Release Notes

MindSpore Lite 2.7.1 Release Notes

Major Features and Improvements

MindSpore Lite is decoupled from MindSpore, and CPU operator libraries and other related dynamic libraries evolve independently of MindSpore.
The cache algorithm in AIGC models supporting image generation is implemented based on graph patterns.

API Change

Added MultiModelRunner and ModelExecutor interfaces, supporting graph mode implementation of the Cache algorithm.

import mindspore_lite as mslite
import numpy as np
dtype_map = {
    mslite.DataType.FLOAT32: np.float32,
    mslite.DataType.INT32: np.int32,
    mslite.DataType.FLOAT16: np.float16,
    mslite.DataType.INT8: np.int8
}
context = mslite.Context()
context.target = ["ascend"]
context.ascend.devcie_id = 0
runner = mslite.MultiModelRunner()
model_path = "path_to_model"
runner.build_from_file(model_path, mslite.ModelType.MINDIR, context)
execs = runner.get_model_executor()
for exec_ in execs:
    exec_inputs = exec_.get_inputs()
    for input_ in exec_inputs:
        data = np.random.randn(*input_.shape).astype(dtype_map[input_.dtype])
        input_.set_data_from_numpy(data)
    exec_.predict(exec_inputs)

The offline conversion tool conver_lite implements subgraph partitioning through the configuration of the SplitGraph parameter and the split_node_name parameter.
```
[SplitGraph]
split_node_name=[[node_name_1],[node_name_2]]
```

Contributors

YeFeng_24,xiong-pan,jjfeing,liuf9,zhangzhugucheng,xu_anyue,yiguangzheng,zxx_xxz,jianghui58,hbhu_bin,chenyihang5,qll1998,yangyingchun1999,liuchengji3,cheng-chao23,gemini524,yangly,yanghui00