apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-10 02:50:19 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[Other] Optimize paddle backend (#1265)

Browse Source 
* Optimize paddle backend

* optimize paddle backend

* add version support
This commit is contained in:
Jason
2023-02-08 19:12:03 +08:00
committed by GitHub
parent 60ba4b06c1
commit a4b0565b9a
10 changed files with 265 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

63
fastdeploy/runtime/runtime.cc
View File

@@ -226,53 +226,24 @@ void Runtime::CreatePaddleBackend() {
option.model_format == ModelFormat::PADDLE,
"Backend::PDINFER only supports model format of ModelFormat::PADDLE.");
#ifdef ENABLE_PADDLE_BACKEND
auto pd_option = PaddleBackendOption();
pd_option.model_file = option.model_file;
pd_option.params_file = option.params_file;
pd_option.enable_mkldnn = option.pd_enable_mkldnn;
pd_option.enable_log_info = option.pd_enable_log_info;
pd_option.mkldnn_cache_size = option.pd_mkldnn_cache_size;
pd_option.use_gpu = (option.device == Device::GPU) ? true : false;
pd_option.use_ipu = (option.device == Device::IPU) ? true : false;
pd_option.gpu_id = option.device_id;
pd_option.delete_pass_names = option.pd_delete_pass_names;
pd_option.cpu_thread_num = option.cpu_thread_num;
pd_option.enable_pinned_memory = option.enable_pinned_memory;
pd_option.external_stream_ = option.external_stream_;
pd_option.model_from_memory_ = option.model_from_memory_;
#ifdef ENABLE_TRT_BACKEND
if (pd_option.use_gpu && option.pd_enable_trt) {
pd_option.enable_trt = true;
pd_option.collect_shape = option.pd_collect_shape;
pd_option.trt_option = option.trt_option;
pd_option.trt_option.gpu_id = option.device_id;
pd_option.trt_option.enable_pinned_memory = option.enable_pinned_memory;
pd_option.trt_disabled_ops_ = option.trt_disabled_ops_;
}
#endif
#ifdef WITH_IPU
if (pd_option.use_ipu) {
auto ipu_option = IpuOption();
ipu_option.ipu_device_num = option.ipu_device_num;
ipu_option.ipu_micro_batch_size = option.ipu_micro_batch_size;
ipu_option.ipu_enable_pipelining = option.ipu_enable_pipelining;
ipu_option.ipu_batches_per_step = option.ipu_batches_per_step;
ipu_option.ipu_enable_fp16 = option.ipu_enable_fp16;
ipu_option.ipu_replica_num = option.ipu_replica_num;
ipu_option.ipu_available_memory_proportion =
option.ipu_available_memory_proportion;
ipu_option.ipu_enable_half_partial = option.ipu_enable_half_partial;
pd_option.ipu_option = ipu_option;
}
#endif
option.paddle_infer_option.model_file = option.model_file;
option.paddle_infer_option.params_file = option.params_file;
option.paddle_infer_option.model_from_memory_ = option.model_from_memory_;
option.paddle_infer_option.device = option.device;
option.paddle_infer_option.device_id = option.device_id;
option.paddle_infer_option.enable_pinned_memory = option.enable_pinned_memory;
option.paddle_infer_option.external_stream_ = option.external_stream_;
option.paddle_infer_option.trt_option = option.trt_option;
option.paddle_infer_option.trt_option.gpu_id = option.device_id;
backend_ = utils::make_unique<PaddleBackend>();
auto casted_backend = dynamic_cast<PaddleBackend*>(backend_.get());
casted_backend->benchmark_option_ = option.benchmark_option;
if (pd_option.model_from_memory_) {
FDASSERT(casted_backend->InitFromPaddle(option.model_file,
option.params_file, pd_option),
"Load model from Paddle failed while initliazing PaddleBackend.");
if (option.model_from_memory_) {
FDASSERT(
casted_backend->InitFromPaddle(option.model_file, option.params_file,
option.paddle_infer_option),
"Load model from Paddle failed while initliazing PaddleBackend.");
ReleaseModelMemoryBuffer();
} else {
std::string model_buffer = "";
@@ -281,9 +252,9 @@ void Runtime::CreatePaddleBackend() {
"Fail to read binary from model file");
FDASSERT(ReadBinaryFromFile(option.params_file, &params_buffer),
"Fail to read binary from parameter file");
FDASSERT(
casted_backend->InitFromPaddle(model_buffer, params_buffer, pd_option),
"Load model from Paddle failed while initliazing PaddleBackend.");
FDASSERT(casted_backend->InitFromPaddle(model_buffer, params_buffer,
option.paddle_infer_option),
"Load model from Paddle failed while initliazing PaddleBackend.");
}
#else
FDASSERT(false,