fix(//core/lowering): Fixes module level fallback recursion

This commit fixes module level fallback by using method calls to determine modules to recurse down too. This should be robust to names other than forward used for methods as well as ignoring functional modules. Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
pytorch · Oct 1, 2021 · 2fc612d · 2fc612d
1 parent 722aa94
commit 2fc612d
Show file tree

Hide file tree

Showing 2 changed files with 37 additions and 15 deletions.
diff --git a/core/lowering/lowering.cpp b/core/lowering/lowering.cpp
@@ -37,7 +37,9 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, LowerInfo lower_info) {
     torch::jit::EliminateCommonSubexpression(g);
   }
   torch::jit::EliminateDeadCode(g);
-  passes::MarkNodesForFallback(g, true);
+  if (lower_info.forced_fallback_modules.size() > 0) {
+    passes::MarkNodesForFallback(g, true);
+  }
   passes::UnpackHardSwish(g);
   passes::EliminateExceptionOrPassPattern(g);
   passes::ReduceToOperation(g);
@@ -60,12 +62,13 @@ void LowerGraph(std::shared_ptr<torch::jit::Graph>& g, LowerInfo lower_info) {
   LOG_GRAPH(*g);
 }
 
-torch::jit::Module LowerModule(
-    const torch::jit::Module& mod,
-    std::string method_name,
-    std::unordered_set<std::string> forced_fallback_modules) {
-  passes::NotateModuleForFallback(mod, "", method_name, forced_fallback_modules);
-  LOG_GRAPH("After MLF notation pass: " << *mod.get_method(method_name).graph());
+torch::jit::Module LowerModule(const torch::jit::Module& mod, std::string method_name, const LowerInfo& lower_info) {
+  std::unordered_set<std::string> forced_fallback_modules(
+      lower_info.forced_fallback_modules.begin(), lower_info.forced_fallback_modules.end());
+  if (forced_fallback_modules.size() > 0) {
+    passes::NotateModuleForFallback(mod, "", method_name, forced_fallback_modules);
+    LOG_GRAPH("After MLF notation pass: " << *mod.get_method(method_name).graph());
+  }
   auto mod_ = torch::jit::freeze_module(mod);
   LOG_GRAPH("After freeze: " << *mod_.get_method(method_name).graph());
   return mod_;
@@ -77,9 +80,7 @@ std::pair<std::shared_ptr<torch::jit::Graph>, std::vector<torch::jit::IValue>> L
     const LowerInfo& lower_info) {
   LOG_DEBUG(lower_info);
   LOG_GRAPH("Before lowering: " << *mod.get_method(method_name).graph());
-  std::unordered_set<std::string> forced_fallback_modules(
-      lower_info.forced_fallback_modules.begin(), lower_info.forced_fallback_modules.end());
-  auto lowered_mod = lower_info.unfreeze_module ? mod : LowerModule(mod, method_name, forced_fallback_modules);
+  auto lowered_mod = lower_info.unfreeze_module ? mod : LowerModule(mod, method_name, lower_info);
   auto g = lowered_mod.get_method(method_name).graph();
 
   LOG_GRAPH("LibTorch Lowering");

diff --git a/core/lowering/passes/module_fallback.cpp b/core/lowering/passes/module_fallback.cpp
@@ -61,8 +61,29 @@ void NotateModuleForFallback(
     LOG_GRAPH("Notated graph: " << *g);
   }
 
-  for (const auto sub_mod : mod.named_children()) {
-    NotateModuleForFallback(sub_mod.value, sub_mod.name, method_name, forced_fallback_modules);
+  if (mod.named_children().size() > 0) {
+    for (const auto n : nodes) {
+      std::string sub_method_name = "";
+      if (n->kind() == torch::jit::prim::CallMethod) {
+        sub_method_name = n->s(c10::Symbol::attr("name"));
+        auto sub_mod_val = n->input(0);
+        auto sub_mod_src_n = sub_mod_val->node();
+        if (!sub_mod_src_n->hasAttributeS("name")) {
+          LOG_GRAPH("Node: " << util::node_info(sub_mod_src_n) << " manages a module with no name, skipping");
+          break;
+        }
+        auto sub_mod_name = sub_mod_src_n->s(c10::Symbol::attr("name"));
+        for (const auto sub_mod : mod.named_children()) {
+          // Theres probably a way to directly access the module we care about
+          if (sub_mod.name == sub_mod_name) {
+            LOG_GRAPH(
+                "Looking at <module>.<method>() next: " << sub_mod_name << "." << sub_method_name
+                                                        << "() (lowering.passes.NotateModuleForFallback)");
+            NotateModuleForFallback(sub_mod.value, sub_mod.name, sub_method_name, forced_fallback_modules);
+          }
+        }
+      }
+    }
   }
 }
 
@@ -74,23 +95,23 @@ void MarkNodesForFallback(std::shared_ptr<torch::jit::Graph>& g, bool delete_del
     auto n = *it;
     if (!mark.top() && n->kind() == torch::jit::prim::Enter && n->hasAttributeS("compilation_edge")) {
       if (n->s(c10::Symbol::attr("compilation_edge")) == "start") {
-        LOG_DEBUG("Starting to mark new segmented block targeted for torch");
+        LOG_GRAPH("Starting to mark new segmented block targeted for torch");
         mark.push(true);
         if (delete_delims) {
           it.destroyCurrent();
         }
       }
     } else if (mark.top() && n->kind() == torch::jit::prim::Enter && n->hasAttributeS("compilation_edge")) {
       if (n->s(c10::Symbol::attr("compilation_edge")) == "start") {
-        LOG_DEBUG("Found the start of another segmented block targeted for torch while actively marking a block");
+        LOG_GRAPH("Found the start of another segmented block targeted for torch while actively marking a block");
         mark.push(true);
         if (delete_delims) {
           it.destroyCurrent();
         }
       }
     } else if (mark.top() && n->kind() == torch::jit::prim::Exit && n->hasAttributeS("compilation_edge")) {
       if (n->s(c10::Symbol::attr("compilation_edge")) == "end") {
-        LOG_DEBUG("Found the end of segmented block targeted for torch while actively marking a block");
+        LOG_GRAPH("Found the end of segmented block targeted for torch while actively marking a block");
         mark.pop();
         if (delete_delims) {
           it.destroyCurrent();