[feature][api] load weight from buffer · Pull Request !454 · MindSpore/mindspore-lite

1. feat: load weight from buffer

依赖mindspore PR: https://gitee.com/mindspore/mindspore/pulls/90472

---
<!--  Thanks for sending a pull request!  Here are some tips for you:

1) If this is your first time, please read our contributor guidelines: https://gitee.com/mindspore/mindspore-lite/blob/master/CONTRIBUTING.md

2) If you want to contribute your code but don't know who will review and merge, please add label `mindspore-assistant` to the pull request, we will find and do it as soon as possible.
-->

**What type of PR is this?**

/kind feature

---

**What does this PR do / why do we need it**:

支持权重分离的模型通过buffer加载

#### 一、新增对外接口

##### 1. C++ API

```c++
/// \brief Build a model from model buffer so that it can run on a device.
///
/// \param[in] model_data Define the buffer read from a model file.
/// \param[in] data_size Define bytes number of model buffer.
/// \param[in] weight_data Define the buffer read from a weight file.
/// \param[in] weight_size Define bytes number of weight buffer.
/// \param[in] model_type Define The type of model file. Options: ModelType::kMindIR, ModelType::kMindIR_Lite. Only
///     ModelType::kMindIR_Lite is valid for Device-side Inference. Cloud-side Inference supports options
///     ModelType::kMindIR and ModelType::kMindIR_Lite, but option ModelType::kMindIR_Lite will be removed in future
///     iterations. 
/// \param[in] model_context Define the context used to store options during execution.
///
/// \return Status. kSuccess: build success, kLiteModelRebuild: build model repeatedly, Other: other types of errors.

Status Build(const void *model_data, size_t data_size, const void *weight_data, size_t weight_size,
               ModelType model_type, const std::shared_ptr<Context> &model_context = nullptr);
```

##### 2. Python API

```python
def build_from_buffer(
self,
model_bytes,
weight_bytes=None,
model_type=None,
context=None,
config_path="",
config_dict: dict = None,
):
"""
Load and build a model from file.

Args:
    model_bytes (Bytes): Bytes of the mindir model when build from buffer.
    weight_bytes (Bytes, optional): Bytes of the separate weight when build from buffer. Default: ``None``.
    model_type (ModelType, optional): Define The type of input model file. Option is ``ModelType.MINDIR``.
        For details, see
        `ModelType <https://mindspore.cn/lite/api/en/master/mindspore_lite/mindspore_lite.ModelType.html>`_ .
    context (Context, optional): Define the context used to transfer options during execution.
        Default: ``None``. ``None`` means the Context with cpu target.
    config_path (str, optional): Define the config file path. the config file is used to transfer user defined
        options during build model. In the following scenarios, users may need to set the parameter.
        For example, "/home/user/config.txt". Default: ``""``.

```

#### 二、实现方案

##### 1. MindIRLoader
MindIRLoader新增接口，接口入参包含结构和权重的buffer， bool LoadMindIR(const void *model_buffer, const size_t &model_size, const void *weight_buffer, const size_t &weight_size, FuncGraphPtr *func_graph, std::string *user_info_string = nullptr);
在获取权重时，针对权重buffer增加分支处理：
1）`weight_buffer == nullptr && mindir_dec_key_ != nullptr`，执行原有解密读权重。
2）`weight_buffer == nullptr && mindir_dec_key_ == nullptr`，执行原有从文件读权重。
3）`weight_buffer != nullptr && mindir_dec_key_ != nullptr`，报错，不支持对权重buffer的解密。
4）`weight_buffer != nullptr && mindir_dec_key_ == nullptr`，通过proto中的权重偏移量和长度获取权重，针对与传入长度不符的情况报错。

##### 2. Model 适配

1）增加对weight_buffer,weight_size的参数校验及传递。
`weight_data == nullptr && weight_size == 0 && !base_path.empty()`，从模型文件获取权重
`weight_data == nullptr && weight_size == 0 && base_path.empty()`，单buffer模型包含权重
`weight_data != nullptr && weight_size != 0 && base_path.empty()`，权重分离模型，双buffer
其他情况报错

2）圈复杂度：对MindIRLoader调用抽离到独立函数。

---

**Which issue(s) this PR fixes**:

Fixes #ID2D6L

---

**Test Plan and Test result：What scenarios were tested, and what were the verification results（Function, performance, reliability, etc.）**：

新增ST用例

#### 1. 正向用例

| 模型  |  权重分离 | 权重输入(buffer)  | 预期  |
|---|---|---|---|
|  sd1.5_unet.onnx_graph.mindir |  是 | sd1.5_unet.onnx_variables/data_0  | 正常加载，计算正确  |
|  bert_model.onnx.mindir |  否 | None  |  正常加载，计算正确 |
|  bert_model.onnx.mindir |  否 | sd1.5_unet.onnx_variables/data_0  |  正常加载，计算正确 |

#### 2. 负向用例

| 模型  |  权重分离 | 权重输入(buffer)  | 预期  |
|---|---|---|---|
|  sd1.5_unet.onnx_graph.mindir |  是 | None  | RuntimeError("build_from_buffer failed! Error is Common error code.")  |
|  sd1.5_unet.onnx_graph.mindir |  是 | emptyfile  | RuntimeError("build_from_buffer failed! Error is Common error code.")  |
|  empty |  - | None  | RuntimeError("build_from_buffer failed, model_bytes is empty.")  |

#### 3. 参数校验

|  输入 |  预期 |
|---|---|
|  {"model_bytes": None} | TypeError（“model_bytes must be bytes”）  |
| {"model_bytes": bytes()}  |  RuntimeError（"build_from_buffer failed, model_bytes is empty." ）|
| {"model_bytes": b"0", "weight_bytes": str()} | TypeError("weight_bytes must be bytes")|
| {"model_bytes": b"0"}| TypeError("model_type must be ModelType") |
| {"model_bytes": b"0", "model_type": mslite.ModelType.MINDIR_LITE} | RuntimeError("build_from_buffer failed, model_type should be MINDIR") |
| {"model_bytes": b"0", "model_type": mslite.ModelType.MINDIR, "context": str()} | TypeError("context must be Context") |
| {"model_bytes": b"0", "model_type": mslite.ModelType.MINDIR, "config_path": 1} | TypeError("config_path must be str") |
| {"model_bytes": b"0", "model_type": mslite.ModelType.MINDIR, "config_dict": 1} | TypeError("config_dict must be dict") |

![新增用例](https://foruda.gitee.com/images/1763540295769151948/49ce504e_15423976.png "屏幕截图")

---

**Self-checklist**:（**请自检，在[ ]内打上x，我们将检视你的完成情况，否则会导致pr无法合入**）

+ - [x] **设计**：PR对应的方案是否已经经过Maintainer评审，方案检视意见是否均已答复并完成方案修改
+ - [x] **测试**：PR中的代码是否已有UT/ST测试用例进行充分的覆盖，新增测试用例是否随本PR一并上库或已经上库
+ - [x] **验证**：PR描述信息中是否已包含对该PR对应的Feature、Refactor、Bugfix的预期目标达成情况的详细验证结果描述
+ - [x] **接口**：是否涉及对外接口变更，相应变更已得到接口评审组织的通过，API对应的注释信息已经刷新正确
+ - [x] **文档**：是否涉及官网文档修改，如果涉及请及时提交资料到Doc仓

MindSpore/mindspore-lite关闭 .gitee-modal { width: 500px !important; }

搜索帮助

MindSpore/mindspore-lite
关闭