已合并
docs: 修正dflow设计文档中模块路径和文件索引与代码不一致 #4307
lining23666创建于 26 天前
docs: 修正dflow设计文档中模块路径和文件索引与代码不一致 #4307
已合并
共 2 个文件变更+30-30
| @@ -8,7 +8,7 @@ This document is for dflow developers, describing the architecture design, core | |||
| 8 | 8 | ||
| 9 | ### Scope | 9 | ### Scope |
| 10 | 10 | ||
| 11 | -Covered modules: `flow_graph/`, `pydflow/`, `compiler/`, `base/`, `deployer/`, `executor/`, `udf/`. Does not cover the `llm_datadist` submodule (large model data distribution is an independent feature). | 11 | +Covered modules: `flow_graph/`, `pydflow/`, `runner/` (including `session/`, `compiler/`, `executor/`), `base/`, `deployer/`, `udf/`. Does not cover the `llm_datadist` submodule (large model data distribution is an independent feature). |
| 12 | 12 | ||
| 13 | Related documents: | 13 | Related documents: |
| 14 | - [udf.md](udf.md) -- UDF submodule independent document | 14 | - [udf.md](udf.md) -- UDF submodule independent document |
| @@ -53,16 +53,16 @@ flowchart TD | |||
| 53 | FG["flow_graph/<br/>Graph construction core: FlowGraph/FlowNode/ProcessPoint"] | 53 | FG["flow_graph/<br/>Graph construction core: FlowGraph/FlowNode/ProcessPoint"] |
| 54 | end | 54 | end |
| 55 | subgraph session | 55 | subgraph session |
| 56 | - SES["session/<br/>DFlowSession API entry<br/>compilation/deployment/execution coordination"] | 56 | + SES["runner/session/<br/>DFlowSession API entry<br/>compilation/deployment/execution coordination"] |
| 57 | end | 57 | end |
| 58 | subgraph Compilation | 58 | subgraph Compilation |
| 59 | - CMP["compiler/<br/>FlowModelBuilder/PNE engine<br/>FlowGraph→FlowModel"] | 59 | + CMP["runner/compiler/<br/>FlowModelBuilder/PNE engine<br/>FlowGraph→FlowModel"] |
| 60 | end | 60 | end |
| 61 | subgraph Deployment | 61 | subgraph Deployment |
| 62 | DEP["deployer/<br/>Multi-node deployment/cross-node communication<br/>fork executor processes"] | 62 | DEP["deployer/<br/>Multi-node deployment/cross-node communication<br/>fork executor processes"] |
| 63 | end | 63 | end |
| 64 | subgraph Execution | 64 | subgraph Execution |
| 65 | - EXEC["executor/<br/>Heterogeneous executors/data alignment<br/>Feed/Fetch"] | 65 | + EXEC["runner/executor/<br/>Heterogeneous executors/data alignment<br/>Feed/Fetch"] |
| 66 | end | 66 | end |
| 67 | subgraph UDF Submodule | 67 | subgraph UDF Submodule |
| 68 | UDF["udf/<br/>User Defined Function framework<br/>(refer to udf.md)"] | 68 | UDF["udf/<br/>User Defined Function framework<br/>(refer to udf.md)"] |
| @@ -80,11 +80,11 @@ flowchart TD | |||
| 80 | |--------|---------------------| | 80 | |--------|---------------------| |
| 81 | | `flow_graph/` | C++ graph construction API: FlowGraph/FlowNode/FlowData/ProcessPoint system | | 81 | | `flow_graph/` | C++ graph construction API: FlowGraph/FlowNode/FlowData/ProcessPoint system | |
| 82 | | `pydflow/` | Python wrapper, @pyflow decorator, PyTorch integration, UDF project auto-generation | | 82 | | `pydflow/` | Python wrapper, @pyflow decorator, PyTorch integration, UDF project auto-generation | |
| 83 | -| `session/` | DFlowSession API entry, coordination hub for compilation, deployment, and execution | | 83 | +| `runner/session/` | DFlowSession API entry, coordination hub for compilation, deployment, and execution | |
| 84 | -| `compiler/` | FlowModelBuilder/PNE engine mechanism, graph optimization passes, compiles FlowGraph to FlowModel | | 84 | +| `runner/compiler/` | FlowModelBuilder/PNE engine mechanism, graph optimization passes, compiles FlowGraph to FlowModel | |
| 85 | | `base/` | Model abstraction (FlowModel/GraphModel/PneModel), ModelRelation, deployment planning, OM serialization | | 85 | | `base/` | Model abstraction (FlowModel/GraphModel/PneModel), ModelRelation, deployment planning, OM serialization | |
| 86 | | `deployer/` | Multi-node master-slave deployment, cross-node gRPC/memory queue communication, subprocess management | | 86 | | `deployer/` | Multi-node master-slave deployment, cross-node gRPC/memory queue communication, subprocess management | |
| 87 | -| `executor/` | Heterogeneous executors, Feed/Fetch, data alignment, exception handling | | 87 | +| `runner/executor/` | Heterogeneous executors, Feed/Fetch, data alignment, exception handling | |
| 88 | | `udf/` | UDF framework: SO loading/registration, state machine scheduling, message abstraction, built-in UDFs (refer to independent document) | | 88 | | `udf/` | UDF framework: SO loading/registration, state machine scheduling, message abstraction, built-in UDFs (refer to independent document) | |
| 89 | 89 | ||
| 90 | --- | 90 | --- |
| @@ -184,7 +184,7 @@ The three externally provided ProcessPoint types correspond to different computa | |||
| 184 | 184 | ||
| 185 | Additionally, `ModelPp` exists in the code (loads pre-compiled OM models, loads directly without compilation); it is an internal experimental feature and does not provide external interfaces. | 185 | Additionally, `ModelPp` exists in the code (loads pre-compiled OM models, loads directly without compilation); it is an internal experimental feature and does not provide external interfaces. |
| 186 | 186 | ||
| 187 | -During `FlowGraph` construction, `MultiThreadGraphBuilder` (8 threads parallel building GraphPp subgraphs) builds the FlowOperator list into a GE `Graph`, and sets `ATTR_NAME_IS_DATA_FLOW_GRAPH = true` to mark this graph as a dflow graph. | 187 | +During `FlowGraph` construction, the FlowOperator list is built into a GE `Graph`, and `ATTR_NAME_IS_DATA_FLOW_GRAPH = true` is set to mark this graph as a dflow graph. |
| 188 | 188 | ||
| 189 | ### 3.2 C++ Runtime Interface | 189 | ### 3.2 C++ Runtime Interface |
| 190 | 190 | ||
| @@ -277,7 +277,7 @@ The compilation layer is located in `dflow/runner/compiler/`, adopting a four-la | |||
| 277 | 277 | ||
| 278 | ```mermaid | 278 | ```mermaid |
| 279 | flowchart TD | 279 | flowchart TD |
| 280 | - A["session/<br/>DFlowSession API entry<br/>+ lifecycle management"] --> B["model/<br/>FlowModelBuilder construction core<br/>+ FlowModelCache caching"] | 280 | + A["runner/session/<br/>DFlowSession API entry<br/>+ lifecycle management"] --> B["model/<br/>FlowModelBuilder construction core<br/>+ FlowModelCache caching"] |
| 281 | B --> C["pne/<br/>ProcessNodeEngine engine abstraction<br/>+ UDF/CPU/NPU three engines"] | 281 | B --> C["pne/<br/>ProcessNodeEngine engine abstraction<br/>+ UDF/CPU/NPU three engines"] |
| 282 | B --> D["data_flow_graph/<br/>Graph parsing + PP loading + compilation passes<br/>+ model relation construction + deployment planning"] | 282 | B --> D["data_flow_graph/<br/>Graph parsing + PP loading + compilation passes<br/>+ model relation construction + deployment planning"] |
| 283 | C --> D | 283 | C --> D |
| @@ -320,11 +320,11 @@ Engines are registered through `REGISTER_PROCESS_NODE_ENGINE` macro + SO plugin | |||
| 320 | 320 | ||
| 321 | **Multi-level caching** avoids repeated compilation: root model cache (graph_key index) + sub-model cache (SHA256 hash matching) + UDF cache (release_info matching, avoiding repeated cmake/make) + buildinfo cache. | 321 | **Multi-level caching** avoids repeated compilation: root model cache (graph_key index) + sub-model cache (SHA256 hash matching) + UDF cache (release_info matching, avoiding repeated cmake/make) + buildinfo cache. |
| 322 | 322 | ||
| 323 | -Subgraphs are compiled in parallel through `MultiThreadGraphBuilder` with multiple threads; FunctionPp uses async cmake/make compilation. The three-level caching coordination ensures incremental compilation efficiency. | 323 | +FunctionPp uses async cmake/make compilation. The three-level caching coordination ensures incremental compilation efficiency. |
| 324 | 324 | ||
| 325 | ### 4.2 Model Abstraction Layer: FlowModel and ModelRelation | 325 | ### 4.2 Model Abstraction Layer: FlowModel and ModelRelation |
| 326 | 326 | ||
| 327 | -`base/model/` defines the dflow model abstraction system, adopting a composition pattern inheritance hierarchy (`base/model/pne_model.h`): | 327 | +`base/model/` defines the dflow model abstraction system, adopting a composition pattern inheritance hierarchy (`inc/data_flow/model/pne_model.h`): |
| 328 | 328 | ||
| 329 | ```mermaid | 329 | ```mermaid |
| 330 | classDiagram | 330 | classDiagram |
| @@ -546,7 +546,7 @@ For UDF Python development, refer to [udf.md](udf.md). | |||
| 546 | 546 | ||
| 547 | #### 4.6.1 How UDF Execution Location (host/device) Is Determined | 547 | #### 4.6.1 How UDF Execution Location (host/device) Is Determined |
| 548 | 548 | ||
| 549 | -The final UDF execution location is determined by the compilation-time attribute chain; the core logic is in `DataFlowGraphAutoDeployer::SelectResourceType` (`compiler/data_flow_graph/data_flow_graph_auto_deployer.cc`): | 549 | +The final UDF execution location is determined by the compilation-time attribute chain; the core logic is in `DataFlowGraphAutoDeployer::SelectResourceType` (`runner/compiler/data_flow_graph/data_flow_graph_auto_deployer.cc`): |
| 550 | 550 | ||
| 551 | | Attribute | Meaning | Setting Location | | 551 | | Attribute | Meaning | Setting Location | |
| 552 | |-----------|---------|-----------------| | 552 | |-----------|---------|-----------------| |
| @@ -567,9 +567,9 @@ Although heavy_load UDF executes on the host CPU, it still needs to specify logi | |||
| 567 | 567 | ||
| 568 | #### 4.6.2 How Users Specify Deployment Location | 568 | #### 4.6.2 How Users Specify Deployment Location |
| 569 | 569 | ||
| 570 | -Users pass the deployment configuration JSON file through the compilation option `ge.experiment.data_flow_deploy_info_path` (`BuildModel` in `compiler/model/flow_model_builder.cc`). The configuration matches FlowNodes by node **deployment name** -- the deployment name preferentially uses the alias; if the node has no alias set, the original node name is used (`GetNodeDeployName` in `data_flow_graph_auto_deployer.cc`). | 570 | +Users pass the deployment configuration JSON file through the compilation option `ge.experiment.data_flow_deploy_info_path` (`BuildModel` in `runner/compiler/model/flow_model_builder.cc`). The configuration matches FlowNodes by node **deployment name** -- the deployment name preferentially uses the alias; if the node has no alias set, the original node name is used (`GetNodeDeployName` in `data_flow_graph_auto_deployer.cc`). |
| 571 | 571 | ||
| 572 | -**Deployment configuration JSON structure** (`ReadDeployInfoFromJsonFile` in `compiler/data_flow_graph/compile_config_json.cc`): | 572 | +**Deployment configuration JSON structure** (`ReadDeployInfoFromJsonFile` in `runner/compiler/data_flow_graph/compile_config_json.cc`): |
| 573 | 573 | ||
| 574 | ```json | 574 | ```json |
| 575 | { | 575 | { |
| @@ -625,7 +625,7 @@ Using a complete Python user workflow as an example, the full chain from code to | |||
| 625 | flowchart TD | 625 | flowchart TD |
| 626 | subgraph Graph Construction Phase | 626 | subgraph Graph Construction Phase |
| 627 | A1["df.init(options)"] --> A2["df.FlowData() / @df.pyflow / GraphProcessPoint"] | 627 | A1["df.init(options)"] --> A2["df.FlowData() / @df.pyflow / GraphProcessPoint"] |
| 628 | - A2 --> A3["df.FlowGraph(outputs=[...])<br/>Reverse traversal extracts nodes+inputs<br/>MultiThreadGraphBuilder→ComputeGraph"] | 628 | + A2 --> A3["df.FlowGraph(outputs=[...])<br/>Reverse traversal extracts nodes+inputs<br/>Build to ComputeGraph"] |
| 629 | end | 629 | end |
| 630 | subgraph Compilation Phase | 630 | subgraph Compilation Phase |
| 631 | B1["graph.feed_data() triggers lazy compilation"] --> B2["DFlowSession.CompileAndLoadGraph"] | 631 | B1["graph.feed_data() triggers lazy compilation"] --> B2["DFlowSession.CompileAndLoadGraph"] |
| @@ -8,7 +8,7 @@ | |||
| 8 | 8 | ||
| 9 | ### 范围 | 9 | ### 范围 |
| 10 | 10 | ||
| 11 | -涵盖模块：`flow_graph/`、`pydflow/`、`compiler/`、`base/`、`deployer/`、`executor/`、`udf/`。不涵盖 `llm_datadist` 子模块（大模型数据分发为独立特性）。 | 11 | +涵盖模块：`flow_graph/`、`pydflow/`、`runner/`（含 `session/`、`compiler/`、`executor/`）、`base/`、`deployer/`、`udf/`。不涵盖 `llm_datadist` 子模块（大模型数据分发为独立特性）。 |
| 12 | 12 | ||
| 13 | 相关文档： | 13 | 相关文档： |
| 14 | - [udf.md](udf.md) — UDF 子模块独立文档 | 14 | - [udf.md](udf.md) — UDF 子模块独立文档 |
| @@ -53,16 +53,16 @@ flowchart TD | |||
| 53 | FG["flow_graph/<br/>构图核心：FlowGraph/FlowNode/ProcessPoint"] | 53 | FG["flow_graph/<br/>构图核心：FlowGraph/FlowNode/ProcessPoint"] |
| 54 | end | 54 | end |
| 55 | subgraph session | 55 | subgraph session |
| 56 | - SES["session/<br/>DFlowSession API入口<br/>编译/部署/执行协调"] | 56 | + SES["runner/session/<br/>DFlowSession API入口<br/>编译/部署/执行协调"] |
| 57 | end | 57 | end |
| 58 | subgraph 编译 | 58 | subgraph 编译 |
| 59 | - CMP["compiler/<br/>FlowModelBuilder/PNE引擎<br/>FlowGraph→FlowModel"] | 59 | + CMP["runner/compiler/<br/>FlowModelBuilder/PNE引擎<br/>FlowGraph→FlowModel"] |
| 60 | end | 60 | end |
| 61 | subgraph 部署 | 61 | subgraph 部署 |
| 62 | DEP["deployer/<br/>多节点部署/跨节点通信<br/>fork executor进程"] | 62 | DEP["deployer/<br/>多节点部署/跨节点通信<br/>fork executor进程"] |
| 63 | end | 63 | end |
| 64 | subgraph 执行 | 64 | subgraph 执行 |
| 65 | - EXEC["executor/<br/>异构执行器/数据对齐<br/>Feed/Fetch"] | 65 | + EXEC["runner/executor/<br/>异构执行器/数据对齐<br/>Feed/Fetch"] |
| 66 | end | 66 | end |
| 67 | subgraph UDF子模块 | 67 | subgraph UDF子模块 |
| 68 | UDF["udf/<br/>用户自定义函数框架<br/>（详见 udf.md）"] | 68 | UDF["udf/<br/>用户自定义函数框架<br/>（详见 udf.md）"] |
| @@ -80,11 +80,11 @@ flowchart TD | |||
| 80 | |------|----------| | 80 | |------|----------| |
| 81 | | `flow_graph/` | C++ 构图 API：FlowGraph/FlowNode/FlowData/ProcessPoint 体系 | | 81 | | `flow_graph/` | C++ 构图 API：FlowGraph/FlowNode/FlowData/ProcessPoint 体系 | |
| 82 | | `pydflow/` | Python 封装、@pyflow 装饰器、PyTorch 集成、UDF 工程自动生成 | | 82 | | `pydflow/` | Python 封装、@pyflow 装饰器、PyTorch 集成、UDF 工程自动生成 | |
| 83 | -| `session/` | DFlowSession API 入口，编译、部署、执行的协调中枢 | | 83 | +| `runner/session/` | DFlowSession API 入口，编译、部署、执行的协调中枢 | |
| 84 | -| `compiler/` | FlowModelBuilder/PNE 引擎机制、图优化 pass，将 FlowGraph 编译为 FlowModel | | 84 | +| `runner/compiler/` | FlowModelBuilder/PNE 引擎机制、图优化 pass，将 FlowGraph 编译为 FlowModel | |
| 85 | | `base/` | 模型抽象（FlowModel/GraphModel/PneModel）、ModelRelation、部署规划、OM 序列化 | | 85 | | `base/` | 模型抽象（FlowModel/GraphModel/PneModel）、ModelRelation、部署规划、OM 序列化 | |
| 86 | | `deployer/` | 多节点主从部署、跨节点 gRPC/内存队列通信、子进程管理 | | 86 | | `deployer/` | 多节点主从部署、跨节点 gRPC/内存队列通信、子进程管理 | |
| 87 | -| `executor/` | 异构执行器、Feed/Fetch、数据对齐、异常处理 | | 87 | +| `runner/executor/` | 异构执行器、Feed/Fetch、数据对齐、异常处理 | |
| 88 | | `udf/` | UDF 框架：SO 加载注册、状态机调度、消息抽象、内置 UDF（详见独立文档） | | 88 | | `udf/` | UDF 框架：SO 加载注册、状态机调度、消息抽象、内置 UDF（详见独立文档） | |
| 89 | 89 | ||
| 90 | --- | 90 | --- |
| @@ -184,7 +184,7 @@ classDiagram | |||
| 184 | 184 | ||
| 185 | 此外代码中还存在 `ModelPp`（加载预编译 OM 模型，直接加载不编译），它是内部实验性特性，未对外提供接口。 | 185 | 此外代码中还存在 `ModelPp`（加载预编译 OM 模型，直接加载不编译），它是内部实验性特性，未对外提供接口。 |
| 186 | 186 | ||
| 187 | -`FlowGraph` 构建时通过 `MultiThreadGraphBuilder`（8 线程并行构建 GraphPp 子图）将 FlowOperator 列表构建为 GE `Graph`，并设置 `ATTR_NAME_IS_DATA_FLOW_GRAPH = true` 标记此图为 dflow 图。 | 187 | +`FlowGraph` 构建时将 FlowOperator 列表构建为 GE `Graph`，并设置 `ATTR_NAME_IS_DATA_FLOW_GRAPH = true` 标记此图为 dflow 图。 |
| 188 | 188 | ||
| 189 | ### 3.2 C++ 运行接口 | 189 | ### 3.2 C++ 运行接口 |
| 190 | 190 | ||
| @@ -277,7 +277,7 @@ GraphPp 是否配置动态 shape 直接决定其走静态还是动态执行路 | |||
| 277 | 277 | ||
| 278 | ```mermaid | 278 | ```mermaid |
| 279 | flowchart TD | 279 | flowchart TD |
| 280 | - A["session/<br/>DFlowSession API 入口<br/>+ 生命周期管理"] --> B["model/<br/>FlowModelBuilder 构建核心<br/>+ FlowModelCache 缓存"] | 280 | + A["runner/session/<br/>DFlowSession API 入口<br/>+ 生命周期管理"] --> B["model/<br/>FlowModelBuilder 构建核心<br/>+ FlowModelCache 缓存"] |
| 281 | B --> C["pne/<br/>ProcessNodeEngine 引擎抽象<br/>+ UDF/CPU/NPU 三引擎"] | 281 | B --> C["pne/<br/>ProcessNodeEngine 引擎抽象<br/>+ UDF/CPU/NPU 三引擎"] |
| 282 | B --> D["data_flow_graph/<br/>图解析 + PP加载 + 编译Pass<br/>+ 模型关系构建 + 部署规划"] | 282 | B --> D["data_flow_graph/<br/>图解析 + PP加载 + 编译Pass<br/>+ 模型关系构建 + 部署规划"] |
| 283 | C --> D | 283 | C --> D |
| @@ -320,11 +320,11 @@ CPU 引擎继承 NPU 引擎仅重写 `GetEngineName`，编译流程完全复用 | |||
| 320 | 320 | ||
| 321 | **多级缓存**避免重复编译：root 模型缓存（graph_key 索引）+ 子模型缓存（SHA256 哈希匹配）+ UDF 缓存（release_info 匹配，避免重复 cmake/make）+ buildinfo 缓存。 | 321 | **多级缓存**避免重复编译：root 模型缓存（graph_key 索引）+ 子模型缓存（SHA256 哈希匹配）+ UDF 缓存（release_info 匹配，避免重复 cmake/make）+ buildinfo 缓存。 |
| 322 | 322 | ||
| 323 | -子图通过 `MultiThreadGraphBuilder` 多线程并行编译，FunctionPp 异步 cmake/make 编译，三级缓存协同保证增量编译效率。 | 323 | +FunctionPp 异步 cmake/make 编译，三级缓存协同保证增量编译效率。 |
| 324 | 324 | ||
| 325 | ### 4.2 模型抽象层：FlowModel 与 ModelRelation | 325 | ### 4.2 模型抽象层：FlowModel 与 ModelRelation |
| 326 | 326 | ||
| 327 | -`base/model/` 定义了 dflow 的模型抽象体系，采用组合模式的继承体系（`base/model/pne_model.h`）： | 327 | +`base/model/` 定义了 dflow 的模型抽象体系，采用组合模式的继承体系（`inc/data_flow/model/pne_model.h`）： |
| 328 | 328 | ||
| 329 | ```mermaid | 329 | ```mermaid |
| 330 | classDiagram | 330 | classDiagram |
| @@ -546,7 +546,7 @@ UDF 的 Python 开发方式详见 [udf.md](udf.md)。 | |||
| 546 | 546 | ||
| 547 | #### 4.6.1 UDF 执行位置（host/device）如何决定 | 547 | #### 4.6.1 UDF 执行位置（host/device）如何决定 |
| 548 | 548 | ||
| 549 | -UDF 的最终执行位置由编译期属性链路决定，核心逻辑在 `DataFlowGraphAutoDeployer::SelectResourceType`（`compiler/data_flow_graph/data_flow_graph_auto_deployer.cc`）： | 549 | +UDF 的最终执行位置由编译期属性链路决定，核心逻辑在 `DataFlowGraphAutoDeployer::SelectResourceType`（`runner/compiler/data_flow_graph/data_flow_graph_auto_deployer.cc`）： |
| 550 | 550 | ||
| 551 | | 属性 | 含义 | 设置位置 | | 551 | | 属性 | 含义 | 设置位置 | |
| 552 | |------|------|----------| | 552 | |------|------|----------| |
| @@ -567,9 +567,9 @@ heavy_load UDF 虽然在 host CPU 执行，但仍需指定 logic_device_id—— | |||
| 567 | 567 | ||
| 568 | #### 4.6.2 用户如何指定部署位置 | 568 | #### 4.6.2 用户如何指定部署位置 |
| 569 | 569 | ||
| 570 | -用户通过编译选项 `ge.experiment.data_flow_deploy_info_path` 传入部署配置 JSON 文件（`compiler/model/flow_model_builder.cc` 的 `BuildModel`）。配置以节点的**部署名**为 key 匹配 FlowNode——部署名优先取 alias，若节点未设置 alias 则用节点原名（`data_flow_graph_auto_deployer.cc` 的 `GetNodeDeployName`）。 | 570 | +用户通过编译选项 `ge.experiment.data_flow_deploy_info_path` 传入部署配置 JSON 文件（`runner/compiler/model/flow_model_builder.cc` 的 `BuildModel`）。配置以节点的**部署名**为 key 匹配 FlowNode——部署名优先取 alias，若节点未设置 alias 则用节点原名（`data_flow_graph_auto_deployer.cc` 的 `GetNodeDeployName`）。 |
| 571 | 571 | ||
| 572 | -**部署配置 JSON 结构**（`compiler/data_flow_graph/compile_config_json.cc` 的 `ReadDeployInfoFromJsonFile`）： | 572 | +**部署配置 JSON 结构**（`runner/compiler/data_flow_graph/compile_config_json.cc` 的 `ReadDeployInfoFromJsonFile`）： |
| 573 | 573 | ||
| 574 | ```json | 574 | ```json |
| 575 | { | 575 | { |
| @@ -625,7 +625,7 @@ heavy_load UDF 虽然在 host CPU 执行，但仍需指定 logic_device_id—— | |||
| 625 | flowchart TD | 625 | flowchart TD |
| 626 | subgraph 构图阶段 | 626 | subgraph 构图阶段 |
| 627 | A1["df.init(options)"] --> A2["df.FlowData() / @df.pyflow / GraphProcessPoint"] | 627 | A1["df.init(options)"] --> A2["df.FlowData() / @df.pyflow / GraphProcessPoint"] |
| 628 | - A2 --> A3["df.FlowGraph(outputs=[...])<br/>反向遍历提取节点+输入<br/>MultiThreadGraphBuilder→ComputeGraph"] | 628 | + A2 --> A3["df.FlowGraph(outputs=[...])<br/>反向遍历提取节点+输入<br/>构建为ComputeGraph"] |
| 629 | end | 629 | end |
| 630 | subgraph 编译阶段 | 630 | subgraph 编译阶段 |
| 631 | B1["graph.feed_data() 触发惰性编译"] --> B2["DFlowSession.CompileAndLoadGraph"] | 631 | B1["graph.feed_data() 触发惰性编译"] --> B2["DFlowSession.CompileAndLoadGraph"] |