From 0dd6a44acb30cc708db22faf56f24c11cee96831 Mon Sep 17 00:00:00 2001 From: Aurelius84 Date: Tue, 27 Jul 2021 10:51:27 +0800 Subject: [PATCH] Revert "Revert "[Dy2Stat] Refactor ExecutorCache logic and pre-support BuildStrategy for pass (#34181)" (#34348)" (#34384) This reverts commit 577fdde579bc9caf06cafe2318b4e673e900ba98. --- paddle/fluid/framework/executor_cache.cc | 60 ++++--- paddle/fluid/framework/executor_cache.h | 153 +++++++----------- paddle/fluid/operators/run_program_op.cc | 4 + paddle/fluid/operators/run_program_op.h | 54 ++++--- paddle/fluid/pybind/pybind.cc | 6 + python/paddle/fluid/core.py | 2 + .../dygraph_to_static/partial_program.py | 16 +- python/paddle/fluid/dygraph/io.py | 4 +- python/paddle/fluid/layers/utils.py | 9 ++ .../dygraph_to_static/CMakeLists.txt | 3 +- .../tests/unittests/test_run_program_op.py | 3 + 11 files changed, 166 insertions(+), 148 deletions(-) diff --git a/paddle/fluid/framework/executor_cache.cc b/paddle/fluid/framework/executor_cache.cc index 17eac8a7228..0ab4bd5a12b 100644 --- a/paddle/fluid/framework/executor_cache.cc +++ b/paddle/fluid/framework/executor_cache.cc @@ -13,6 +13,7 @@ // limitations under the License. #include "paddle/fluid/framework/executor_cache.h" +#include "paddle/fluid/framework/op_info.h" namespace paddle { namespace framework { @@ -25,11 +26,11 @@ namespace framework { namespace details { -static ExecutionStrategy GetExecutionStrategy( - const ExecutorInfoCache::CacheKey &cache_key) { +static ExecutionStrategy GetExecutionStrategy(const platform::Place &place) { framework::ExecutionStrategy execution_strategy; - switch (cache_key.device_type_) { + auto device_type = platform::Place2DeviceType(place); + switch (device_type) { case platform::DeviceType::CPU: { execution_strategy.num_threads_ = 2; break; @@ -46,9 +47,9 @@ static ExecutionStrategy GetExecutionStrategy( } default: PADDLE_THROW(platform::errors::Unavailable("Unsupported Device type %d.", - cache_key.device_type_)); + device_type)); } - execution_strategy.use_device_ = cache_key.device_type_; + execution_strategy.use_device_ = device_type; return execution_strategy; } @@ -136,20 +137,14 @@ ExecutorInfoCache &ExecutorInfoCache::Instance() { return g_exe_cache_info_map; } -void ExecutorInfoCache::Finalize() { - // NOTE(Aurelius84): DO NOT perform finalize in destructor - // to avoid problems caused by destructor order of static - // object. - info_map_.clear(); -} - -CacheInfo GetExecutorInfoFromCache(const ExecutorInfoCache::CacheKey &cache_key, +CacheInfo GetExecutorInfoFromCache(const ProgramDesc &program_desc, + const platform::Place &place, + int64_t start_op_index, int64_t end_op_index, + bool is_grad, int64_t program_id, framework::Scope *scope) { auto &cached_exe_info = framework::ExecutorInfoCache::Instance(); - if (!cached_exe_info.Has(cache_key)) { - VLOG(1) << "create exe_info for " << cache_key.DebugString(); - + if (!cached_exe_info.Has(program_id, is_grad)) { // TODO(Aurelius84): Consider to use LRU algorithm to replace this. if (cached_exe_info.Size() > 4u /* max_cached_size*/) { VLOG(2) << "The cached info size has exceeded max_cached_size: 4, clear " @@ -157,29 +152,28 @@ CacheInfo GetExecutorInfoFromCache(const ExecutorInfoCache::CacheKey &cache_key, cached_exe_info.Finalize(); } - framework::BuildStrategy build_strategy; - auto execution_strategy = details::GetExecutionStrategy(cache_key); + VLOG(1) << "create exe_info for " << program_id << " is_grad: " << is_grad; + auto execution_strategy = details::GetExecutionStrategy(place); + auto &build_strategy = cached_exe_info.GetBuildStrategy(program_id); + // 2. Construct Graph and ParallelExecutor. auto graph = std::make_shared( - *cache_key.program_desc_, cache_key.start_op_index_, - cache_key.end_op_index_); + program_desc, start_op_index, end_op_index); auto parallel_executor = std::make_shared( - cache_key.place_, scope, execution_strategy, build_strategy, - graph.get()); + place, scope, execution_strategy, build_strategy, graph.get()); parallel_executor->PrepareVariables(scope); - framework::ExecutorInfoCache::ValueType cache_val = {parallel_executor, - graph}; - cached_exe_info.Insert(cache_key, cache_val); - - bool is_new_created = true; - return std::make_pair(parallel_executor, is_new_created); + // 3. Insert value into cached map. + auto &cached_value = cached_exe_info.GetMutable(program_id, is_grad); + cached_value.executor_ = parallel_executor; + cached_value.graph_ = std::move(graph); + return std::make_pair(parallel_executor, /*is_new_created=*/true); } else { - VLOG(1) << "get exe_info from cache by: " << cache_key.DebugString(); - bool is_new_created = false; - auto cache_val = cached_exe_info.GetMutable(cache_key); - auto parallel_executor = cache_val.first; + VLOG(1) << "get exe_info from cache by: " << program_id + << " is_grad: " << is_grad; + auto &cached_value = cached_exe_info.GetMutable(program_id, is_grad); + auto ¶llel_executor = cached_value.executor_; // update op_handle scope_map in pe->executor_->Graph std::unordered_map scope_map = { {parallel_executor->GetLocalScopes().front(), scope}}; @@ -187,7 +181,7 @@ CacheInfo GetExecutorInfoFromCache(const ExecutorInfoCache::CacheKey &cache_key, // need to recreate tmp variables in new scope parallel_executor->PrepareVariables(scope); - return std::make_pair(parallel_executor, is_new_created); + return std::make_pair(parallel_executor, /*is_new_created=*/false); } } diff --git a/paddle/fluid/framework/executor_cache.h b/paddle/fluid/framework/executor_cache.h index 3e08579ba42..8207b56fc04 100644 --- a/paddle/fluid/framework/executor_cache.h +++ b/paddle/fluid/framework/executor_cache.h @@ -45,121 +45,92 @@ void ParseSafeEagerDeletionSkipVars( std::vector* skip_eager_delete_vars); } // namespace details -class ExecutorInfoCache { + +class ExecutorInfo { public: - struct CacheKey { - CacheKey(const ProgramDesc* program_desc, const platform::Place& place, - int64_t start_op_index, int64_t end_op_index, bool is_grad) - : program_desc_(program_desc), - place_(place), - start_op_index_(start_op_index), - end_op_index_(end_op_index), - is_grad_(is_grad) { - device_type_ = platform::Place2DeviceType(place); - PADDLE_ENFORCE_NOT_NULL(program_desc_, - "program_desc should not be null."); - } - - std::string DebugString() const { - std::stringstream ss; - - ss << "\n CacheKey(program_desc: " << program_desc_; - ss << ", start_op_index: " << start_op_index_; - ss << ", end_op_index: " << end_op_index_; - ss << ", is_grad: " << is_grad_; - ss << ", device_type: " << device_type_ << ")"; - - return ss.str(); - } - - const ProgramDesc* program_desc_; - platform::Place place_; - int64_t start_op_index_; - int64_t end_op_index_; - bool is_grad_; - platform::DeviceType device_type_; - }; + struct CacheValue { + std::shared_ptr executor_{nullptr}; + std::shared_ptr graph_{nullptr}; - using KeyType = size_t; - using ValueType = - std::pair, std::shared_ptr>; - - struct KeyHasher { - size_t operator()(const CacheKey& key) const noexcept { - size_t seed = 10; - auto* prog_desc = key.program_desc_; - /* - * Note(Aurelius84): DO NOT use only ProgramDesc* to calculate hash value - * because a new program will hold same pointer address after an older - * program is destructed with a small probability. Add op size while - * hashing because program may contains at least one block. - */ - hash_combine(&seed, prog_desc); - for (size_t i = 0; i < prog_desc->Size(); ++i) { - hash_combine(&seed, &prog_desc->Block(i)); - hash_combine(&seed, prog_desc->Block(i).OpSize()); - } - hash_combine(&seed, static_cast(key.device_type_)); - hash_combine(&seed, key.start_op_index_); - hash_combine(&seed, key.end_op_index_); - hash_combine(&seed, key.is_grad_); - VLOG(3) << "hash value is : " << seed - << " of key: " << key.DebugString(); - return seed; - } - - template - void hash_combine(size_t* seed, const T& val) const { - std::hash hasher; - (*seed) ^= hasher(val) + 0x9e3779b9 + ((*seed) << 6) + ((*seed >> 2)); - } + std::vector skip_eager_delete_vars_; }; + bool IsAvailable(bool is_grad) { + const auto& executor = + is_grad ? backward_info_.executor_ : forward_info_.executor_; + return executor != nullptr; + } + + CacheValue& GetMutable(bool is_grad) { + return is_grad ? backward_info_ : forward_info_; + } + + private: + CacheValue forward_info_; + CacheValue backward_info_; +}; + +class ExecutorInfoCache { + public: static ExecutorInfoCache& Instance(); - ValueType GetMutable(const CacheKey& key) { - auto key_val = key_hash_func_(key); + const BuildStrategy& GetBuildStrategy(int64_t program_id) { + // If not found, insert build_strategy with default value. + return strategy_map_[program_id]; + } + + void SetBuildStrategy(int64_t program_id, + const BuildStrategy& build_strategy) { PADDLE_ENFORCE_EQ( - Has(key_val), true, - platform::errors::NotFound("%s doesn't exist in ExecutorInfoCache", - key.DebugString())); - return info_map_[key_val]; + strategy_map_.count(program_id), 0, + platform::errors::PreconditionNotMet( + "program_id: %s already exist in ExecutorInfoCache", program_id)); + strategy_map_[program_id] = build_strategy; } - bool Has(const CacheKey& key) const { - auto key_val = key_hash_func_(key); - return Has(key_val); + bool Has(int64_t program_id, bool is_grad) { + return info_map_.find(program_id) != info_map_.end() && + info_map_[program_id].IsAvailable(is_grad); } - bool Has(const KeyType& key) const { - return info_map_.find(key) != info_map_.end(); + ExecutorInfo::CacheValue& GetMutable(int64_t program_id, bool is_grad) { + return info_map_[program_id].GetMutable(is_grad); } - void Insert(const CacheKey& key, ValueType value) { - auto key_val = key_hash_func_(key); - PADDLE_ENFORCE_EQ( - Has(key_val), false, - platform::errors::NotFound("%s has existed in ExecutorInfoCache", - key.DebugString())); - info_map_.insert({key_val, value}); + void UpdateSkipEagerDeleteVars(int64_t program_id, bool is_grad, + const std::vector& skip_vars) { + auto& cached_value = GetMutable(program_id, is_grad); + cached_value.skip_eager_delete_vars_ = std::move(skip_vars); + } + + std::vector& SkipEagerDeleteVars(int64_t program_id, + bool is_grad) { + auto& cached_value = GetMutable(program_id, is_grad); + return cached_value.skip_eager_delete_vars_; } size_t Size() const { return info_map_.size(); } - void Finalize(); + void Finalize() { + // NOTE(Aurelius84): DO NOT perform finalize in destructor + // to avoid problems caused by destructor order of static + // object. + info_map_.clear(); + strategy_map_.clear(); + } private: - ExecutorInfoCache() = default; - DISABLE_COPY_AND_ASSIGN(ExecutorInfoCache); - - KeyHasher key_hash_func_; - std::unordered_map info_map_; + std::unordered_map info_map_; + std::unordered_map strategy_map_; }; using CacheInfo = std::pair, bool /*is_new_created*/>; -CacheInfo GetExecutorInfoFromCache(const ExecutorInfoCache::CacheKey& cache_key, +CacheInfo GetExecutorInfoFromCache(const ProgramDesc& program_desc, + const platform::Place& place, + int64_t start_op_index, int64_t end_op_index, + bool is_grad, int64_t program_id, framework::Scope* scope); } // namespace framework diff --git a/paddle/fluid/operators/run_program_op.cc b/paddle/fluid/operators/run_program_op.cc index 69b2c5b7380..80758e1718b 100644 --- a/paddle/fluid/operators/run_program_op.cc +++ b/paddle/fluid/operators/run_program_op.cc @@ -103,6 +103,10 @@ class RunProgramOpMaker : public framework::OpProtoAndCheckerMaker { "(bool, default false) Set to true for inference only, false " "for training.") .SetDefault(false); + AddAttr( + "program_id", + "(int64_t)" + "The unique hash id used as cache key for ExecutorInfoCache."); AddComment(R"DOC( RunProgram operator. diff --git a/paddle/fluid/operators/run_program_op.h b/paddle/fluid/operators/run_program_op.h index 69345dc8a08..ac352876e78 100644 --- a/paddle/fluid/operators/run_program_op.h +++ b/paddle/fluid/operators/run_program_op.h @@ -188,6 +188,7 @@ class RunProgramOpKernel : public framework::OpKernel { auto start_op_index = ctx.Attr("start_op_index"); auto end_op_index = ctx.Attr("end_op_index"); auto is_test = ctx.Attr("is_test"); + auto program_id = ctx.Attr("program_id"); // NOTE(chenweihang): In order not to add new variable type, use vector // here. Originally, here can use scope directly. @@ -215,23 +216,27 @@ class RunProgramOpKernel : public framework::OpKernel { if (end_op_index > start_op_index) { auto *program = ctx.Attr("global_block")->Program(); - auto cache_key = framework::ExecutorInfoCache::CacheKey( - program, ctx.GetPlace(), start_op_index, end_op_index, - /*is_grad=*/false); - auto cache_info = framework::GetExecutorInfoFromCache(cache_key, &scope); + auto cache_info = framework::GetExecutorInfoFromCache( + *program, ctx.GetPlace(), start_op_index, end_op_index, + /*is_grad=*/false, program_id, &scope); auto ¶llel_executor = cache_info.first; + // all out_vars are skip_eager_var + auto &skip_eager_delete_vars = + framework::ExecutorInfoCache::Instance().SkipEagerDeleteVars( + program_id, false); if (cache_info.second /*is_new_created*/) { parallel_executor->SkipMemoryReuse(/*scope_idx=*/0, input_var_names); + skip_eager_delete_vars.insert(skip_eager_delete_vars.end(), + output_var_names.begin(), + output_var_names.end()); + skip_eager_delete_vars.insert(skip_eager_delete_vars.end(), + dout_var_names.begin(), + dout_var_names.end()); + framework::details::ParseSafeEagerDeletionSkipVars( + *program, end_op_index, output_var_names, &skip_eager_delete_vars); } // Step 3. run ops - // all out_vars are skip_eager_var - std::vector skip_eager_delete_vars(output_var_names); - skip_eager_delete_vars.insert(skip_eager_delete_vars.end(), - dout_var_names.begin(), - dout_var_names.end()); - framework::details::ParseSafeEagerDeletionSkipVars( - *program, end_op_index, output_var_names, &skip_eager_delete_vars); parallel_executor->RunWithoutFetch(skip_eager_delete_vars); } // Step 4. Get Output @@ -280,6 +285,7 @@ class RunProgramGradOpKernel : public framework::OpKernel { auto *block = ctx.Attr("global_block"); auto orig_end_op_index = ctx.Attr("end_op_index"); + auto program_id = ctx.Attr("program_id"); // NOTE: skip `shape` and `fill_constant` op created by // fluid.backward.gradients, one forward output will generate one `shape` // and `fill_constant` @@ -305,24 +311,30 @@ class RunProgramGradOpKernel : public framework::OpKernel { if (end_op_index > start_op_index) { // Step 2. prepare executor and scope auto *program = ctx.Attr("global_block")->Program(); - auto cache_key = framework::ExecutorInfoCache::CacheKey( - program, ctx.GetPlace(), start_op_index, end_op_index, - /*is_grad*/ true); - auto cache_info = framework::GetExecutorInfoFromCache(cache_key, &scope); + auto cache_info = framework::GetExecutorInfoFromCache( + *program, ctx.GetPlace(), start_op_index, end_op_index, + /*is_grad*/ true, program_id, &scope); auto ¶llel_executor = cache_info.first; - parallel_executor->SkipMemoryReuse(/*scope_idx=*/0, - output_grad_var_names); + auto &skip_eager_delete_vars = + framework::ExecutorInfoCache::Instance().SkipEagerDeleteVars( + program_id, true); + if (cache_info.second /*is_new_created*/) { + parallel_executor->SkipMemoryReuse(/*scope_idx=*/0, + output_grad_var_names); + + skip_eager_delete_vars.insert(skip_eager_delete_vars.end(), + input_grad_var_names.begin(), + input_grad_var_names.end()); + framework::details::AppendSkipDeletionVars(param_grad_names, + &skip_eager_delete_vars); + } details::ShareVarsIntoScope(output_grad_vars, output_grad_var_names, &scope); // Debug info: scope info when run end VLOG(3) << framework::GenScopeTreeDebugInfo(out_scope_vec->front()); - std::vector skip_eager_delete_vars(input_grad_var_names); - framework::details::AppendSkipDeletionVars(param_grad_names, - &skip_eager_delete_vars); - // Step 3. run ops parallel_executor->RunWithoutFetch( /*skip_eager_delete_vars=*/skip_eager_delete_vars); diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index d646e06d8a4..f4976670190 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -3011,6 +3011,12 @@ All parameter, weight, gradient are variables in Paddle. optimization passes should be defined in this way. BuildStrategy cannot be updated after being finalized.)DOC"); + m.def("_set_cached_executor_build_strategy", + [](int64_t program_id, const BuildStrategy &build_strategy) { + auto &cached_exe_info = framework::ExecutorInfoCache::Instance(); + cached_exe_info.SetBuildStrategy(program_id, build_strategy); + }); + pe.def(py::init &, const std::vector &, const std::string &, Scope *, std::vector &, const ExecutionStrategy &, diff --git a/python/paddle/fluid/core.py b/python/paddle/fluid/core.py index 1f26b3d0bf0..befcd0901dc 100644 --- a/python/paddle/fluid/core.py +++ b/python/paddle/fluid/core.py @@ -273,6 +273,7 @@ if avx_supported(): from .core_avx import _cuda_synchronize from .core_avx import _is_compiled_with_heterps from .core_avx import _promote_types_if_complex_exists + from .core_avx import _set_cached_executor_build_strategy from .core_avx import _device_synchronize from .core_avx import _get_current_stream if sys.platform != 'win32': @@ -324,6 +325,7 @@ if load_noavx: from .core_noavx import _cuda_synchronize from .core_noavx import _is_compiled_with_heterps from .core_noavx import _promote_types_if_complex_exists + from .core_noavx import _set_cached_executor_build_strategy from .core_noavx import _device_synchronize from .core_noavx import _get_current_stream if sys.platform != 'win32': diff --git a/python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py b/python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py index 8d155d20601..9e6fdd8ed92 100644 --- a/python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py +++ b/python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py @@ -24,6 +24,8 @@ from paddle.fluid.dygraph.dygraph_to_static import logging_utils from paddle.fluid.dygraph.dygraph_to_static.return_transformer import RETURN_NO_VALUE_MAGIC_NUM from paddle.fluid.layers.utils import flatten from paddle.fluid.layers.utils import pack_sequence_as +from paddle.fluid.layers.utils import _hash_with_id +from paddle.fluid.compiler import BuildStrategy import paddle.compat as cpt from paddle import _C_ops @@ -162,6 +164,14 @@ class PartialProgramLayer: return train_program + @LazyInitialized + def _infer_program_id(self): + return _hash_with_id(self._infer_program, self) + + @LazyInitialized + def _train_program_id(self): + return _hash_with_id(self._train_program, self) + def _verify_program(self, main_program): """ Verify that the program parameter is initialized, prune some unused params, @@ -228,7 +238,7 @@ class PartialProgramLayer: attrs = ('global_block', self.program.desc.block(0), 'start_op_index', 0, 'end_op_index', self._infer_program.desc.block(0).op_size(), - 'is_test', not self.training) + 'is_test', not self.training, 'program_id', self.program_id) _C_ops.run_program( self._valid_vars(in_vars), self._valid_vars(self._params), @@ -242,6 +252,10 @@ class PartialProgramLayer: def program(self): return self._train_program if self.training else self._infer_program + @property + def program_id(self): + return self._train_program_id if self.training else self._infer_program_id + def _prepare(self, inputs): """ Prepare inputs, outputs, attrs. diff --git a/python/paddle/fluid/dygraph/io.py b/python/paddle/fluid/dygraph/io.py index d5ad3a88e8c..2318a08462d 100644 --- a/python/paddle/fluid/dygraph/io.py +++ b/python/paddle/fluid/dygraph/io.py @@ -27,6 +27,7 @@ from paddle.fluid import backward from paddle.fluid import unique_name from paddle.fluid.dygraph import layers from paddle.fluid.layers import nn +from paddle.fluid.layers.utils import _hash_with_id from paddle.fluid.dygraph.base import switch_to_static_graph from paddle.fluid.framework import in_dygraph_mode @@ -824,7 +825,8 @@ def _run_dygraph(instance, input, program_holder): 'global_block': trace_program.block(0), 'start_op_index': 0, 'end_op_index': end_op_index, - 'is_test': instance._is_test + 'is_test': instance._is_test, + 'program_id': _hash_with_id(trace_program) }) # NOTE: [ why need set param's gradient type here ] # if user set sparse gradient mode, the param's gradient diff --git a/python/paddle/fluid/layers/utils.py b/python/paddle/fluid/layers/utils.py index d5c0310d165..f877a00c511 100644 --- a/python/paddle/fluid/layers/utils.py +++ b/python/paddle/fluid/layers/utils.py @@ -78,6 +78,15 @@ def is_sequence(seq): not isinstance(seq, six.string_types)) +def _hash_with_id(*args): + """ + Return int hash value calculated by id(arg) or tuple(id1,id2, ...). + """ + assert len(args) > 0 + info = tuple([id(v) for v in args]) + return hash(info) & 0xfffffff + + def _sorted(dict_): """ Returns a sorted list of the dict keys, with error if keys not sortable. diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/CMakeLists.txt b/python/paddle/fluid/tests/unittests/dygraph_to_static/CMakeLists.txt index 18f2ffc2382..bdeb5c42a37 100644 --- a/python/paddle/fluid/tests/unittests/dygraph_to_static/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/CMakeLists.txt @@ -1,5 +1,6 @@ file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") +set(GC_ENVS FLAGS_eager_delete_tensor_gb=0.0) list(REMOVE_ITEM TEST_OPS test_lac) # NOTE(Aurelius84): In case of Windows CI, if open ON_INFER, RWLOCK of Scope will @@ -10,7 +11,7 @@ if(NOT ON_INFER) endif() foreach(TEST_OP ${TEST_OPS}) - py_test_modules(${TEST_OP} MODULES ${TEST_OP}) + py_test_modules(${TEST_OP} MODULES ${TEST_OP} ENVS ${GC_ENVS}) endforeach(TEST_OP) set_tests_properties(test_se_resnet PROPERTIES TIMEOUT 900) diff --git a/python/paddle/fluid/tests/unittests/test_run_program_op.py b/python/paddle/fluid/tests/unittests/test_run_program_op.py index 81490642fa8..b3d0845a4fb 100644 --- a/python/paddle/fluid/tests/unittests/test_run_program_op.py +++ b/python/paddle/fluid/tests/unittests/test_run_program_op.py @@ -23,6 +23,7 @@ import paddle import paddle.fluid as fluid from paddle import compat as cpt from paddle.fluid import core, framework, executor +from paddle.fluid.layers.utils import _hash_with_id paddle.enable_static() @@ -174,6 +175,7 @@ class RunProgramOpTest(unittest.TestCase): def calc_dygraph_output(self, place): self.program_desc, self.fwd_op_num = self.get_program_desc() self.attrs = self.prepare_attrs() + self.attrs['program_id'] = _hash_with_id(self.program_desc) with fluid.dygraph.guard(place): inputs = self.prepare_dygraph_input(place) @@ -189,6 +191,7 @@ class RunProgramOpTest(unittest.TestCase): def calc_dygraph_grad(self, place): self.program_desc, self.fwd_op_num = self.get_program_desc() self.attrs = self.prepare_attrs() + self.attrs['program_id'] = _hash_with_id(self.program_desc) with fluid.dygraph.guard(place): # Step 1. run forward -- GitLab