()]
+                    )
 
             # If users don't want to print input density, disable the input density observer to avoid overhead
             # when looping through inputs during training.

orttraining/orttraining/python/training/ortmodule/_io.py

@@ -245,10 +245,8 @@ def _expand_inputs(current_input, non_none_inputs, name=""):
             if PrimitiveType.is_primitive_type(inp):
                 inp = PrimitiveType.get_tensor(inp, device)
 
-            found, embedding_density, label_density = rt_inspector.inspect_input(name, inp)
+            found, _, label_density = rt_inspector.inspect_input(name, inp)
             if found:
-                if embedding_density < 100:
-                    embed_sparsity_results[name] = embedding_density
                 if label_density < 100:
                     label_sparsity_results[name] = label_density
             result.append(inp)

orttraining/orttraining/python/training/ortmodule/_runtime_inspector.py

@@ -136,7 +136,6 @@ def initialize(self, model: ModelProto, user_input_names: List[str]) -> None:
                     if output_name != "":
                         self._tensor_to_node_map[output_name] = node
 
-            self._initialize_embedding_padding_inspector(model, user_input_names)
             self._initialize_loss_label_padding_inspector(model, user_input_names)
 
             self._is_initialized = True
@@ -354,33 +353,6 @@ def _inspect_embed_label_input(self, name, data):
         found = False
         min_embed_density = 100
         min_label_density = 100
-        if (
-            len(self._embedding_graph_input_to_padding_idx_map) > 0
-            and name in self._embedding_graph_input_to_padding_idx_map
-            and isinstance(data, torch.Tensor)
-        ):
-            for padding_idx in self._embedding_graph_input_to_padding_idx_map[name]:
-                valid_token = torch.count_nonzero(data - padding_idx)
-                valid_token_per_batch = "N/A"
-                if data.dim() > 1:
-                    valid_token_per_batch = str(torch.count_nonzero(data - padding_idx, dim=1).tolist())
-                total_token = data.numel()
-                embed_density = float(valid_token) / float(total_token) * 100
-                if embed_density < 90:
-                    min_embed_density = min(min_embed_density, embed_density)
-                self._stats.append(
-                    [
-                        self._current_step,
-                        "EMBED",
-                        name,
-                        padding_idx,
-                        embed_density,
-                        valid_token,
-                        total_token,
-                        valid_token_per_batch,
-                    ]
-                )
-                found = True
 
         if (
             len(self._loss_label_graph_input_to_ignore_idx_map) > 0

orttraining/orttraining/python/training/ortmodule/_training_manager.py

@@ -253,6 +253,8 @@ def forward(self, *inputs, **kwargs):
             ):
                 self.time_tracker.start(ORTModuleInitPhase.EndToEnd)
 
+                self._check_embedding_sparsity()
+
                 build_gradient_graph = self._export_model(*inputs, **kwargs)
 
                 if build_gradient_graph: