add cuda_graph for mts_gpu_benchmark (#1012)

Summary: Pull Request resolved: #1012 Add cuda_graph enablement for AIT. Also leave the hook for AOTI GPU trace after enabling cudagraph: https://fburl.com/perfdoctor/yybf0z60 log information for verification I0710 17:58:44.425707 2013974 AITModelImpl.cpp:148] AITModelImpl: loading .so lib /tmp/benchmark_529602.1720659401/_run_on_acc_0/_run_on_acc_0-ait_engine.so I0710 17:58:44.425731 2013974 AITModelImpl.cpp:149] AITModelImpl: num_runtimes: 1,use_cuda_graph: 1 Reviewed By: guowentian Differential Revision: D59617284
facebookincubator · Jul 12, 2024 · 9af8be6 · 9af8be6
1 parent a1769f2
commit 9af8be6
Show file tree

Hide file tree

Showing 7 changed files with 15 additions and 2 deletions.
diff --git a/fx2ait/fx2ait/ait_splitter.py b/fx2ait/fx2ait/ait_splitter.py
@@ -218,6 +218,7 @@ def _lower_model_to_backend(
                 torch.float16,
                 torch.float,
                 1,  # num_runtimes
+                False,
             ),
             interpreter_result,
         )

diff --git a/fx2ait/fx2ait/csrc/AITModel.cpp b/fx2ait/fx2ait/csrc/AITModel.cpp
@@ -64,7 +64,8 @@ static auto registerAITModel =
              std::vector<std::string>,
              std::optional<at::ScalarType>,
              std::optional<at::ScalarType>,
-             int64_t>())
+             int64_t,
+             bool>())
         .def("forward", &AITModel::forward)
         .def("profile", &AITModel::profile)
         .def("get_library_path", &AITModel::libraryPath)

diff --git a/fx2ait/fx2ait/csrc/AITModelImpl.cpp b/fx2ait/fx2ait/csrc/AITModelImpl.cpp
@@ -145,7 +145,9 @@ AITModelImpl::AITModelImpl(
       floating_point_input_dtype_(input_dtype),
       floating_point_output_dtype_(output_dtype),
       use_cuda_graph_(use_cuda_graph) {
-  LOG(INFO) << "Loading .so lib " << model_path;
+  LOG(INFO) << "AITModelImpl: loading .so lib " << model_path;
+  LOG(INFO) << "AITModelImpl: num_runtimes: " << num_runtimes
+            << ",use_cuda_graph: " << use_cuda_graph;
   TORCH_CHECK(handle_, "could not dlopen ", model_path, ": ", dlerror());
   TORCH_CHECK(num_runtimes > 0, "num_runtimes must be positive");
 

diff --git a/fx2ait/fx2ait/test/test_fx2ait.py b/fx2ait/fx2ait/test/test_fx2ait.py
@@ -68,6 +68,7 @@ def _test_fx2ait_impl(self, test_serialization=False, test_cuda_graph=False):
                 torch.float16,
                 torch.float16,
                 1,  # num_runtimes
+                False,
             )
         )
         ait_mod.engine.use_cuda_graph = test_cuda_graph
@@ -140,6 +141,7 @@ def forward(self, a, b, c, d):
                 torch.float16,
                 torch.float16,
                 1,  # num_runtimes
+                False,
             ),
             interp_result,
         )

diff --git a/fx2ait/fx2ait/tools/ait_minimizer.py b/fx2ait/fx2ait/tools/ait_minimizer.py
@@ -42,6 +42,7 @@ def lower_mod_default(
             torch.float16,
             torch.float16,
             1,  # num_runtimes
+            False,
         ),
         interpreter_result,
     )

diff --git a/fx2ait/fx2ait/tools/common_aten2ait.py b/fx2ait/fx2ait/tools/common_aten2ait.py
@@ -163,6 +163,7 @@ def run_test(
                 torch.float16,
                 torch.float,
                 1,  #  num_runtimes
+                False,
             ),
             interp_result,
         )
@@ -256,6 +257,7 @@ def run_test_with_dynamic_shape(
                 torch.float16,
                 torch.float,
                 1,  #  num_runtimes
+                False,
             ),
             interp_result,
         )
@@ -375,6 +377,7 @@ def benchmark(f, args):
                     torch.float16,
                     torch.float,
                     1,  #  num_runtimes
+                    False,
                 ),
                 interp_result,
             )

diff --git a/fx2ait/fx2ait/tools/common_fx2ait.py b/fx2ait/fx2ait/tools/common_fx2ait.py
@@ -199,6 +199,7 @@ def run_test(
                         torch_dtype,
                         torch.float,
                         1,  #  num_runtimes
+                        False,
                     ),
                     interp_result,
                 )
@@ -329,6 +330,7 @@ def run_test_with_dynamic_shape(
                         torch.float16,
                         torch.float,
                         1,  #  num_runtimes
+                        False,
                     ),
                     interp_result,
                 )
@@ -467,6 +469,7 @@ def benchmark(f, args):
                 torch.float16,
                 torch.float,
                 1,  #  num_runtimes
+                False,
             ),
             interp_result,
         )