forked from mindspore-Ecosystem/mindspore
fix some buf: model parallel runner
This commit is contained in:
parent
60587fcb49
commit
3523d11f45
|
@ -657,6 +657,9 @@ std::vector<MSTensor> ModelPool::GetInputs() {
|
|||
return {};
|
||||
}
|
||||
tensor->SetShape(model_pool_inputs_.at(i).Shape());
|
||||
tensor->SetFormat(model_pool_inputs_.at(i).format());
|
||||
tensor->SetAllocator(model_pool_inputs_.at(i).allocator());
|
||||
tensor->SetQuantParams(model_pool_inputs_.at(i).QuantParams());
|
||||
inputs.push_back(*tensor);
|
||||
delete tensor;
|
||||
}
|
||||
|
@ -677,6 +680,9 @@ std::vector<MSTensor> ModelPool::GetOutputs() {
|
|||
return {};
|
||||
}
|
||||
tensor->SetShape(model_pool_outputs_.at(i).Shape());
|
||||
tensor->SetFormat(model_pool_outputs_.at(i).format());
|
||||
tensor->SetAllocator(model_pool_outputs_.at(i).allocator());
|
||||
tensor->SetQuantParams(model_pool_outputs_.at(i).QuantParams());
|
||||
outputs.push_back(*tensor);
|
||||
delete tensor;
|
||||
}
|
||||
|
@ -802,6 +808,20 @@ Status ModelPool::InitAdvancedStrategy(const char *model_buf, size_t size, int b
|
|||
return kSuccess;
|
||||
}
|
||||
auto advanced_thread_num = kDefaultThreadsNum;
|
||||
if (numa_available_ && static_cast<size_t>(advanced_thread_num) > numa_physical_cores_.front().size()) {
|
||||
MS_LOG(WARNING) << "not use advanced strategy | numa available: " << numa_available_ << " | "
|
||||
<< " advanced thread num[" << advanced_thread_num << "] is more than one numa physical core num["
|
||||
<< numa_physical_cores_.front().size() << "]";
|
||||
use_advanced_strategy_ = false;
|
||||
return kSuccess;
|
||||
}
|
||||
if (!numa_available_ && advanced_thread_num > lite::GetCoreNum()) {
|
||||
MS_LOG(WARNING) << "not use advanced strategy | numa available: " << numa_available_ << " | "
|
||||
<< " advanced thread num[" << advanced_thread_num << "] is more than all core num["
|
||||
<< lite::GetCoreNum() << "]";
|
||||
use_advanced_strategy_ = false;
|
||||
return kSuccess;
|
||||
}
|
||||
model_pool_info_[ADVANCED].use_numa = numa_available_;
|
||||
auto status = SetDefaultOptimalModelNum(advanced_thread_num, ADVANCED);
|
||||
if (status != kSuccess) {
|
||||
|
|
Loading…
Reference in New Issue