fix: Bugfix in Linear-to-AddMM Fusion Lowering Pass

core/lowering/passes/linear_to_addmm.cpp

@@ -3,6 +3,7 @@
 #include "core/util/prelude.h"
 #include "torch/csrc/jit/api/function_impl.h"
 #include "torch/csrc/jit/ir/alias_analysis.h"
+#include "torch/csrc/jit/ir/irparser.h"
 #include "torch/csrc/jit/jit_log.h"
 #include "torch/csrc/jit/passes/constant_propagation.h"
 #include "torch/csrc/jit/passes/dead_code_elimination.h"
@@ -16,26 +17,58 @@ namespace core {
 namespace lowering {
 namespace passes {
 
-void replaceLinearWithBiasNonePattern(std::shared_ptr<torch::jit::Graph> graph) {
+void replaceLinear(torch::jit::Block* block) {
   // Define the decomposition function for aten::linear for the case where bias (mat2) is None.
   static torch::jit::CompilationUnit decompose_funcs(R"SCRIPT(
      def linear(self: Tensor, mat1: Tensor, mat2: Tensor):
          return torch.matmul(self, mat1.t())
      )SCRIPT");
 
-  // Iterate through nodes and search for aten::linear nodes where bias is not a Tensor (includes bias=None case)
-  auto block = graph->block();
+  // Define graph format for aten::linear with Tensor-type bias
+  std::string fused_linear = R"IR(
+        graph(%input, %weight, %bias):
+            %1: int = prim::Constant[value=1]()
+            %weight = aten::t(%weight)
+            %mm: Tensor = aten::matmul(%input, %weight)
+            %b_f: Tensor = trt::const(%bias)
+            %out: Tensor = aten::add(%b_f, %mm, %1)
+            return (%out))IR";
+
+  // Iterate through nodes in block, seaching for aten::linear
   for (auto it = block->nodes().begin(); it != block->nodes().end(); it++) {
     auto n = *it;
-    if (n->kind().toQualString() == std::string("aten::linear")) {
+
+    // Recursively explore nested blocks, such as those arising from prim::If
+    for (auto block : n->blocks()) {
+      replaceLinear(block);
+    }
+
+    if ((n->kind().toQualString() == std::string("aten::linear")) && (n->inputs().size() >= 3)) {
       auto input_values = n->inputs();
-      // input_values[2] is the bias. If none, replace it with the decomposed linear graph.
+
+      // input_values[2] is the bias
+      // If Tensor, replace with fused-bias decomposed graph
+      // Otherwise, replace it with the no-bias decomposed linear graph.
       if (input_values[2]->type()->isSubtypeOf(c10::TensorType::get())) {
-        continue;
+        torch::jit::WithInsertPoint guard(*it);
+
+        // Initialize new fused subgraph from IR code above
+        auto fused_g = std::make_shared<torch::jit::Graph>();
+        torch::jit::parseIR(fused_linear, fused_g.get());
+
+        // Insert subgraph in place of aten::linear, replacing inputs and outputs accordingly
+        torch::jit::Value* new_output = insertGraph(*it->owningGraph(), *fused_g, it->inputs()).at(0);
+        new_output->setType(it->output()->type());
+        it->output()->replaceAllUsesWith(new_output);
+        it.destroyCurrent();
       } else {
         torch::jit::WithInsertPoint guard(*it);
+
+        // Initialized decomposed graph without bias term
         std::shared_ptr<torch::jit::Graph> d_graph = toGraphFunction(decompose_funcs.get_function("linear")).graph();
         torch::jit::Value* new_output = insertGraph(*it->owningGraph(), *d_graph, it->inputs()).at(0);
+
+        // Insert function in place of aten::linear, replacing inputs and outputs accordingly
         new_output->setType(it->output()->type());
         it->output()->replaceAllUsesWith(new_output);
         it.destroyCurrent();
@@ -45,27 +78,8 @@ void replaceLinearWithBiasNonePattern(std::shared_ptr<torch::jit::Graph> graph)
 }
 
 void LinearToAddMM(std::shared_ptr<torch::jit::Graph>& graph) {
-  // TensorRT implicitly adds a flatten layer infront of FC layers if necessary
-  std::string flatten_linear_pattern = R"IR(
-        graph(%input, %weight, %bias):
-            %res = aten::linear(%input, %weight, %bias)
-            return (%res))IR";
-
-  std::string fused_linear = R"IR(
-        graph(%input, %weight_t, %bias):
-            %1: int = prim::Constant[value=1]()
-            %weight = aten::t(%weight_t)
-            %mm: Tensor = aten::matmul(%input, %weight)
-            %b_f: Tensor = trt::const(%bias)
-            %out: Tensor = aten::add(%b_f, %mm, %1)
-            return (%out))IR";
-
-  // First find and replace aten::linear nodes with non-tensor bias values.
-  replaceLinearWithBiasNonePattern(graph);
-
-  torch::jit::SubgraphRewriter flatten_linear_to_linear;
-  flatten_linear_to_linear.RegisterRewritePattern(flatten_linear_pattern, fused_linear);
-  flatten_linear_to_linear.runOnGraph(graph);
+  // Recursively find and replace all instances of aten::linear with the corresponding decomposed form
+  replaceLinear(graph->block());
 }
 
 } // namespace passes

tests/core/lowering/test_linear_to_addmm.cpp

@@ -57,3 +57,81 @@ TEST(LoweringPasses, LinearToAddMMBiasNone) {
 
   ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
 }
+
+TEST(LoweringPasses, LinearToAddMMBiasNoneGraphRun) {
+  std::string source_graph = R"IR(
+    graph(%input, %weight):
+      %biasNone : None = prim::Constant()
+      %true : bool = prim::Constant[value=1]()
+      %invalid_weight : Tensor = aten::t(%weight)
+      %4 : Tensor = prim::If(%true)
+        block0():
+          %res = aten::linear(%input, %weight, %biasNone)
+          -> (%res)
+        block1():
+          %res = aten::linear(%input, %invalid_weight, %biasNone)
+          -> (%res)
+      return (%4))IR";
+
+  // This regression test case ensures the Linear-to-AddMM lowering pass satisfies two constraints for non-Tensor bias:
+  // 1. It recursively resolves sub-blocks within the node, replacing sub-blocks to be converted as well
+  // 2. It does not pre-evaluate branches of the block which may have invalid operations
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, g.get());
+
+  auto in_0 = at::rand({8, 7}, {at::kCUDA});
+  auto in_1 = at::rand({8, 7}, {at::kCUDA});
+
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {in_0, in_1});
+
+  torch_tensorrt::core::lowering::passes::LinearToAddMM(g);
+
+  LOG_DEBUG(g);
+
+  params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {in_0, in_1});
+
+  ASSERT_TRUE(
+      torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0].reshape_as(jit_results[0]), 2e-6));
+}
+
+TEST(LoweringPasses, LinearToAddMMBiasGraphRun) {
+  std::string source_graph = R"IR(
+    graph(%input, %weight, %bias):
+      %true : bool = prim::Constant[value=1]()
+      %invalid_weight : Tensor = aten::t(%weight)
+      %4 : Tensor = prim::If(%true)
+        block0():
+          %res = aten::linear(%input, %weight, %bias)
+          -> (%res)
+        block1():
+          %res = aten::linear(%input, %invalid_weight, %bias)
+          -> (%res)
+      return (%4))IR";
+
+  // This regression test case ensures the Linear-to-AddMM lowering pass satisfies two constraints for Tensor bias:
+  // 1. It recursively resolves sub-blocks within the node, replacing sub-blocks to be converted as well
+  // 2. It does not pre-evaluate branches of the block which may have invalid operations
+
+  auto g = std::make_shared<torch::jit::Graph>();
+  torch::jit::parseIR(source_graph, g.get());
+
+  auto in_0 = at::rand({8, 7}, {at::kCUDA});
+  auto in_1 = at::rand({8, 7}, {at::kCUDA});
+  auto in_2 = at::rand({8, 8}, {at::kCUDA});
+
+  auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {in_0, in_1, in_2});
+
+  torch_tensorrt::core::lowering::passes::LinearToAddMM(g);
+
+  LOG_DEBUG(g);
+
+  params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
+  auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {in_0, in_1, in_2});
+
+  ASSERT_TRUE(
+      torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0].reshape_as(jit_results[0]), 2e-6));
+}