enable serialize prepacked weights into data file #22256
Conversation
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
| import os | ||
|
|
||
| import numpy as np | ||
| import onnx |
Check notice
Code scanning / CodeQL
Module is imported with 'import' and 'import from' Note test
| void MatMulNBits<T1>::ConvertPrepackWeightIntoTensor(const onnxruntime::Tensor& tensor) { | ||
| if (!packed_tensor_) { | ||
| std::vector<int64_t> weights_dims = {static_cast<int64_t>((packed_b_size_ - 1) / tensor.DataType()->Size()) + 1}; | ||
| packed_tensor_ = new Tensor(tensor.DataType(), |
Check warning
Code scanning / PREfast
Avoid calling new and delete explicitly, use std::make_unique instead (r.11). Warning
| void MatMulNBits<T1>::ConvertPrepackWeightIntoTensor(const onnxruntime::Tensor& tensor) { | ||
| if (!packed_tensor_) { | ||
| std::vector<int64_t> weights_dims = {static_cast<int64_t>((packed_b_size_ - 1) / tensor.DataType()->Size()) + 1}; | ||
| packed_tensor_ = new Tensor(tensor.DataType(), |
Check warning
Code scanning / PREfast
Avoid calling new and delete explicitly, use std::make_unique instead (r.11). Warning
| packed_b_.get(), | ||
| OrtMemoryInfo(CPU, OrtAllocatorType::OrtDeviceAllocator)); | ||
| } else { | ||
| packed_tensor_ = new Tensor(packed_tensor_->DataType(), |
Check warning
Code scanning / PREfast
Avoid calling new and delete explicitly, use std::make_unique instead (r.11). Warning
| packed_b_.get(), | ||
| OrtMemoryInfo(CPU, OrtAllocatorType::OrtDeviceAllocator)); | ||
| } else { | ||
| packed_tensor_ = new Tensor(packed_tensor_->DataType(), |
Check warning
Code scanning / PREfast
Avoid calling new and delete explicitly, use std::make_unique instead (r.11). Warning
Description
part of #21448
This change is intend to save CPU memory during model load for inference.
Added session option save_prepacked_constant_initializers, with save_prepacked_constant_initializers turn on:
Tested with model Phi-3-mini-instruct-onnx,
with ORT 1.12.0:
with this change:
Peak memory usage dropped from 5.438 GB to 2.726GB.
This change takes advantage of ORT loads external initializer with mmap on CPU. Prepack will use extra memory on heap, omit prepack process can save this part of memory (roughly same size as external initializers).
next step:
Change all the kernels on CPU with PrePack method implemented and test properly. Will do in next PR.
Motivation and Context