diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index 48f94a1f05614d4b797562ac67cdb9828fd0456e..37202f869508c283e1b464942cadc0ebe3eef39c 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -79,9 +79,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { BuildStrategy strategy_; }; -std::shared_ptr BuildStrategy::CreatePassesFromStrategy() - const { +std::shared_ptr BuildStrategy::CreatePassesFromStrategy( + bool finalize_strategy) const { + if (is_finalized_) { + return pass_builder_; + } pass_builder_.reset(new ParallelExecutorPassBuilder(*this)); + if (finalize_strategy) { + is_finalized_ = true; + } return pass_builder_; } @@ -95,10 +101,8 @@ std::unique_ptr BuildStrategy::Apply( #else const bool use_cuda) const { #endif - // Create a default one if not initialized by user. - if (!pass_builder_) { - CreatePassesFromStrategy(); - } + // Create a default one if not finalized by user. + CreatePassesFromStrategy(false); std::unique_ptr graph(new ir::Graph(main_program)); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index 6c7b54db8f610aa34cd51dcbc13063290cae3ac0..fc2641dbd48274b43db0b1f156e3e1128f96772e 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -75,12 +75,20 @@ struct BuildStrategy { bool remove_unnecessary_lock_{false}; + // NOTE: + // Before you add new options, think if it's a general strategy that works + // with other strategy. If not, the strategy should be created through + // CreatePassesFromStrategy and the pass can be managed separately. + // User normally doesn't need to call this API. // The PassBuilder allows for more customized insert, remove of passes // from python side. // A new PassBuilder is created based on configs defined above and // passes are owned by the PassBuilder. - std::shared_ptr CreatePassesFromStrategy() const; + std::shared_ptr CreatePassesFromStrategy( + bool finalize_strategy) const; + + bool IsFinalized() const { return is_finalized_; } // Apply the passes built by the pass_builder_. The passes will be // applied to the Program and output an ir::Graph. @@ -97,6 +105,7 @@ struct BuildStrategy { #endif private: + mutable bool is_finalized_ = false; mutable std::shared_ptr pass_builder_; }; diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 592c40cf1ce07e28f3fe4a1ca0b11d249600e5f1..0d059d8aea756fbf85932d4d918ba7f77b772ccd 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -668,9 +668,9 @@ All parameter, weight, gradient are variables in Paddle. [](ir::Pass &self, const std::string &name, const std::string &attr) { self.Set(name, new std::string(attr)); }) - .def("set_int", [](ir::Pass &self, const std::string &name, int val) { - self.Set(name, new int(val)); - }); + .def("set_int", [](ir::Pass &self, const std::string &name, + int val) { self.Set(name, new int(val)); }) + .def("type", &ir::Pass::Type); py::class_> pb( m, "PassBuilder"); @@ -810,6 +810,7 @@ All parameter, weight, gradient are variables in Paddle. "reduce_strategy", [](const BuildStrategy &self) { return self.reduce_; }, [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.reduce_ = strategy; }, R"DOC(The type is STR, there are two reduce strategies in ParallelExecutor, @@ -823,6 +824,7 @@ All parameter, weight, gradient are variables in Paddle. [](const BuildStrategy &self) { return self.gradient_scale_; }, [](BuildStrategy &self, BuildStrategy::GradientScaleStrategy strategy) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.gradient_scale_ = strategy; }, R"DOC(The type is STR, there are three ways of defining :math:`loss@grad` in @@ -834,6 +836,7 @@ All parameter, weight, gradient are variables in Paddle. "debug_graphviz_path", [](const BuildStrategy &self) { return self.debug_graphviz_path_; }, [](BuildStrategy &self, const std::string &path) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.debug_graphviz_path_ = path; }, R"DOC(The type is STR, debug_graphviz_path indicate the path that @@ -843,6 +846,7 @@ All parameter, weight, gradient are variables in Paddle. "enable_data_balance", [](const BuildStrategy &self) { return self.enable_data_balance_; }, [](BuildStrategy &self, bool b) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.enable_data_balance_ = b; }) // FIXME(chengudo): enable_data_balance seems not important .def_property( @@ -851,6 +855,7 @@ All parameter, weight, gradient are variables in Paddle. return self.enable_sequential_execution_; }, [](BuildStrategy &self, bool b) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.enable_sequential_execution_ = b; }, R"DOC(The type is BOOL. If set True, the execution order of ops would be the same as what is in the program. Default False.)DOC") @@ -860,6 +865,7 @@ All parameter, weight, gradient are variables in Paddle. return self.remove_unnecessary_lock_; }, [](BuildStrategy &self, bool b) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.remove_unnecessary_lock_ = b; }, R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default False.)DOC") @@ -869,15 +875,19 @@ All parameter, weight, gradient are variables in Paddle. return self.fuse_elewise_add_act_ops_; }, [](BuildStrategy &self, bool b) { + PADDLE_ENFORCE(!self.IsFinalized(), "BuildStrategy is finlaized."); self.fuse_elewise_add_act_ops_ = b; }, R"DOC(The type is BOOL, fuse_elewise_add_act_ops indicate whether to fuse elementwise_add_op and activation_op, it may make the execution faster. Default False)DOC") - .def("_create_passes_from_strategy", + .def("_finalize_strategy_and_create_passes", [](BuildStrategy &self) -> std::shared_ptr { - return self.CreatePassesFromStrategy(); - }); + return self.CreatePassesFromStrategy(true); + }, + R"DOC(Allow user to customized passes. Normally model-specific + optimization passes should be defined in this way. BuildStrategy + cannot be updated after being finalized.)DOC"); pe.def(py::init &, const std::unordered_set &, diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py index 4b8a215190a90c974a9ecc8658d044c59b80c989..97e7ee6229f081ff67ca3e2aedcad0a2e3d9cabf 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_base.py +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -105,7 +105,7 @@ class TestDistRunnerBase(object): build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce if args.batch_merge_repeat > 1: - pass_builder = build_stra._create_passes_from_strategy() + pass_builder = build_stra._finalize_strategy_and_create_passes() mypass = pass_builder.insert_pass( len(pass_builder.all_passes()) - 2, "multi_batch_merge_pass") mypass.set_int("num_repeats", args.batch_merge_repeat) diff --git a/python/paddle/fluid/tests/unittests/test_pass_builder.py b/python/paddle/fluid/tests/unittests/test_pass_builder.py index 288c5f6a1f6b1760ca40c0c653e4c0726b799519..5a3ec8ff0180281babeaa006133b3ff9dc6d8083 100644 --- a/python/paddle/fluid/tests/unittests/test_pass_builder.py +++ b/python/paddle/fluid/tests/unittests/test_pass_builder.py @@ -94,7 +94,12 @@ class TestPassBuilder(unittest.TestCase): def test_parallel_testing_with_new_strategy(self): build_strategy = fluid.BuildStrategy() - pass_builder = build_strategy._create_passes_from_strategy() + self.assertFalse(build_strategy.fuse_elewise_add_act_ops) + build_strategy.fuse_elewise_add_act_ops = True + pass_builder = build_strategy._finalize_strategy_and_create_passes() + self.assertTrue("fuse_elewise_add_act_pass" in + [p.type() for p in pass_builder.all_passes()]) + origin_len = len(pass_builder.all_passes()) viz_pass = pass_builder.append_pass("graph_viz_pass")