Support MatMulBnb4 in PaddingElimination (#18646)

Also support Cast pattern between input and embedding node for sparsity
inspecting

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

@@ -282,7 +282,8 @@ void IterateSubgraphFromNode(Graph& graph,
       ORT_ENFORCE(subgraph.find(cur->MutableInputDefs()[0]) != subgraph.end());
       subgraph.insert(cur->MutableOutputDefs()[0]);
       PushAllOutputNode(graph, to_visit, cur, visited);
-    } else if (graph_utils::IsSupportedOptypeVersionAndDomain(*cur, "MatMul", {1, 9, 13})) {
+    } else if (graph_utils::IsSupportedOptypeVersionAndDomain(*cur, "MatMul", {1, 9, 13}) ||
+               graph_utils::IsSupportedOptypeVersionAndDomain(*cur, "MatMulBnb4", {1}, kMSDomain)) {
       if (subgraph.find(cur->MutableInputDefs()[0]) != subgraph.end()) {
         // If shape of [batch_size, seqlen, ...] is propagated from the first argument of MatMul.
         // The dim size of the first argument must be larger than 2 to propagate the first two dims to the output.

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

@@ -157,12 +157,7 @@ def _initialize_embedding_padding_inspector(self, model, user_input_names):
         self._embedding_graph_input_to_padding_idx_map.clear()
 
         for node in model.graph.node:
-            if not (
-                node.domain == "org.pytorch.aten"
-                and node.op_type == "ATen"
-                and node.input[1] in user_input_names
-                and len(node.input) >= 3
-            ):
+            if not (node.domain == "org.pytorch.aten" and node.op_type == "ATen" and len(node.input) >= 3):
                 continue
 
             found = [attr for attr in node.attribute if attr.name == "operator"]
@@ -194,10 +189,29 @@ def _initialize_embedding_padding_inspector(self, model, user_input_names):
             if padding_idx < 0:
                 continue
 
-            if node.input[1] not in self._embedding_graph_input_to_padding_idx_map:
-                self._embedding_graph_input_to_padding_idx_map[node.input[1]] = set()
+            # Given the input arg of embedding node, find the corresponding user input that feeds into the data.
+            # Will iterate the args recursively if some subgraph pattern is found between the input and the embedding,
+            # such as Input -> Cast -> Cast -> Embedding.
+            # TODO: This is a workaround for the case that the input of embedding is a list of Cast nodes which is found
+            # in Llama-2. We need to find a general way to handle all types of subgraph parttern between input and embedding.
+            def _get_embedding_graph_input(node_arg):
+                if node_arg in user_input_names:
+                    return node_arg
+                input_node = self._try_get_node_from_its_output(node_arg)
+                if input_node.op_type == "Cast":
+                    return _get_embedding_graph_input(input_node.input[0])
+                else:
+                    self._logger.warning(f"Cannot find embedding input {node_arg}")
+                    return None
+
+            embedding_graph_input = _get_embedding_graph_input(node.input[1])
+            if embedding_graph_input is None:
+                continue
+
+            if embedding_graph_input not in self._embedding_graph_input_to_padding_idx_map:
+                self._embedding_graph_input_to_padding_idx_map[embedding_graph_input] = set()
 
-            self._embedding_graph_input_to_padding_idx_map[node.input[1]].add(padding_idx)
+            self._embedding_graph_input_to_padding_idx_map[embedding_graph_input].add(padding_idx)
 
     def _initialize_loss_label_padding_inspector(self, model, user_input_names):
         """Register loss label input padding inspector.