Add attention op insertion code

shark/shark_importer.py

@@ -312,6 +312,51 @@ def get_f16_inputs(inputs, is_f16, f16_input_mask):
     return tuple(f16_masked_inputs)
 
 
+def insert_attention_block(fx_g):
+    import torch
+
+    unary_ops = [
+        torch.ops.aten._unsafe_view,
+        torch.ops.aten.view,
+        torch.ops.aten.expand,
+        torch.ops.aten.clone,
+    ]
+
+    def traverse(node):
+        while node.target in unary_ops:
+            node = node.args[0]
+        return node
+
+    for node in fx_g.graph.nodes:
+        if node.target in [torch.ops.aten.bmm]:
+            outer_bmm = node
+            node = traverse(outer_bmm.args[0])
+            if node.target in [torch.ops.aten._softmax]:
+                softmax_node = node
+                node = traverse(softmax_node.args[0])
+                if node.target in [torch.ops.aten.bmm]:
+                    inner_bmm = node
+                    value = outer_bmm.args[1]
+                    key = inner_bmm.args[1]
+                    with fx_g.graph.inserting_before(outer_bmm):
+                        key = fx_g.graph.call_function(
+                            torch.ops.aten.transpose,
+                            args=(key, -2, -1),
+                            kwargs={},
+                        )
+                        query = inner_bmm.args[0]
+                        new_node = fx_g.graph.call_function(
+                            torch.ops.aten.scaled_dot_product_attention,
+                            args=(query, key, value),
+                            kwargs={},
+                        )
+                        outer_bmm.append(new_node)
+                        outer_bmm.append(key)
+                        outer_bmm.replace_all_uses_with(new_node)
+
+    fx_g.graph.lint()
+
+
 # Upcasts the block/list of ops.
 def add_upcast(fx_g):
     import torch
@@ -640,6 +685,8 @@ def strip_overloads(gm):
 
     strip_overloads(fx_g)
 
+    insert_attention_block(fx_g)
+
     if is_f16:
         fx_g = fx_g.half()
         transform_fx(fx_g)
@@ -659,6 +706,7 @@ def strip_overloads(gm):
         return ts_graph
 
     inputs = get_f16_inputs(inputs, is_f16, f16_input_mask)
+
     mlir_importer = SharkImporter(
         ts_graph,
         inputs,