# 知识库

**Repository Path**: chen-junhao1234/knowledge-base

## Basic Information

- **Project Name**: 知识库
- **Description**: graphify仓库
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-06
- **Last Updated**: 2026-07-09

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 图与框架知识库

## 知识

### 诊断背景

> 从 `skill/knowledge/background/` 复制过来的诊断系统背景知识文档，供人工阅读。

#### ACLGraph 背景文档

- [ACLGraph 设计与实现](知识/诊断背景/aclgraph/aclgraph-design.md)
- [ACLGraph 分层架构图文字提取](知识/诊断背景/aclgraph/aclgraph-architecture-overview.md)
- [CUDA Graph Launch 收益图文字提取](知识/诊断背景/aclgraph/cuda-graph-launch-benefit.md)
- [Graph Update Task Group 时序图文字提取](知识/诊断背景/aclgraph/graph-update-task-group-flow.md)
- [Multi-Stream Capture Fork-Join 时序图文字提取](知识/诊断背景/aclgraph/multistream-capture-flow.md)
- [Graph Tree 运行时图文字提取](知识/诊断背景/aclgraph/graph-tree-runtime.md)
- [Graph Tree 子图内存池共享图文字提取](知识/诊断背景/aclgraph/graph-tree-mempool-sharing.md)

#### Super Kernel 背景文档

- [ACLGraph + Super Kernel 使用指南](知识/诊断背景/super_kernel/super-kernel-guide.md)
- [Super Kernel 配置选项速查](知识/诊断背景/super_kernel/super-kernel-config-options.md)
- [Super Kernel 调试流程](知识/诊断背景/super_kernel/super-kernel-debug-flow.md)
- [Super Kernel 融合断开排查](知识/诊断背景/super_kernel/super-kernel-fusion-break-guide.md)
- [Super Kernel 盘古配置参考](知识/诊断背景/super_kernel/super-kernel-pangu-config.md)

#### 多档位 / 调试 / 存档

- [多档位编码约束](知识/诊断背景/multi-tier/multi-tier-encoding-constraints.md)
- [FX 图打印指南](知识/诊断背景/debug/fx-graph-printing-guide.md)
- [日志开启方式速查](知识/诊断背景/debug/logging-enablement-guide.md)
- [ACLGraph 通用定位手段速查（存档）](知识/诊断背景/archive/aclgraph-debug-toolkit.md)
- [静态 Kernel 诊断流程（存档）](知识/诊断背景/archive/static-kernel-diagnostic-flow.md)
- [静态 Kernel 控核排查（存档）](知识/诊断背景/archive/static-kernel-core-control.md)
- [Capture 动态 shape 资源耗尽（存档）](知识/诊断背景/archive/capture-dynamic-shape-resource-exhaustion.md)
- [图与算子联调定位手册（存档）](知识/诊断背景/archive/graph-op-debug-playbook.md)

### 入门

- [TorchAir大学习](知识/入门/TorchAir大学习.md)
- [【组内培训】torch_npu](知识/入门/【组内培训】torch_npu.md)

### ACLGraph

- [ACLGraph设计与实现](知识/ACLGraph/ACLGraph设计与实现.md)
- [ACLGraph内存复用方案](知识/ACLGraph/ACLGraph内存复用方案.md)
- [ACLGraph 重捕显存释放与 use_count 0 断言根因分析](知识/ACLGraph/ACLGraph%20重捕显存释放与%20%60use_count%20%200%60%20断言根因分析.md)
- [AclGraph问题定位指导](知识/ACLGraph/AclGraph问题定位指导.md)
- [aclgraph使能常见问题分析](知识/ACLGraph/aclgraph使能常见问题分析.md)
- [aclgraph支持中间节点tensor dump](知识/ACLGraph/aclgraph支持中间节点tensor%20dump.md)
- [A5适配ACLGraph叠加基础特性](知识/ACLGraph/A5适配ACLGraph叠加基础特性.md)
- [ACL Graph in vLLM](知识/ACLGraph/ACL%20Graph%20in%20vLLM.md)
- [vLLM中ACLGraph的应用解析](知识/ACLGraph/vLLM中ACLGraph的应用解析.md)
- [SGLang入图开发总结](知识/ACLGraph/SGLang入图开发总结.md)
- [TorchAir Dump 图功能和代码梳理](知识/ACLGraph/TorchAir%20Dump%20图功能和代码梳理.md)
- [TorchAir 中 ACLGraph 的代码流程梳理](知识/ACLGraph/TorchAir%20中%20ACLGraph%20的代码流程梳理.md)
- [Torchair使用指导一本通2.0 part2](知识/ACLGraph/Torchair使用指导一本通2.0-part2.md)
- [Torchair精度问题定位](知识/ACLGraph/Torchair精度问题定位.md)
- [Trace断图 torch._dynamo.exc.Unsupported](知识/ACLGraph/Trace断图torch._dynamo.exc.Unsupported.md)
- [多ACLGraphWrapper实例下recapture报错](知识/ACLGraph/多ACLGraphWrapper实例下recapture报错.md)
- [被capture禁止的操作](知识/ACLGraph/被capture禁止的操作.md)
- [torch_npu print_npugraph_tensor](知识/ACLGraph/torch_npu-print_npugraph_tensor.md)
- [npugraph dfx](知识/ACLGraph/npugraph%20dfx.md)
- [omni v1 盘古v2 RL训推过程aclgraph recapture内存优化](知识/ACLGraph/omni%20v1%20盘古v2%20RL训推过程aclgraph%20recapture内存优化.md)

### GE

- [模型入图介绍](知识/GE/模型入图介绍.md)
- [[组内培训] 图编译&图加载](知识/GE/%5B组内培训%5D%20图编译%26图加载.md)
- [图执行](知识/GE/图执行.md)
- [Parse](知识/GE/Parse.md)

### torch.compile

- [torch.compile 技术剖析（一）：PyTorch 2.x 编译系统详解](知识/torch.compile/torch.compile%20技术剖析（一）：PyTorch%202.x%20编译系统详解.md)
- [torch.compile 技术剖析（二）：拆解 FX Graph 优化流水线](知识/torch.compile/torch.compile%20技术剖析（二）：拆解%20FX%20Graph%20优化流水线.md)
- [torch.compile 技术剖析（三）：Lowering](知识/torch.compile/torch.compile%20技术剖析（三）：Lowering.md)
- [torch.compile 技术剖析（四）：Scheduler 的融合决策](知识/torch.compile/torch.compile%20技术剖析（四）：Scheduler%20的融合决策.md)
- [torch.compile 技术剖析（五）：Triton Codegen](知识/torch.compile/torch.compile%20技术剖析（五）：Triton%20Codegen.md)
- [torch compile 中dynamo, inductor, aut-autograd 关系](知识/torch.compile/torch%20compile%20中dynamo,inductor,aut-autograd%20关系.md)
- [torch.compile+Inductor+AutoFuse介绍](知识/torch.compile/torch.compile%2BInductor%2BAutoFuse介绍.md)
- [Inductor+ AF 介绍以及适配流程](知识/torch.compile/Inductor%2B%20AF%20介绍以及适配流程.md)
- [torchtitan对接inductor+AF适配](知识/torch.compile/torchtitan对接inductor%2BAF适配.md)
- [读《TORCH.FX PRACTICAL PROGRAM CAPTURE AND TRANSFORMATION FOR DEEP LEARNING IN PYTHON》](知识/torch.compile/读《TORCH.FX%20PRACTICAL%20PROGRAM%20CAPTURE%20AND%20TRANSFORMATION%20FOR%20DEEP%20LEARNING%20IN%20PYTHON》.md)

### Kernel

- [静态kernel报错](知识/Kernel/静态kernel报错.md)
- [max-autotune模式动静态图介绍及典型问题定位指导](知识/Kernel/max-autotune模式动静态图介绍及典型问题定位指导.md)
- [TorchAir-Ge期望模型编成GE静态图实际编译为动态图的分析指导](知识/Kernel/TorchAir-Ge期望模型编成GE静态图实际编译为动态图的分析指导.md)
- [盘古V2模型superkernel、多流、权重预取、控核](知识/Kernel/盘古V2模型superkernel、多流、权重预取、控核.md)

### 算子

- [单算子入图后输出校验和精度对比指导](知识/算子/单算子入图后输出校验和精度对比指导.md)
- [原地操作算子入图指导](知识/算子/原地操作算子入图指导.md)
- [自定义算子支持Torchair图模式开发指导](知识/算子/自定义算子支持Torchair图模式开发指导.md)
- [[SRS] Eager 模式 dump 指定名称的算子](知识/算子/%5BSRS%5D%20Eager%20模式%20dump%20指定名称的算子.md)

### 多流并行

- [图内多流并行](知识/多流并行/图内多流并行.md)
- [多流精度record_stream接口](知识/多流并行/多流精度record_stream接口.md)

### 开发流程

- [代码合入流程](知识/开发流程/代码合入流程.md)
- [代码提交规范](知识/开发流程/代码提交规范.md)
- [案例贡献规范](知识/开发流程/案例贡献规范.md)

## 案例

| 编号 | 标题 | 阶段 | 关键词 |
|------|------|------|--------|
| [001](案例/case-001-RL场景多档位入图报错.md) | RL场景多档位入图报错 | capture/compile | 多档位, AICPU tiling下沉 |
| [002](案例/case-002-FULL_DECODE_ONLY多档位入图报错.md) | FULL_DECODE_ONLY多档位入图报错 | capture | 多档位, key error |
| [003](案例/case-003-capture-size-mismatch.md) | capture size mismatch | capture | size mismatch, cudagraph |
| [004](案例/case-004-npugraph控核多档位精度问题.md) | npugraph控核多档位精度问题 | replay | 控核, 多档位, 精度 |
| [005](案例/case-005-推理分支适配sleep-wakeup报错.md) | 推理分支适配sleep/wakeup报错 | replay | sleep, wakeup, 参数更新 |
| [006](案例/case-006-capture阶段StreamSynchronize报错.md) | capture阶段StreamSynchronize报错 | capture | StreamSynchronize, 多流 |
| [007](案例/case-007-静态kernel-FIA算子性能劣化.md) | 静态kernel FIA算子性能劣化 | replay | 静态kernel, FIA, 性能 |
| [008](案例/case-008-静态kernel-FIA-ATTR编译报错.md) | 静态kernel FIA/ATTR编译报错 | compile | 静态kernel, 可选输出 |
| [009](案例/case-009-静态kernel-非连续stride传入.md) | 静态kernel非连续stride传入 | compile | 静态kernel, stride |
| [010](案例/case-010-静态kernel-CreateView转一维处理.md) | 静态kernel CreateView转一维处理 | compile | 静态kernel, CreateView |
| [011](案例/case-011-静态kernel-算子修改const.md) | 静态kernel算子修改const | compile | 静态kernel, const |
| [012](案例/case-012-静态kernel-自定义算子mhcpre未执行.md) | 静态kernel自定义算子mhcpre未执行 | replay | 静态kernel, double截断 |
| [013](案例/case-013-静态kernel-盘古V2-92B-coredump.md) | 静态kernel盘古V2 92B coredump | compile | 静态kernel, coredump, ValueDepend |
| [014](案例/case-014-静态kernel-aclgraph叠加aicpu下沉精度.md) | 静态kernel aclgraph叠加aicpu下沉精度 | replay | 静态kernel, aicpu下沉, 精度 |
| [016](案例/case-016-superkernel-静态kernel叠加sk编译launchKernel.md) | superkernel静态kernel叠加sk编译 | compile | superkernel, launchKernel |
| [017](案例/case-017-superkernel-matmulV3全核同步.md) | superkernel matmulV3全核同步 | compile | superkernel, 全核同步 |
| [018](案例/case-018-superkernel-DCCI性能优化.md) | superkernel DCCI性能优化 | replay | superkernel, DCCI, 性能 |
| [019](案例/case-019-superkernel-融合断开死锁.md) | superkernel融合断开死锁 | compile | superkernel, 死锁 |
| [020](案例/case-020-superkernel-ScatterBlockUpdate全核同步.md) | superkernel ScatterBlockUpdate全核同步 | compile | superkernel, 全核同步 |
| [021](案例/case-021-capture-force-eager-配置写入失败.md) | capture force_eager 配置写入失败 | capture | force_eager, CompilerConfig schema 不兼容, CANN 26.1 |
| [022](案例/case-022-capture-107025-mxfp8多流未join.md) | capture 107025 mxfp8 多流 side stream 未 join | capture | 107025, mxfp8 多流, shared_experts_stream 未 join |
| [023](案例/case-023-capture-aot-mutation-guard-view-dtype.md) | capture aot_autograd 拒绝不同 dtype view 的 in-place mutation | capture | aot_autograd mutation guard, kv_cache.view(int8), opaque custom op |
| [024](案例/case-024-capture-npu-rotary-mul-561103-multitier.md) | capture npu_rotary_mul 多档位失败 | capture | 561103, npu_rotary_mul, tiling 未刷新, graph break |
| [025](案例/case-025-runtime-a5-mte-out-of-range-i64-addr固化.md) | runtime A5 MTE 越界 / i64 tensor 地址固化 | runtime | 507011, MTE out of range, A5, i64 tensor 地址固化, PR !1499 |
| [026](案例/case-026-torch-compile-hsdp-hook-mutation-conflict.md) | torch.compile + HSDP Forward Hook mutation tracking 冲突 | torch_compile | HSDP, mutation tracking, dynamo.disable, fully_shard |
| [027](案例/case-027-torch-compile-orig-mod-keyerror.md) | 手动 compile 子模块致 accelerator.prepare KeyError: '_orig_mod' | torch_compile | KeyError, _orig_mod, has_compiled_regions, OptimizedModule |
| [028](案例/case-028-torch-compile-qwen3next-gdn-inlined-multitier.md) | torch.compile Qwen3Next GDN 被 Dynamo 内联致多档位失败 | torch_compile | ERR01001, gdn_attention, Dynamo 内联, torch.ops.vllm, 多档位 |