dynamo benchmarking

ghstack-source-id: 7ea1be2
Pull Request resolved: #265

Author: PaliC

multipy/runtime/CMakeLists.txt

@@ -10,7 +10,7 @@ project(MultipyRuntime)
 # set ABI by default to 0
 
 option(BUILD_CUDA_TESTS "Set to ON in order to build cuda tests. By default we do not" OFF)
-option(GDB_ON "Sets to debug mode (for gdb), defaults to OFF" OFF)
+option(GDB_ON "Sets to debug mode (for gdb), defaults to OFF" ON)
 
 if(GDB_ON)
   set(CMAKE_BUILD_TYPE Debug)
@@ -87,9 +87,9 @@ set(INTERPRETER_TEST_SOURCES
 set(INTERPRETER_TEST_SOURCES_GPU
   ${DEPLOY_DIR}/test_deploy_gpu.cpp
 )
-
-# TODO: Currently tests can only be done when ABI=1 as the testing infrustructure
-# used by ASSERT_TRUE requires ABI=1 in Github actions, we should fix this!
+set(INTERPRETER_TEST_SOURCES_COMPAT
+  ${DEPLOY_DIR}/test_deploy_compat.cpp
+)
 
 add_executable(test_deploy ${INTERPRETER_TEST_SOURCES})
 # target_compile_definitions(test_deploy PUBLIC TEST_CUSTOM_LIBRARY)
@@ -99,6 +99,15 @@ target_link_libraries(test_deploy
 )
 target_include_directories(test_deploy PRIVATE ${CMAKE_SOURCE_DIR}/../..)
 
+add_executable(test_deploy_compat ${INTERPRETER_TEST_SOURCES_COMPAT})
+# target_compile_definitions(test_deploy_compat PUBLIC TEST_CUSTOM_LIBRARY)
+target_include_directories(test_deploy_compat PRIVATE ${PYTORCH_ROOT}/torch)
+target_link_libraries(test_deploy_compat
+  PUBLIC "-Wl,--no-as-needed -rdynamic" gtest dl torch_deploy_interface c10 torch_cpu
+)
+target_include_directories(test_deploy_compat PRIVATE ${CMAKE_SOURCE_DIR}/../..)
+
+
 if(BUILD_CUDA_TESTS)
   LINK_DIRECTORIES("${PYTORCH_ROOT}/torch/lib")
   add_executable(test_deploy_gpu ${INTERPRETER_TEST_SOURCES_GPU})

multipy/runtime/example/benchmark.cpp

@@ -178,6 +178,8 @@ struct Benchmark {
       manager.debugLimitInterpreters(1);
     } else if (strategy == "multi_python") {
       manager.debugLimitInterpreters(n_threads_);
+    } else if (strategy == "dynamo+deploy"){
+      manager.debugLimitInterpreters(n_threads_);
     }
   }
 
@@ -295,6 +297,7 @@ int main(int argc, char* argv[]) {
   cuda = std::string(argv[2]) == "cuda";
   // NOLINTNEXTLINE(cppcoreguidelines-init-variables)
   bool jit_enable = std::string(argv[3]) == "jit";
+  bool inductor_enable = std::string(argv[3]) == "inductor";
   Report::report_header(std::cout);
   torch::deploy::InterpreterManager manager(max_thread);
 
@@ -311,19 +314,25 @@ int main(int argc, char* argv[]) {
       if (n_thread > max_thread) {
         continue;
       }
-      for (std::string strategy : {"one_python", "multi_python", "jit"}) {
+      for (std::string strategy : {"one_python", "multi_python", "jit", "dynamo+deploy"}) {
         if (strategy == "jit") {
           if (!jit_enable) {
             continue;
           }
           if (!exists(model_file + "_jit")) {
             continue;
           }
-        }
-        if (strategy == "one_python") {
+        } else if (strategy == "one_python") {
           Benchmark b(manager, 1, strategy, model_file);
           Report r = b.run();
           r.report(std::cout);
+        } else if (strategy == "dynamo+deploy") {
+          if (!exists(model_file + "_dynamo")) {
+            continue;
+          }
+          Benchmark b(manager, n_thread, strategy, model_file + "_dynamo");
+          Report r = b.run();
+          r.report(std::cout);
         } else {
           Benchmark b(manager, n_thread, strategy, model_file);
           Report r = b.run();

multipy/runtime/example/examples.py

@@ -9,6 +9,7 @@
 import torch
 import torch.nn as nn
 from torch import Tensor
+import torch._dynamo as torchdynamo
 
 
 class Simple(torch.nn.Module):
@@ -128,6 +129,10 @@ def forward(self, x):
 def resnet18():
     return ResNet(BasicBlock, [2, 2, 2, 2])
 
+@torchdynamo.optimize("eager")
+def resnet18_dynamo():
+    return resnet18(); 
+
 
 class BatchedModel(nn.Module):
     def forward(self, input1: Tensor, input2: Tensor) -> Tuple[Tensor, Tensor]:

multipy/runtime/example/generate_examples.py

@@ -21,6 +21,7 @@
         multi_return_metadata,
         MultiReturn,
         resnet18,
+        resnet18_dynamo,
         Simple,
     )
 except ImportError:
@@ -30,6 +31,7 @@
         multi_return_metadata,
         MultiReturn,
         resnet18,
+        resnet18_dynamo,
         Simple,
     )
 
@@ -60,34 +62,39 @@ def save(
     name,
     model,
     model_jit=None,
+    model_dynamo=None,
     eg=None,
     featurestore_meta=None,
     text_in_extra_file=None,
     binary_in_extra_file=None,
 ):
-    with PackageExporter(str(p / name)) as e:
-        e.mock("iopath.**")
-        e.intern("**")
-        e.save_pickle("model", "model.pkl", model)
-        if eg:
-            e.save_pickle("model", "example.pkl", eg)
-        if featurestore_meta:
-            # TODO(whc) can this name come from buck somehow,
-            # so it's consistent with predictor_config_constants::METADATA_FILE_NAME()?
-            e.save_text("extra_files", "metadata.json", featurestore_meta)
-        if text_in_extra_file:
-            e.save_text("extra_files", "text", text_in_extra_file)
-        if binary_in_extra_file:
-            e.save_binary("extra_files", "binary", binary_in_extra_file)
-
+    def package_model(name, model):
+        with PackageExporter(str(p / name)) as e:
+            e.mock("iopath.**")
+            e.intern("**")
+            e.save_pickle("model", "model.pkl", model)
+            if eg:
+                e.save_pickle("model", "example.pkl", eg)
+            if featurestore_meta:
+                # TODO(whc) can this name come from buck somehow,
+                # so it's consistent with predictor_config_constants::METADATA_FILE_NAME()?
+                e.save_text("extra_files", "metadata.json", featurestore_meta)
+            if text_in_extra_file:
+                e.save_text("extra_files", "text", text_in_extra_file)
+            if binary_in_extra_file:
+                e.save_binary("extra_files", "binary", binary_in_extra_file)
+                
+    package_model(name, model)
+    if model_dynamo:
+        package_model(name + "_dynamo", model_dynamo)
     if model_jit:
         model_jit.save(str(p / (name + "_jit")))
+    
 
 
 parser = argparse.ArgumentParser(description="Generate Examples")
 parser.add_argument("--install_dir", help="Root directory for all output files")
 
-
 if __name__ == "__main__":
     args = parser.parse_args()
     if args.install_dir is None:
@@ -98,9 +105,11 @@ def save(
 
     resnet = resnet18()
     resnet.eval()
+    resnet_dynamo = resnet18_dynamo()
     resnet_eg = torch.rand(1, 3, 224, 224)
     resnet_traced = torch.jit.trace(resnet, resnet_eg)
-    save("resnet", resnet, resnet_traced, (resnet_eg,))
+    save("resnet", resnet_dynamo, resnet_traced, resnet_dynamo, (resnet_eg,))
+    # save("resnet", resnet, resnet_traced, resnet_dynamo, (resnet_eg,))
 
     simple = Simple(10, 20)
     save(
@@ -117,6 +126,7 @@ def save(
         "multi_return",
         multi_return,
         torch.jit.script(multi_return),
+        None,
         (torch.rand(10, 20),),
         multi_return_metadata,
     )
@@ -149,4 +159,4 @@ def save(
         e.add_dependency("tensorrt")
         e.mock("iopath.**")
         e.intern("**")
-        e.save_pickle("make_trt_module", "model.pkl", make_trt_module)
+        e.save_pickle("make_trt_module", "model.pkl", make_trt_module)

multipy/runtime/test_compat.py

@@ -8,7 +8,6 @@
 
 import torch
 
-
 class TestCompat(unittest.TestCase):
     def test_torchvision(self):
         import torchvision  # noqa: F401
@@ -21,6 +20,9 @@ def test_pytorch3d(self):
 
     def test_hf_tokenizers(self):
         import tokenizers  # noqa: F401
+resnet_path = "multipy/runtime/example/generated/resnet";
+# resnet_path = "multipy/runtime/example/generated/resnet_dynamo";
+resnet_jit_path = "multipy/runtime/example/generated/resnet_jit";
 
     @unittest.skip("torch.Library is not supported")
     def test_torchdynamo_eager(self):

multipy/runtime/test_compat_dynamo.py

@@ -0,0 +1,28 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import unittest
+
+import torch
+
+
+class TestCompat(unittest.TestCase):
+
+    def test_torchdynamo(self):
+        resnet_path = "multipy/runtime/example/generated/resnet";
+        resnet_jit_path = "multipy/runtime/example/generated/resnet_jit";
+        importer_dynamo = torch.package.PackageImporter(resnet_path)
+        model_dynamo = importer_dynamo.load_pickle("model", "model.pkl")
+        eg_dynamo = importer_dynamo.load_pickle("model", "example.pkl")
+        model_jit = torch.jit.load(resnet_jit_path)
+        model_dynamo.forward(eg_dynamo[0])
+        print(model_dynamo(eg_dynamo[0]), model_jit(eg_dynamo[0]))
+        self.assertTrue(
+            torch.allclose(model_dynamo(eg_dynamo[0]), model_jit(eg_dynamo[0]))
+        )
+
+if __name__ == "__main__":
+    unittest.main()

multipy/runtime/test_deploy_compat.cpp

@@ -0,0 +1,105 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+#include <ATen/Parallel.h>
+#include <gtest/gtest.h>
+#include <libgen.h>
+#include <cstring>
+
+#include <c10/util/irange.h>
+#include <libgen.h>
+#include <multipy/runtime/deploy.h>
+#include <torch/script.h>
+#include <torch/torch.h>
+
+#include <future>
+#include <iostream>
+#include <string>
+
+void compare_torchpy_jit(const char* model_filename, const char* jit_filename) {
+  // Test
+
+  torch::deploy::InterpreterManager m(2);
+  torch::deploy::Package p = m.loadPackage(model_filename);
+  auto model = p.loadPickle("model", "model.pkl");
+  at::IValue eg;
+  {
+    auto I = p.acquireSession();
+    eg = I.self.attr("load_pickle")({"model", "example.pkl"}).toIValue();
+  }
+
+  at::Tensor output = model(eg.toTupleRef().elements()).toTensor();
+
+  // Reference
+  auto ref_model = torch::jit::load(jit_filename);
+  at::Tensor ref_output =
+      ref_model.forward(eg.toTupleRef().elements()).toTensor();
+  ASSERT_TRUE(ref_output.allclose(output, 1e-03, 1e-05));
+}
+const char* resnet_path = "multipy/runtime/example/generated/resnet";
+// const char* resnet_path = "multipy/runtime/example/generated/resnet_dynamo";
+const char* resnet_jit_path = "multipy/runtime/example/generated/resnet_jit";
+
+const char* path(const char* envname, const char* path) {
+  const char* e = getenv(envname);
+  return e ? e : path;
+}
+
+TEST(TorchpyTest, ResNetWithDynamo) {
+  compare_torchpy_jit(
+      path("RESNET", resnet_path),
+      path("RESNET_JIT", resnet_jit_path));
+}
+
+TEST(TorchpyTest, ThreadedResnetModelWithDynamo) {
+  size_t nthreads = 3;
+  torch::deploy::InterpreterManager manager(nthreads);
+
+  torch::deploy::Package p = manager.loadPackage(path("RESNET", resnet_path));
+  auto model = p.loadPickle("model", "model.pkl");
+  auto ref_model = torch::jit::load(path("RESNET_JIT", resnet_jit_path));
+
+  auto input = torch::ones({10, 20});
+
+  std::vector<at::Tensor> outputs;
+
+  std::vector<std::future<at::Tensor>> futures;
+  for (const auto i : c10::irange(nthreads)) {
+    (void)i;
+    futures.push_back(std::async(std::launch::async, [&model]() {
+      auto input = torch::ones({10, 10, 10});
+      for (const auto j : c10::irange(100)) {
+        (void)j;
+        model({input.alias()}).toTensor();
+      }
+      auto result = model({input.alias()}).toTensor();
+      return result;
+    }));
+  }
+  for (const auto i : c10::irange(nthreads)) {
+    outputs.push_back(futures[i].get());
+  }
+
+  // Generate reference
+  auto ref_output = ref_model.forward({input.alias()}).toTensor();
+
+  // Compare all to reference
+  for (const auto i : c10::irange(nthreads)) {
+    ASSERT_TRUE(ref_output.equal(outputs[i]));
+  }
+}
+
+int main(int argc, char* argv[]) {
+  ::testing::InitGoogleTest(&argc, argv);
+  char tempeh[256];
+  getcwd(tempeh, 256);
+  std::cout << "Current working directory: " << tempeh << std::endl;
+  int rc = RUN_ALL_TESTS();
+  char tmp[256];
+  getcwd(tmp, 256);
+  std::cout << "Current working directory: " << tmp << std::endl;
+  return rc;
+}