fix: Implement a patch for gelu schema change in older NGC containers

Signed-off-by: Dheeraj Peri <[email protected]>

Author: peri044

Diff for: core/lowering/passes/reduce_gelu.cpp

@@ -8,10 +8,17 @@ namespace passes {
 
 void ReduceGelu(std::shared_ptr<torch::jit::Graph>& graph) {
   std::string gelu_pattern = R"IR(
-        graph(%x):
+        graph(%x : Tensor):
             %out : Tensor = aten::gelu(%x)
             return (%out))IR";
 
+  // This gelu_approximate_pattern schema exists in 21.11, 21.12, 22.01 containers of pytorch. These container versions use
+  // an unmerged PR in pytorch : https://github.com/pytorch/pytorch/pull/61439. We reduce this to regular Gelu.
+  std::string gelu_approximate_pattern = R"IR(
+        graph(%x : Tensor, %approx):
+            %out : Tensor = aten::gelu(%x, %approx)
+            return (%out))IR";
+
   std::string gelu_reduce_pattern = R"IR(
     graph(%x.1 : Tensor):
         %6 : float = prim::Constant[value=0.044714999999999998]()
@@ -30,11 +37,36 @@ void ReduceGelu(std::shared_ptr<torch::jit::Graph>& graph) {
         %15 : Tensor = aten::mul(%7, %14)
         return (%15))IR";
 
+  // This is same as gelu_reduce_pattern except for an additional input %approx.
+  // SubgraphRewriter only works as expected if the number of inputs to gelu_approximate_pattern
+  // and gelu_reduce_multi_input_pattern are same.
+  std::string gelu_reduce_multi_input_pattern = R"IR(
+    graph(%x.1 : Tensor, %approx):
+        %6 : float = prim::Constant[value=0.044714999999999998]()
+        %5 : float = prim::Constant[value=0.79788456080000003]()
+        %4 : float = prim::Constant[value=1.]()
+        %3 : float = prim::Constant[value=0.5]()
+        %2 : int = prim::Constant[value=1]()
+        %7 : Tensor = aten::mul(%x.1, %3)
+        %8 : Tensor = aten::mul(%x.1, %5)
+        %9 : Tensor = aten::mul(%x.1, %6)
+        %10 : Tensor = aten::mul(%9, %x.1)
+        %11 : Tensor = aten::add(%10, %4, %2)
+        %12 : Tensor = aten::mul(%8, %11)
+        %13 : Tensor = aten::tanh(%12)
+        %14 : Tensor = aten::add(%13, %4, %2)
+        %15 : Tensor = aten::mul(%7, %14)
+        return (%15))IR";
+
   // replace aten::gelu with pointwise operations
   torch::jit::SubgraphRewriter map_gelu_to_pointwise_ops;
   map_gelu_to_pointwise_ops.RegisterRewritePattern(gelu_pattern, gelu_reduce_pattern);
   map_gelu_to_pointwise_ops.runOnGraph(graph);
 
+  torch::jit::SubgraphRewriter map_gelu_approximate_to_pointwise_ops;
+  map_gelu_approximate_to_pointwise_ops.RegisterRewritePattern(gelu_approximate_pattern, gelu_reduce_multi_input_pattern);
+  map_gelu_approximate_to_pointwise_ops.runOnGraph(graph);
+
   LOG_GRAPH("Post lowering of [aten::gelu] -> " << *graph);
 }
 

Diff for: tests/core/lowering/test_reduce_gelu.cpp

@@ -40,3 +40,38 @@ TEST(LoweringPasses, ReduceGeluCorrectly) {
 
   ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
 }
+
+TEST(LoweringPasses, ReduceGeluApproximateCorrectly) {
+  std::string source_graph = R"IR(
+    graph(%x, %approx):
+        %out : Tensor = aten::gelu(%x, %approx)
+        return (%out))IR";
+  std::string target_graph = R"IR(
+    graph(%x.1 : Tensor, %approx):
+        %6 : float = prim::Constant[value=0.044714999999999998]()
+        %5 : float = prim::Constant[value=0.79788456080000003]()
+        %4 : float = prim::Constant[value=1.]()
+        %3 : float = prim::Constant[value=0.5]()
+        %2 : int = prim::Constant[value=1]()
+        %7 : Tensor = aten::mul(%x.1, %3)
+        %8 : Tensor = aten::mul(%x.1, %5)
+        %9 : Tensor = aten::mul(%x.1, %6)
+        %10 : Tensor = aten::mul(%9, %x.1)
+        %11 : Tensor = aten::add(%10, %4, %2)
+        %12 : Tensor = aten::mul(%8, %11)
+        %13 : Tensor = aten::tanh(%12)
+        %14 : Tensor = aten::add(%13, %4, %2)
+        %15 : Tensor = aten::mul(%7, %14)
+        return (%15))IR";
+
+  torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
+      torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
+  auto sg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, &*sg);
+  torch_tensorrt::core::lowering::passes::ReduceGelu(sg);
+
+  auto tg = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(target_graph, &*tg);
+
+  ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
+}