Support calling custom method names via MODULE_METHOD_NAME (fixes triton-inference-server/server#5209)

Signed-off-by: Christian Bruckdorfer <[email protected]>

Author: iceychris

README.md

@@ -176,6 +176,20 @@ key: "ENABLE_CACHE_CLEANING"
 }
 ```
 
+* `MODULE_METHOD_NAME`: String flag to specify which method on the PyTorch model is being called.
+Default value is `forward`.
+
+The section of model config file specifying this parameter will look like:
+
+```
+parameters: {
+key: "MODULE_METHOD_NAME"
+    value: {
+    string_value:"custom_method"
+    }
+}
+```
+
 * Additional Optimizations: Three additional boolean parameters are available to disable
 certain Torch optimizations that can sometimes cause latency regressions in models with
 complex execution modes and dynamic shapes. If not specified, all are enabled by default.

src/libtorch.cc

@@ -56,6 +56,9 @@
 #include <cuda_runtime_api.h>
 #endif  // TRITON_ENABLE_GPU
 
+// Default forward method to call on PyTorch modules
+const std::string DEFAULT_MODULE_METHOD_NAME = "forward";
+
 //
 // PyTorch C++ (LibTorch) Backend that implements the TRITONBACKEND API.
 //
@@ -105,6 +108,7 @@ class ModelState : public BackendModel {
   {
     return model_outputs_;
   }
+  const std::string& ModuleMethodName() { return module_method_name_; }
 
  private:
   ModelState(TRITONBACKEND_Model* triton_model);
@@ -147,6 +151,10 @@ class ModelState : public BackendModel {
   // is specified both in the output section and state section, it indicates
   // that the backend must return the output state to the client too.
   std::map<std::string, std::pair<int64_t, int64_t>> model_outputs_;
+
+  // Method to call on PyTorch Module.
+  // Defaults to DEFAULT_MODULE_METHOD_NAME.
+  std::string module_method_name_;
 };
 
 TRITONSERVER_Error*
@@ -224,7 +232,8 @@ ModelState::ModelState(TRITONBACKEND_Model* triton_model)
       enable_inference_mode_(true), enable_cache_cleaning_(false),
       enable_weight_sharing_(false), enable_tensor_fuser_pair_({false, true}),
       enable_jit_profiling_pair_({false, true}),
-      enable_jit_executor_pair_({false, true})
+      enable_jit_executor_pair_({false, true}),
+      module_method_name_(DEFAULT_MODULE_METHOD_NAME)
 {
 }
 
@@ -465,6 +474,30 @@ ModelState::ParseParameters()
            " for model instance '" + Name() + "'")
               .c_str());
     }
+
+    // If 'MODULE_METHOD_NAME' is not present in 'parameters' then
+    // 'module_method_name_' is set to 'DEFAULT_MODULE_METHOD_NAME' ('forward').
+    std::string module_method_name = DEFAULT_MODULE_METHOD_NAME;
+    err = GetParameterValue(params, "MODULE_METHOD_NAME", &module_method_name);
+    if (err != nullptr) {
+      if (TRITONSERVER_ErrorCode(err) != TRITONSERVER_ERROR_NOT_FOUND) {
+        return err;
+      } else {
+        LOG_MESSAGE(
+            TRITONSERVER_LOG_INFO,
+            (std::string("module_method_name is not specified") +
+             " for model instance '" + Name() + "'")
+                .c_str());
+        TRITONSERVER_ErrorDelete(err);
+      }
+    } else {
+      module_method_name_ = module_method_name;
+      LOG_MESSAGE(
+          TRITONSERVER_LOG_INFO,
+          (std::string("module_method_name is ") + module_method_name_ +
+           " for model instance '" + Name() + "'")
+              .c_str());
+    }
   }
 
   return nullptr;
@@ -886,7 +919,19 @@ ModelInstanceState::ValidateInputs(const size_t expected_input_cnt)
   // configuration specifies only those.
   std::vector<std::string> allowed_inputs;
 
-  const torch::jit::Method& method = torch_model_->get_method("forward");
+  // First check if method exists in the model and throw an error if absent
+  const auto methodNameToExecute = model_state_->ModuleMethodName();
+  const auto optionalMethodHandle = torch_model_->find_method(methodNameToExecute);
+  if (!optionalMethodHandle.has_value()) {
+    return TRITONSERVER_ErrorNew(
+        TRITONSERVER_ERROR_INVALID_ARG,
+        (std::string("unable to find method '") +
+         methodNameToExecute + "' in model '" + model_path_ + "'")
+            .c_str());
+  }
+
+  // Get the method schema and validate the inputs
+  const torch::jit::Method& method = optionalMethodHandle.value();
   const auto& schema = method.function().getSchema();
   const std::vector<c10::Argument>& arguments = schema.arguments();
 
@@ -1529,18 +1574,24 @@ ModelInstanceState::Execute(
     torch::NoGradGuard no_grad;
 
     // If input is a dictionary, prepare dictionary from 'input_tensors'.
+    std::string module_method_name = model_state_->ModuleMethodName();
+    std::vector<c10::IValue> inputs;
     if (is_dict_input_) {
-      torch::Dict<std::string, torch::Tensor> input_dict;
+      c10::Dict<std::string, at::Tensor> dict;
       for (auto& input_index : input_index_map_) {
         torch::jit::IValue ival = (*input_tensors)[input_index.second];
-        input_dict.insert(input_index.first, ival.toTensor());
+        dict.insert(input_index.first, ival.toTensor());
       }
-      std::vector<torch::jit::IValue> input_dict_ivalue = {input_dict};
-      model_outputs_ = torch_model_->forward(input_dict_ivalue);
+      inputs.push_back(dict);
     } else {
-      model_outputs_ = torch_model_->forward(*input_tensors);
+      for (auto& input_tensor : *input_tensors) {
+        inputs.push_back(input_tensor.toTensor());
+      }
     }
 
+    // Actually run the method on the model.
+    model_outputs_ = torch_model_->get_method(module_method_name)(inputs);
+
     if (model_outputs_.isTuple()) {
       auto model_outputs_tuple = model_outputs_.toTuple();
       size_t op_index = 0;