Magpie Kernel Evaluator
amd/skills
Benchmarks LLM inference and drives GPU kernel optimization with Magpie.
A skill your agent uses when working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate…
Install Paddle Design Compiler
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install PaddlePaddle/Paddle paddle-design-compiler --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .claude/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.
Claude Code skills documentation · loads skills from .claude/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .claude/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compilerType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install PaddlePaddle/Paddle paddle-design-compiler --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .agents/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .agents/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .agents/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install PaddlePaddle/Paddle paddle-design-compiler --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .cursor/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .cursor/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/PaddlePaddle/Paddle.git --path .agents/skills/paddle-design-compiler--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install PaddlePaddle/Paddle paddle-design-compiler --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .gemini/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .gemini/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install PaddlePaddle/Paddle paddle-design-compilerInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .github/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .github/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .github/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install PaddlePaddle/Paddle paddle-design-compiler --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/PaddlePaddle/Paddle.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/.agents/skills/paddle-design-compiler .opencode/skills/paddle-design-compiler && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "paddle-design-compiler" agent skill from https://github.com/PaddlePaddle/Paddle/tree/develop/.agents/skills/paddle-design-compiler into .opencode/skills/paddle-design-compiler/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "paddle-design-compiler", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Facts
- Skill name
paddle-design-compiler- GitHub stars
- 24k
- Token cost
- ~3.6k tokens
- SKILL.md length
- 666 words
- Files
- 7 (incl. references)
- Skills in repo
- 11
- Repo updated
- First seen
- Licence
- Apache-2.0
At a glance
A skill your agent uses when working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate…
- Working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture
- SKILL.md covers 全链路概览, SOT（Symbolic Opcode Translator）, PIR（Paddle Intermediate… and CINN 编译与执行, plus 3 more sections
- Calls python
- PIR (Paddle IR) for SSA-based intermediate representation
What it does
Paddle Design Compiler is an agent skill from PaddlePaddle/Paddle. Use when working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate representation, CINN for fused CUDA kernel generation, operator decomposition (Prim), or the end-to-end flow from Python eager code to optimized GPU execution.
Its SKILL.md is about 3.6k tokens, which your agent loads only when the skill is triggered. The skill folder holds 7 other files, including reference files (for example `references/cinn-pipeline.md`, `references/control-flow.md` and `references/executor.md`).
It sits in AI & LLM Engineering, covering Translation, GPU and accelerator computing and Deep learning. It works with Python and CUDA. The repository describes itself as: PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）. The licence is Apache-2.0.
When your agent uses it
- Working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture
- PIR (Paddle IR) for SSA-based intermediate representation
- CINN for fused CUDA kernel generation
- Operator decomposition (Prim)
Example prompts
- “/paddle-design-compiler”
Requirements
- Python 3
What it can do on your machine
Read from SKILL.md and the folder at commit 5434c21. It shows what the files ask for, not the result of running them.
Tool permissions
Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.
From allowed-tools in the SKILL.md frontmatter.
Runs code
Shell commands in SKILL.md call:
python
From the folder's file list and the shell code blocks in SKILL.md.
Network
No URLs in SKILL.md.
From URLs in SKILL.md, links to its own repository left out.
Credentials
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Context cost
Paddle Design Compiler loads about 3.6k tokens when it runs, and up to ~13k if it reads all its reference files. Until then it costs about 89 tokens; SKILL.md has 666 words of instructions outside code blocks.
Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.
Safety
Auto-check passed
The automated check found no risky patterns in SKILL.md.
Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.
SKILL.md
The full file from PaddlePaddle/Paddle at commit 5434c21, republished under its Apache-2.0 licence (© PaddlePaddle). 666 words, ~3,619 tokens.
Download SKILL.mdSave it as
.claude/skills/paddle-design-compiler/SKILL.md (or your agent's skills folder). This skill also uses 6 other files; get the full folder from GitHub.- name
- paddle-design-compiler
- description
- Use when working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate representation, CINN for fused CUDA kernel generation, operator decomposition (Prim), or the end-to-end flow from Python eager code to optimized GPU execution.
Paddle 3.0 编译器全链路
Paddle 3.0 的编译器体系通过 SOT → PIR → CINN 三阶段流水线，将用户的动态图 Python 代码编译为高性能 GPU Kernel，实现「动态图编写、编译器加速」的开发体验。
全链路概览
用户 Python 代码（动态图 eager mode）
│
▼ Stage 0: SOT 图捕获
PEP 523 eval_frame 拦截 → OpcodeExecutor 字节码模拟
→ FunctionGraph / StatementIR
→ paddle.jit.to_static(full_graph=True) 编译子图
→ pir::Program
│
▼ Stage 1: PIR Pass 优化
pir::Program（SSA 形式的 pd_op.* 算子图）
│ ├── ShapeOptimizationPass（InferSymbolicShape 动态 shape 符号推导）
│ ├── 组合算子分解 (DecompInterface → primitive operators)
│ └── 通用 Pass 优化（常量折叠、死代码消除等）
│
▼ Stage 2: CINN 编译
├── PdOpToCinnOpPass / PdOpToDynamicShapeCinnOpPass（算子映射）
├── add_cinn_pass → cinn_op.group（算子融合）
├── OpLower（Compute + Schedule）→ LoweredFunc
├── CodeGenCUDA_Dev → CUDA source → NVRTC → CUfunction
├── CompilationCache（编译缓存，相同子图复用已编译 Kernel）
│
▼ Stage 3: 执行
PirInterpreter 调度 → CinnJitInstruction → cuLaunchKernelSOT → PIR 的衔接：SOT 捕获的 StatementIR 被包装为 Python 函数后，通过 paddle.jit.to_static(full_graph=True) 再次走 AST Transformer 路径编译为 pir::Program（参见 python/paddle/jit/sot/symbolic/compile_cache.py）。这意味着 SOT 负责"图捕获"，而 to_static 负责"图编译"。
SOT（Symbolic Opcode Translator）
SOT 是 Paddle 3.0 的动转静前端，在 Python VM 字节码层面拦截和模拟执行用户代码，精确捕获 Tensor 计算子图。相比旧的 AST Transformer 方案，SOT 能处理 numpy/Tensor 互操作、动态控制流、第三方库调用等复杂场景。
核心机制
Python Frame
│
▼ PEP 523 eval_frame 拦截
PyInterpreterState.eval_frame
│
▼
OpcodeExecutor（模拟 Python VM 字节码执行）
│ ├─ Variable 体系 (TensorVariable, ConstantVariable, ContainerVariable, ...)
│ ├─ Tracker 追踪来源 → 生成 Guard（缓存有效性校验）
│ └─ SideEffect 记录副作用（全局变量修改、可变对象修改）
▼
FunctionGraph / StatementIR（记录算子调用）
│
├─ 无 fallback: 完整子图 → to_static(full_graph=True) → pir::Program
└─ fallback: 子图切分 → 可静态化部分编译 + 不可静态化部分 Python 执行| 组件 | 说明 |
|---|---|
| OpcodeExecutor | 模拟 Python VM 执行字节码，不真正计算，而是追踪 Tensor 操作 |
| Variable 体系 | 将 Python 对象包装为 Variable（TensorVariable / ConstantVariable / ContainerVariable / CallableVariable） |
| Tracker | 记录 Variable 来源（provenance），形成 DAG，用于生成 Guard |
| Guard | Callable[[FrameType], bool]，判断当前帧输入是否满足编译假设，用于缓存命中判断 |
| FunctionGraph | 收集 Tensor 相关操作，输出 StatementIR |
| StatementIR | 4 种语句类型（call_api / call_method / call_sir / call_layer），最终经 to_static(full_graph=True) 编译为 Program |
| SideEffect | 记录并回放模拟执行中对全局变量和可变对象的修改，保证语义等价 |
| OpcodeInlineExecutor | 跨函数边界模拟执行，实现子图跨函数融合 |
Fallback 场景
| 缩写 | 全称 | 场景 |
|---|---|---|
| DDCF | Data-Dependent Control Flow | 控制流条件依赖 Tensor 值（如 if x.sum() > 0） |
| UNSPS | Unsupported Simulation | 无法模拟的 Python 操作（如某些 C 扩展、.numpy()） |
| CDBL | Custom Blacklist | 用户或框架标记的不转换函数（如产生 -1 shape 的算子） |
| UNIMP | Unimplemented Opcode | 尚未实现模拟的字节码指令 |
Fallback 是安全兜底：任何无法处理的情况退化为部分子图编译 + 部分 Python 执行，不会导致报错。
使用方式
# full_graph=False（默认）：启用 SOT 模式（字节码级别捕获 + 自动 fallback）
# full_graph=True：使用传统 AST Transformer（要求整图可转）
net = paddle.jit.to_static(net) # 默认 full_graph=False，即 SOT 模式
output = net(x)PIR（Paddle Intermediate Representation）
PIR 是 Paddle 3.0 的统一中间表示，采用 MLIR 风格的 SSA 设计，替代旧的 ProgramDesc/OpDesc 体系。
核心概念
| 概念 | 关键类 | 说明 |
|---|---|---|
| Type | TypeID / AbstractType / TypeStorage / Type | 统一类型系统：TypeID 用 static 变量地址做唯一标识，Type 本质是指向 TypeStorage 的指针，相等性通过指针比较 O(1) |
| Value | ValueImpl / OpResultImpl / OpOperandImpl | SSA 值系统：OpResult 是算子输出（inline 0-5 / out-of-line），OpOperand 通过侵入式双向链表管理 use-chain |
| Operation | Operation（连续内存布局） | 核心执行单元：`[OutOfLineResults |
| Block/Region | Block / Region | Block 持有 Operation 列表 + BlockArgument + terminator；Region 是 Block 的容器，约束 Value 作用域 |
| Dialect | BuiltinDialect / PaddleDialect / CinnDialect | 模块化容器：聚合一组 Type、Attribute、Op 定义，支持独立注册与扩展 |
| Trait/Interface | OpTraitBase / concept-model 多态 | Trait 是静态标记，Interface 通过 concept-model 实现多态分派，替代 C++ 虚函数 |
核心 Dialect
| Dialect | 职责 | 典型内容 |
|---|---|---|
BuiltinDialect | PIR 内置基础类型 | Float32Type, Int64Type, VectorType, DenseTensorType |
PaddleDialect | Paddle 算子定义 | pd_op.matmul, pd_op.relu, pd_op.conv2d |
CinnDialect | CINN 编译器专用 | cinn_op.group, cinn_op.yield, cinn_op.generate_shape |
ControlFlowDialect | 控制流辅助 | cf.yield, cf.stack_create, cf.tuple_push, cf.tuple_pop |
PaddleDialect（控制流部分） | 控制流算子 | pd_op.if, pd_op.while |
PIR Program 结构
Program
├── weights: unordered_map<string, shared_ptr<Parameter>>
└── ModuleOp (顶层 Operation)
└── Region[0]
└── Block[0]
├── builtin.parameter("w") → %0 (从权重表读取参数)
├── pd_op.matmul(%input, %0) → %1
├── pd_op.if(%cond) → %2
│ ├── Region[0] (then)
│ │ └── Block[0]: pd_op.relu(%1) → cf.yield
│ └── Region[1] (else)
│ └── Block[0]: pd_op.tanh(%1) → cf.yield
└── builtin.set_parameter(%2, "out")组合算子分解（Prim）
将高层算子分解为基础算子（primitive operators），降低编译器 / 分布式 / 新硬件适配成本：
- 前向分解：
DecompInterface→call_decomp_rule()→composite.h - 反向分解（VJP）：两条路径——
VjpInterface经call_vjp()处理前向 op 的反向；DecompVjpInterface经call_decomp_vjp()分解反向 op。规则实现均在details.h - CustomVJP：为 sigmoid、log_softmax 等数值敏感算子提供手写反向
PIR Pass 框架
PIR 提供 MLIR 风格的 Pass 基础设施，用于图优化：
Pass：单个优化 Pass 基类，通过Run(Operation*)执行PassManager：管理 Pass 执行顺序，支持嵌套 PipelinePatternRewritePass：基于 Pattern Matching 的重写 Pass，通过RewritePattern定义匹配和替换规则
CINN 编译与执行
CINN（Compiler Infrastructure for Neural Networks）将 PIR Program 中的算子子图编译为高性能 CUDA Kernel，由 PirInterpreter 调度执行。当前默认走动态 shape主线。
编译流水线（含动态 shape）
PIR Program (pd_op.*)
│
▼ Stage 1: Frontend（前端）
├── ShapeOptimizationPass（InferSymbolicShape 符号推导）
├── PdOpToCinnOpPass / PdOpToDynamicShapeCinnOpPass（算子映射）
├── add_broadcast_to_elementwise_pass（显式 broadcast 插入）
└── add_cinn_pass → cinn_op.group（按 OpPatternKind 融合）
│
▼ Stage 2: Lowering（后端下降）
├── PirCompiler → CompilationTask（per GroupOp）
│ ├── CompilationCache 查询（命中则跳过编译）
│ ├── OpLower：Compute → AST IR → Schedule
│ ├── DynamicShapeGroupScheduler（动态 shape 调度）
│ └── LowerToAstVec → LoweredFunc
│
▼ Stage 3: CodeGen（代码生成）
├── ir::Module → CodeGenCUDA_Dev → CUDA __global__ source
└── nvrtc::Compiler → PTX → cubin → CUfunction
│
▼ Stage 4: Execution（执行）
└── cinn_runtime.jit_kernel (CINNKernelInfo: fn_ptr + symbol_args_map)
└── CinnJitInstruction → cuLaunchKernelOpPatternKind 融合规则
| Kind | 含义 | 典型算子 |
|---|---|---|
kElementWise | 逐元素计算 | relu, add, multiply |
kBroadcast | 含广播语义 | broadcast_to |
kInjective | 单射映射 | reshape, transpose, slice |
kReduction | 规约操作 | reduce_sum, reduce_max |
kOutFusible | 规约但输出可继续融合 | softmax 中间步骤 |
kNonFusible | 不可融合 | custom_call, sort |
Group-level Schedule（DynamicShapeGroupScheduler）
| 步骤 | 说明 |
|---|---|
DoLoopAlignment | 对齐各算子的循环范围 |
DoComputeInline | 将简单计算内联到消费者 |
OptimizeReduction | 优化规约算子的并行策略 |
DoHorizontalLoopFusion | 水平融合：合并独立的并行循环 |
DoVerticalLoopFusion | 垂直融合：合并生产者-消费者循环 |
BindCudaAxis | 绑定循环到 CUDA threadIdx/blockIdx |
AllocateStorage | 分配 shared memory 和 local buffer |
编译缓存（CompilationCache）
CINN 对已编译的 GroupOp 结果进行缓存（基于 FusionInfo hash），相同结构的子图可直接复用已编译的 Kernel，避免重复编译开销。
执行（PirInterpreter）
编译完成的 Kernel 最终由 PirInterpreter 调度执行：
StandaloneExecutor
└─ PirInterpreter (per Job)
│
├─ Build（首次 Run，结果缓存）
│ ├── 为每个 Op 构建 Instruction（Kernel 选择 + 数据传输插入）
│ ├── 构建算子依赖 DAG → 传递性边消除
│ ├── PirStreamAnalyzer 流调度分类（direct / event / sync）
│ └── Variable 引用计数 → GC 生命周期管理
│
└─ Scheduling（每次 Run）
├── dep_count=0 的 Instruction 推入 work queue
├── 线程池并行派发 → kernel launch
├── 跨 stream 同步：cudaEventRecord + cudaEventWait
└── ref_count=0 时回收 Variable 内存CINN 编译产物通过 CinnJitInstruction 执行：从 CINNKernelInfo 获取 fn_ptr，收集输入输出 device pointer，调用 cuLaunchKernel。非 CINN 算子则通过 PHI Kernel 常规路径执行。
Show full SKILL.md (261 more words)Show less
调试速查
| 场景 | 应关注的文件 |
|---|---|
| SOT 捕获失败 / fallback 过多 | python/paddle/jit/sot/opcode_translator/executor/opcode_executor.py — 检查未支持的 opcode |
| SOT SIR 到 Program 编译失败 | python/paddle/jit/sot/symbolic/compile_cache.py — to_static(full_graph=True) 环节 |
| PIR 动态 shape 推导错误 | paddle/pir/src/dialect/shape/transforms/shape_optimization_pass.cc |
| CINN 融合策略问题 | paddle/cinn/hlir/dialect/operator/transforms/add_cinn_pass.cc |
| CINN 动态 shape 算子映射 | paddle/cinn/hlir/dialect/operator/transforms/pd_to_cinn_pass.cc — PdOpToDynamicShapeCinnOpPass |
| CINN 编译缓存命中 / 未命中 | paddle/cinn/hlir/framework/pir/compilation_cache.cc |
| CINN Schedule 调试 | paddle/cinn/ir/group_schedule/dy_shape_group_scheduler.cc |
| CINN CodeGen CUDA 源码 | paddle/cinn/backends/codegen_cuda_dev.cc |
| 执行器 Kernel 启动 | paddle/fluid/framework/new_executor/pir_interpreter.cc |
| 执行器依赖分析 / 调度 | paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc |
| 执行器 Variable 内存泄漏 | paddle/fluid/framework/new_executor/garbage_collector/ |
什么场景看什么文件
| 场景 | 参考文档 |
|---|---|
| SOT 架构设计（eval_frame / OpcodeExecutor / Guard / Fallback） | references/sot-design.md |
| PIR 类型系统、Dialect、Trait/Interface 设计 | references/pir-basics.md |
| PIR Program/Value/Operation 内存结构、ProgramTranslator | references/pir-program.md |
| CINN 从 GroupOp 到 CUDA Kernel 的完整编译流程 | references/cinn-pipeline.md |
| PIR 控制流（IfOp/WhileOp）、反向 Stack 机制 | references/control-flow.md |
| PIR 执行器（PirInterpreter）、Instruction 调度、Stream 分析、GC | references/executor.md |
源码入口
SOT
| 模块 | 路径 |
|---|---|
| to_static 入口（full_graph 分发） | python/paddle/jit/api.py |
| eval_frame 入口 | python/paddle/jit/sot/opcode_translator/eval_frame_callback.py |
| OpcodeExecutor | python/paddle/jit/sot/opcode_translator/executor/opcode_executor.py |
| OpcodeInlineExecutor | python/paddle/jit/sot/opcode_translator/executor/opcode_inline_executor.py |
| Variable 体系 | python/paddle/jit/sot/opcode_translator/executor/variables/ |
| Tracker | python/paddle/jit/sot/opcode_translator/executor/tracker.py |
| Guard | python/paddle/jit/sot/opcode_translator/executor/guard.py |
| FunctionGraph | python/paddle/jit/sot/opcode_translator/executor/function_graph.py |
| StatementIR | python/paddle/jit/sot/symbolic/statement_ir.py |
| SIR 编译缓存 | python/paddle/jit/sot/symbolic/compile_cache.py |
| SideEffect | python/paddle/jit/sot/opcode_translator/executor/side_effects.py |
| 符号 Shape 推导 | python/paddle/jit/sot/symbolic_shape/ |
PIR
| 模块 | 路径 |
|---|---|
| PIR 核心 | paddle/pir/include/core/ — type.h, value.h, operation.h, block.h, program.h |
| IRContext / StorageManager | paddle/pir/src/core/ir_context.cc, storage_manager.cc |
| Dialect 基类 | paddle/pir/include/core/dialect.h |
| PaddleDialect | paddle/fluid/pir/dialect/operator/ir/op_dialect.h |
| 控制流 Dialect | paddle/pir/include/dialect/control_flow/ir/cf_op.h, cf_type.h |
| 控制流 Op 实现 | paddle/fluid/pir/dialect/operator/ir/control_flow_op.h |
| Shape Dialect | paddle/pir/include/dialect/shape/ |
| ShapeOptimizationPass | paddle/pir/src/dialect/shape/transforms/shape_optimization_pass.cc |
| InferSymbolicShape 接口 | paddle/pir/include/dialect/shape/interface/infer_symbolic_shape/ |
| Pass 框架 | paddle/pir/include/pass/pass.h, pass_manager.h |
| Pattern Rewrite | paddle/pir/include/pattern_rewrite/pattern_match.h |
| DecompInterface（Prim 前向分解接口） | paddle/fluid/pir/dialect/operator/interface/decomp.h |
组合算子（Prim）
| 模块 | 路径 |
|---|---|
| 前向分解规则 | paddle/fluid/primitive/decomp_rule/decomp_rule/composite.h |
| 反向分解规则（VJP） | paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h |
| 分解调度入口 | paddle/fluid/primitive/base/decomp_trans.cc |
| Primitive 基础算子 | paddle/fluid/primitive/primitive/primitive.h |
| VJP 接口 | paddle/fluid/primitive/vjp_interface/vjp.h |
| Backend 适配 | paddle/fluid/primitive/backend/backend.h |
CINN
| 模块 | 路径 |
|---|---|
| CINN 总入口 Pass | paddle/cinn/hlir/dialect/operator/transforms/add_cinn_pass.cc |
| 算子映射（含动态 shape） | paddle/cinn/hlir/dialect/operator/transforms/pd_to_cinn_pass.cc |
| 算子融合 | paddle/cinn/hlir/dialect/operator/transforms/cinn_group_cluster_pass.cc |
| PirCompiler | paddle/cinn/hlir/framework/pir_compiler.cc |
| OpLower 实现 | paddle/cinn/hlir/framework/pir/op_lowering_impl.cc |
| 编译任务 | paddle/cinn/hlir/framework/pir/compilation_task.cc |
| 编译缓存 | paddle/cinn/hlir/framework/pir/compilation_cache.cc |
| DynamicShapeGroupScheduler | paddle/cinn/ir/group_schedule/dy_shape_group_scheduler.cc |
| CodeGen | paddle/cinn/backends/codegen_cuda_dev.cc |
| NVRTC 编译 | paddle/cinn/backends/nvrtc/nvrtc_util.cc |
| CINNKernelInfo 定义 | paddle/cinn/hlir/framework/pir/utils.h |
| JitKernelOp 定义 | paddle/cinn/hlir/dialect/runtime/ir/jit_kernel_op.h |
| AST IR 节点 | paddle/cinn/ir/ |
| Schedule 原语 | paddle/cinn/ir/schedule/ |
执行器（PIR-based）
| 模块 | 路径 |
|---|---|
| Python Executor 入口 | python/paddle/base/executor.py |
| StandaloneExecutor | paddle/fluid/framework/new_executor/standalone_executor.cc |
| InterpreterCore 统一入口 | paddle/fluid/framework/new_executor/interpretercore.cc |
| PirInterpreter | paddle/fluid/framework/new_executor/pir_interpreter.cc |
| ProgramInterpreter（旧 IR 兼容） | paddle/fluid/framework/new_executor/program_interpreter.cc |
| PirStreamAnalyzer | paddle/fluid/framework/new_executor/interpreter/stream_analyzer.cc |
| Instruction 定义 | paddle/fluid/framework/new_executor/instruction/ |
| CinnJitInstruction | paddle/fluid/framework/new_executor/instruction/ |
| Scope（变量容器） | paddle/fluid/framework/scope.cc |
| GC 实现 | paddle/fluid/framework/new_executor/garbage_collector/ |
© PaddlePaddle, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file
Files
SKILL.md and 6 other files (references) in .agents/skills/paddle-design-compiler of PaddlePaddle/Paddle.
- SKILL.md
- references/cinn-pipeline.md
- references/control-flow.md
- references/executor.md
- references/pir-basics.md
- references/pir-program.md
- references/sot-design.md
Open the folder on GitHubat commit 5434c21
Compare with similar skills
Paddle Design Compiler next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Paddle Design Compiler this skillPaddlePaddle/Paddle | 24k | — | ~3.6k | Automated safety check: Pass | Apache-2.0 | |
| Magpie Kernel Evaluatoramd/skills | 406 | — | ~2.3k | Automated safety check: Pass | MIT | |
| Hyperpod Version Checkerawslabs/agent-plugins | 915 | — | ~910 | Automated safety check: Pass | Apache-2.0 | |
| Triton Langmohitmishra786/low-level-dev-skills | 253 | — | ~1.8k | Automated safety check: Pass | MIT | |
| Megatron-LM on SLURMNVIDIA/Megatron-LM | 18k | — | ~1.8k | Automated safety check: Pass | Apache-2.0 | |
| Cuda Index Widthpytorch/pytorch | 104k | — | ~1.6k | Automated safety check: Pass | Custom licence |
Similar skills
- Official
Hyperpod Version Checker
awslabs/agent-plugins
Check and compare software component versions on SageMaker HyperPod cluster nodes - NVIDIA drivers, CUDA toolkit, cuDNN, NCCL, EFA, AWS OFI NCCL, GDRCopy, MPI, Neuron SDK (Trainium/Inferentia)…
Triton Lang
mohitmishra786/low-level-dev-skills
Triton language skill for Python GPU kernel authoring. An agent skill from mohitmishra786/low-level-dev-skills.
- Official
Megatron-LM on SLURM
NVIDIA/Megatron-LM
Shows how to launch distributed Megatron-LM training on a SLURM cluster: sbatch skeleton, torch.distributed.run setup, CUDA_DEVICE_MAX_CONNECTIONS rules and failure diagnosis.
Cuda Index Width
pytorch/pytorch
Choose 32-bit vs 64-bit index math in PyTorch CUDA kernels. An agent skill from pytorch/pytorch.
Graphsignal
graphsignal/graphsignal
Profile AI inference workloads (vLLM, SGLang, TensorRT-LLM, PyTorch, any GPU application) with the Graphsignal profiler and read the results from its local /signals JSON endpoint.
More from PaddlePaddle/Paddle
All 11 skills in this repoPaddle Build
PaddlePaddle/Paddle
A skill your agent uses when needing to compile, rebuild, or install Paddle from source after code changes.
Paddle Design Distributed
PaddlePaddle/Paddle
A skill your agent uses when working with Paddle's distributed training system: understanding parallelism strategies (DP, ZeRO, TP, PP, SP), semi-automatic parallel with ProcessMesh + shardtensor…
Paddle Eager Graph
PaddlePaddle/Paddle
A skill your agent uses when navigating Paddle eager-mode (dynamic graph) source code, tracing forward/backward execution, debugging autograd issues, understanding PyLayer, or investigating…
Paddle Phi Kernel
PaddlePaddle/Paddle
A skill your agent uses when working with Paddle's PHI kernel system: registering new kernels, debugging kernel selection/dispatch, understanding code auto-generation from YAML, or implementing…
Paddle Debug
PaddlePaddle/Paddle
在 Paddle 代码库中定位问题并输出高质量调试报告的专用技能。当遇到以下场景时优先使用：(1) Paddle 框架 bug 调试，(2) 算子实现问题排查，(3) 训练脚本异常诊断，(4) 分布式训练故障定位，(5) CUDA/GPU 相关错误处理，(6) 需要生成结构化调试报告。
Paddle Op Dev
PaddlePaddle/Paddle
PaddlePaddle (飞桨) C++ 算子开发指南。提供从 YAML 配置、InferMeta 函数、Kernel 实现、Python API 封装、单元测试到编译验证的完整算子开发流程指导。在以下场景使用此 skill：(1) 为 Paddle 框架新增 C++ 算子 (2) 修改或调试已有 Paddle 算子 (3) 编写算子的 YAML…
Related
Categories
Questions about Paddle Design Compiler
What does Paddle Design Compiler do?
A skill your agent uses when working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate…. Paddle Design Compiler is an agent skill from PaddlePaddle/Paddle.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture, PIR (Paddle IR) for SSA-based intermediate representation, CINN for fused CUDA kernel generation, operator decomposition (Prim), or the end-to-end flow from Python eager code to optimized GPU execution.
When should I use Paddle Design Compiler?
Paddle Design Compiler fits situations like: working with Paddle 3.0 compiler full pipeline: SOT (Symbolic Opcode Translator) for bytecode-level dy2st graph capture; PIR (Paddle IR) for SSA-based intermediate representation; CINN for fused CUDA kernel generation; operator decomposition (Prim).
How do I install Paddle Design Compiler in Claude Code?
Run `npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a claude-code`. Or copy the skill folder (.agents/skills/paddle-design-compiler in PaddlePaddle/Paddle) into .claude/skills/paddle-design-compiler in your project. Claude Code loads it when a task matches its description.
How do I install Paddle Design Compiler in Codex?
Run `npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a codex`. Or copy the skill folder (.agents/skills/paddle-design-compiler in PaddlePaddle/Paddle) into .agents/skills/paddle-design-compiler in your project. Codex loads it when a task matches its description.
Can I use Paddle Design Compiler in Cursor, Gemini CLI or GitHub Copilot?
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add PaddlePaddle/Paddle --skill paddle-design-compiler -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/paddle-design-compiler, .gemini/skills/paddle-design-compiler, .github/skills/paddle-design-compiler and .opencode/skills/paddle-design-compiler in your project.
What does Paddle Design Compiler need to run?
Going by SKILL.md and its folder, Paddle Design Compiler needs the command-line tools its instructions call (python). Our summary lists: Python 3.
Does Paddle Design Compiler access the network?
SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.
Is Paddle Design Compiler safe to install?
Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.
What licence does Paddle Design Compiler use?
Paddle Design Compiler is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.
How many tokens does Paddle Design Compiler use?
About 3.6k tokens (SKILL.md is roughly 14k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 9.7k tokens, read only when the agent opens those files.
What are the alternatives to Paddle Design Compiler?
Skills that share tags, products or a category with Paddle Design Compiler: Magpie Kernel Evaluator (amd/skills, 406 stars), Hyperpod Version Checker (awslabs/agent-plugins, 915 stars), Triton Lang (mohitmishra786/low-level-dev-skills, 253 stars) and Megatron-LM on SLURM (NVIDIA/Megatron-LM, 18k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
Who maintains Paddle Design Compiler?
PaddlePaddle (a GitHub organization) maintains it in PaddlePaddle/Paddle, which has 24,121 GitHub stars. The repository holds 11 skills in this directory. The repository was last updated on October 9, 2026.
Source: PaddlePaddle/Paddle on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.