Skip to content

Commit dc0702f

Browse files
TeslaZhaoTeslaZhao
authored
Using DistConfig in Paddle Inference (#41128)
* Pass compat of conv_transpose_bias_mkldnn_fuse_pass * Fix a bug of strided_slice op, about the axes parameter access memory out of bounds * Fix a bug of strided_slice op, about the axes parameter access memory out of bounds * Fix a bug of transpose op, about accessing memory out of bounds of the perm param * op:transpose_op supports bool type * op:transpose_op supports bool type * Keep strided_slice op behavior consistent with slice op when starts input is less than -rank * Using DistConfig in inference
1 parent a6bf221 commit dc0702f

File tree

5 files changed

+107
-124
lines changed

5 files changed

+107
-124
lines changed

‎paddle/fluid/distributed/ps/table/CMakeLists.txt‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -34,10 +34,9 @@ ${RPC_DEPS} graph_edge graph_node device_context string_helper
3434
simple_threadpool xxhash generator ${EXTERN_DEP})
3535

3636
set_source_files_properties(tensor_accessor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
37-
set_source_files_properties(tensor_table.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
3837

3938
cc_library(tensor_accessor SRCS tensor_accessor.cc DEPS ${TABLE_DEPS} eigen3 ps_framework_proto device_context)
40-
cc_library(tensor_table SRCS tensor_table.cc DEPS eigen3 ps_framework_proto executor scope device_context tensor ${TABLE_DEPS})
39+
cc_library(tensor_table SRCS DEPS eigen3 ps_framework_proto executor scope device_context tensor ${TABLE_DEPS})
4140
set_source_files_properties(table.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
4241

4342
set_source_files_properties(sparse_sgd_rule.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
@@ -54,6 +53,7 @@ cc_library(memory_sparse_table SRCS memory_sparse_table.cc DEPS ps_framework_pro
5453

5554
set_source_files_properties(memory_sparse_geo_table.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
5655
cc_library(memory_sparse_geo_table SRCS memory_sparse_geo_table.cc DEPS ps_framework_proto ${TABLE_DEPS} common_table)
56+
5757
cc_library(table SRCS table.cc DEPS memory_sparse_table memory_sparse_geo_table common_table tensor_accessor tensor_table ps_framework_proto string_helper device_context gflags glog boost)
5858

5959
target_link_libraries(table -fopenmp)

‎paddle/fluid/distributed/ps/table/tensor_table.cc‎

Lines changed: 1 addition & 106 deletions
Original file line numberDiff line numberDiff line change
@@ -16,110 +16,5 @@
1616

1717
DECLARE_double(eager_delete_tensor_gb);
1818
namespace paddle {
19-
namespace distributed {
20-
21-
int32_t TensorTable::set_program_env(
22-
framework::Scope *scope, platform::Place place,
23-
const std::vector<framework::ProgramDesc> *sub_program) {
24-
scope_ = scope;
25-
place_ = place;
26-
executor_ = new framework::Executor(place_);
27-
sub_program_ = sub_program;
28-
return 0;
29-
}
30-
31-
int32_t GlobalStepTable::initialize() {
32-
auto _program_config = _config.tensor();
33-
auto trainers_ = _config.common().trainer_num();
34-
FLAGS_eager_delete_tensor_gb = -1;
35-
// Get Config
36-
if (_program_config.has_startup_program_id()) {
37-
startup_program_id_ = _program_config.startup_program_id();
38-
}
39-
if (_program_config.has_main_program_id()) {
40-
main_program_id_ = _program_config.main_program_id();
41-
}
42-
if (_program_config.has_feed_var_name()) {
43-
feed_var_name_ = _program_config.feed_var_name();
44-
}
45-
if (_program_config.has_fetch_var_name()) {
46-
fetch_var_name_ = _program_config.fetch_var_name();
47-
}
48-
49-
// Run startup program
50-
if (startup_program_id_ != -1) {
51-
std::map<std::string, const framework::LoDTensor *> fake_feed;
52-
std::map<std::string, framework::FetchType *> fake_fetch;
53-
auto startup_program_desc = sub_program_->at(startup_program_id_);
54-
auto ctx = executor_->Prepare(startup_program_desc, 0);
55-
executor_->RunPreparedContext(ctx.get(), scope_, false);
56-
}
57-
58-
if (main_program_id_ != -1) {
59-
// Run main porgram, if program is used for learning decay
60-
auto main_program_desc = sub_program_->at(main_program_id_);
61-
auto main_ctx = executor_->Prepare(main_program_desc, 0);
62-
exec_context_ = std::move(main_ctx);
63-
executor_->RunPreparedContext(exec_context_.get(), scope_, false);
64-
// init decay_counters
65-
decay_counters_.reserve(trainers_);
66-
for (int32_t i = 0; i < trainers_; ++i) {
67-
decay_counters_[i] = 0;
68-
}
69-
}
70-
71-
return 0;
72-
}
73-
74-
int32_t GlobalStepTable::set_table_map(
75-
std::unordered_map<uint32_t, std::shared_ptr<Table>> *table_map) {
76-
auto *lr_var = scope_->FindVar(fetch_var_name_);
77-
auto *lr_tensor = lr_var->GetMutable<framework::LoDTensor>();
78-
auto *lr_value = lr_tensor->mutable_data<float>(platform::CPUPlace());
79-
VLOG(3) << "GlobalStepTable::set_table_map set global lr: " << *lr_value;
80-
81-
for (auto iter = table_map->begin(); iter != table_map->end(); iter++) {
82-
auto table_id = iter->first;
83-
if (table_id == _config.table_id()) {
84-
continue;
85-
}
86-
iter->second->set_global_lr(lr_value);
87-
}
88-
return 0;
89-
}
90-
91-
int32_t GlobalStepTable::push_dense(const int64_t *values,
92-
const int32_t trainer_id) {
93-
return _run_program(values, trainer_id);
94-
}
95-
96-
int32_t GlobalStepTable::_run_program(const int64_t *values,
97-
const uint32_t trainer_id) {
98-
FLAGS_eager_delete_tensor_gb = -1;
99-
auto counter = decay_counters_.at(trainer_id);
100-
counter += int(values[0]);
101-
decay_counters_.at(trainer_id) = counter;
102-
103-
auto *global_step_var = scope_->FindVar(feed_var_name_);
104-
auto *tensor = global_step_var->GetMutable<framework::LoDTensor>();
105-
auto *value = tensor->mutable_data<int64_t>(platform::CPUPlace());
106-
107-
auto global_counter = 0;
108-
for (auto &trainer_counter : decay_counters_) {
109-
global_counter += trainer_counter.second;
110-
}
111-
112-
// Todo: hard code for increment op
113-
value[0] = global_counter - 1;
114-
VLOG(3) << "GlobalStepTable::_run_program global_counter " << value[0];
115-
116-
executor_->RunPreparedContext(exec_context_.get(), scope_, false, false);
117-
auto *lr_var = scope_->FindVar(fetch_var_name_);
118-
auto *lr_tensor = lr_var->GetMutable<framework::LoDTensor>();
119-
auto *lr_value = lr_tensor->mutable_data<float>(platform::CPUPlace());
120-
VLOG(3) << "GlobalStepTable::LR value: " << lr_value[0];
121-
return 0;
122-
}
123-
124-
} // namespace distributed
19+
namespace distributed {} // namespace distributed
12520
} // namespace paddle

‎paddle/fluid/distributed/ps/table/tensor_table.h‎

Lines changed: 102 additions & 15 deletions
Original file line numberDiff line numberDiff line change
@@ -37,6 +37,8 @@ struct ExecutorPrepareContext;
3737
} // namespace framework
3838
} // namespace paddle
3939

40+
DECLARE_double(eager_delete_tensor_gb);
41+
4042
namespace paddle {
4143
namespace distributed {
4244

@@ -66,9 +68,9 @@ class TensorTable : public Table {
6668

6769
virtual void *get_shard(size_t shard_idx) { return 0; }
6870

69-
virtual int32_t initialize_shard() { return 0; };
71+
virtual int32_t initialize_shard() { return 0; }
7072

71-
virtual int32_t flush() { return 0; };
73+
virtual int32_t flush() { return 0; }
7274

7375
virtual int32_t load(const std::string &path, const std::string &param) {
7476
return 0;
@@ -77,18 +79,23 @@ class TensorTable : public Table {
7779
return 0;
7880
}
7981

80-
virtual void clear(){};
82+
virtual void clear() {}
8183

82-
virtual int32_t initialize() override { return 0; };
84+
int32_t initialize() override { return 0; }
8385

84-
virtual int32_t push_dense(const int64_t *values,
85-
const int32_t trainer_id) override {
86+
int32_t push_dense(const int64_t *values, const int32_t trainer_id) override {
8687
return 0;
87-
};
88+
}
8889

89-
virtual int32_t set_program_env(
90+
int32_t set_program_env(
9091
framework::Scope *scope, platform::Place place,
91-
const std::vector<framework::ProgramDesc> *sub_program) override;
92+
const std::vector<framework::ProgramDesc> *sub_program) override {
93+
scope_ = scope;
94+
place_ = place;
95+
executor_ = new framework::Executor(place_);
96+
sub_program_ = sub_program;
97+
return 0;
98+
}
9299

93100
protected:
94101
framework::Executor *executor_;
@@ -135,7 +142,7 @@ class DenseTensorTable : public TensorTable {
135142

136143
/*----------------------------------------------------------------------*/
137144

138-
virtual int32_t initialize() override { return 0; }
145+
int32_t initialize() override { return 0; }
139146

140147
int32_t push_dense(const float *values, size_t num) override { return 0; }
141148

@@ -189,18 +196,98 @@ class GlobalStepTable : public DenseTensorTable {
189196

190197
/*----------------------------------------------------------------------*/
191198

192-
int32_t initialize() override;
199+
int32_t initialize() override {
200+
auto _program_config = _config.tensor();
201+
auto trainers_ = _config.common().trainer_num();
202+
FLAGS_eager_delete_tensor_gb = -1;
203+
// Get Config
204+
if (_program_config.has_startup_program_id()) {
205+
startup_program_id_ = _program_config.startup_program_id();
206+
}
207+
if (_program_config.has_main_program_id()) {
208+
main_program_id_ = _program_config.main_program_id();
209+
}
210+
if (_program_config.has_feed_var_name()) {
211+
feed_var_name_ = _program_config.feed_var_name();
212+
}
213+
if (_program_config.has_fetch_var_name()) {
214+
fetch_var_name_ = _program_config.fetch_var_name();
215+
}
216+
217+
// Run startup program
218+
if (startup_program_id_ != -1) {
219+
std::map<std::string, const framework::LoDTensor *> fake_feed;
220+
std::map<std::string, framework::FetchType *> fake_fetch;
221+
auto startup_program_desc = sub_program_->at(startup_program_id_);
222+
auto ctx = executor_->Prepare(startup_program_desc, 0);
223+
executor_->RunPreparedContext(ctx.get(), scope_, false);
224+
}
225+
226+
if (main_program_id_ != -1) {
227+
// Run main porgram, if program is used for learning decay
228+
auto main_program_desc = sub_program_->at(main_program_id_);
229+
auto main_ctx = executor_->Prepare(main_program_desc, 0);
230+
exec_context_ = std::move(main_ctx);
231+
executor_->RunPreparedContext(exec_context_.get(), scope_, false);
232+
// init decay_counters
233+
decay_counters_.reserve(trainers_);
234+
for (int32_t i = 0; i < trainers_; ++i) {
235+
decay_counters_[i] = 0;
236+
}
237+
}
238+
}
193239

194240
int32_t push_dense(const float *values, size_t num) override { return 0; }
195241

196-
int32_t push_dense(const int64_t *values, const int32_t trainer_id);
242+
int32_t push_dense(const int64_t *values, const int32_t trainer_id) {
243+
return _run_program(values, trainer_id);
244+
}
197245

198-
int32_t set_table_map(
199-
std::unordered_map<uint32_t, std::shared_ptr<Table>> *table_map) override;
246+
int32_t set_table_map(std::unordered_map<uint32_t, std::shared_ptr<Table>>
247+
*table_map) override {
248+
auto *lr_var = scope_->FindVar(fetch_var_name_);
249+
auto *lr_tensor = lr_var->GetMutable<framework::LoDTensor>();
250+
auto *lr_value = lr_tensor->mutable_data<float>(platform::CPUPlace());
251+
VLOG(3) << "GlobalStepTable::set_table_map set global lr: " << *lr_value;
252+
253+
for (auto iter = table_map->begin(); iter != table_map->end(); iter++) {
254+
auto table_id = iter->first;
255+
if (table_id == _config.table_id()) {
256+
continue;
257+
}
258+
iter->second->set_global_lr(lr_value);
259+
}
260+
return 0;
261+
}
200262

201263
private:
202264
virtual int32_t _run_program(const int64_t *values,
203-
const uint32_t trainer_id);
265+
const uint32_t trainer_id) {
266+
FLAGS_eager_delete_tensor_gb = -1;
267+
auto counter = decay_counters_.at(trainer_id);
268+
counter += int(values[0]);
269+
decay_counters_.at(trainer_id) = counter;
270+
271+
auto *global_step_var = scope_->FindVar(feed_var_name_);
272+
auto *tensor = global_step_var->GetMutable<framework::LoDTensor>();
273+
auto *value = tensor->mutable_data<int64_t>(platform::CPUPlace());
274+
275+
auto global_counter = 0;
276+
for (auto &trainer_counter : decay_counters_) {
277+
global_counter += trainer_counter.second;
278+
}
279+
280+
// Todo: hard code for increment op
281+
value[0] = global_counter - 1;
282+
VLOG(3) << "GlobalStepTable::_run_program global_counter " << value[0];
283+
284+
executor_->RunPreparedContext(exec_context_.get(), scope_, false, false);
285+
auto *lr_var = scope_->FindVar(fetch_var_name_);
286+
auto *lr_tensor = lr_var->GetMutable<framework::LoDTensor>();
287+
auto *lr_value = lr_tensor->mutable_data<float>(platform::CPUPlace());
288+
VLOG(3) << "GlobalStepTable::LR value: " << lr_value[0];
289+
return 0;
290+
}
204291

205292
private:
206293
std::unordered_map<int, int64_t> decay_counters_;

‎paddle/fluid/inference/CMakeLists.txt‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -93,7 +93,7 @@ if (WITH_CRYPTO)
9393
endif (WITH_CRYPTO)
9494

9595
if (WITH_PSCORE)
96-
set(SHARED_INFERENCE_DEPS ${SHARED_INFERENCE_DEPS} fleet ps_service)
96+
set(SHARED_INFERENCE_DEPS ${SHARED_INFERENCE_DEPS} fleet ps_service tensor_table)
9797
endif ()
9898

9999
if (WITH_ONNXRUNTIME)

‎paddle/fluid/inference/api/paddle_inference_api.h‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -45,6 +45,7 @@ namespace paddle_infer {
4545

4646
using PrecisionType = paddle::AnalysisConfig::Precision;
4747
using Config = paddle::AnalysisConfig;
48+
using DistConfig = paddle::DistConfig;
4849

4950
///
5051
/// \class Predictor

0 commit comments

Comments
 (0)