mindspore-ai

矩阵乘法矩阵乘法 A[M, K] @ B[K, N] = C[M, N]中，大K维度矩阵乘法(K>>M,N)优化：针对M/N较小但K极大(如M=N=256,K=131072)的场景，Split-K切分K维度并行化、Workspace+Reduce替代全局同步，实现显著性能提升

Triton Ascend hard API restrictions and forbidden syntax. MUST-follow rules that apply to every kernel: forbidden control flow (return/break/continue/lambda/while), tensor slice/index restrictions, scalar conversion rules, BLOCK_SIZE upper bound. Violating…

Triton Ascend 性能优化通用策略: BLOCK_SIZE 选择 (1024-2048 for elementwise, must be <65536), grid configuration (use VEC_CORE_NUM / CUBE_CORE_NUM, 2D/3D grid for matmul / conv / reduce, 1D grid + inner loop for elementwise / pointwise), 256B alignment for memory…

通过 adaptive_search 或 evolve 搜索式 workflow 生成优化算子。 后台 silent mode 执行，轮询监控进度。

适用于归约(reduce)类算子和含归约子步骤的复合算子（如归一化）的优化指南。典型算子包括：sum, mean, max, min, prod, argmax, argmin, cumsum, cumprod, softmax, logsoftmax, layernorm, rmsnorm, groupnorm, instancenorm, batchnorm, l1norm, l2norm, frobeniusnorm, var, std, average_pooling, sum_pooling…

CPU C++ 算子核心概念、标准结构模式、KernelBench 代码规范和内嵌扩展方法

ARM CPU 架构性能优化技巧、NEON SIMD 向量化、数值稳定性和调试策略

x64 CPU 架构性能优化技巧、SIMD/AVX 向量化、数值稳定性和调试策略

把开源算法模型适配到 MindSpore Lite 部署管线：按网络结构拆分导出 ONNX、ONNX Runtime 推理验证、ONNX→MindIR 转换、MindSpore Lite 推理实现，并交付文档与常见问题。用户想把某个开源模型迁移到 MSLite 部署时调用。

MindSpore Lite MindIR模型精度问题定位技能。覆盖从CANN Profiling对比分析定位可疑算子、配置算子Dump、到最终对比Dump数据确定精度差异根因算子的全流程。

开源 LLM 模型的多卡张量并行部署：Megatron 风格权重分片导出 → ONNX→MindIR 转换（离线 acl / GE online 双路径）→ 多进程 HCCL 推理。覆盖 1p/2p/4p 三种模式，含 Custom 算子（AllReduce/Scatter/RMSNorm 等）开发与通信机制、300I Duo 与 800I A2 差异、已知限制与修复方案。用户想把模型部署到多卡/多芯 Ascend 推理时调用。

MindSpore Lite（Ascend）模型性能优化总攻略。做基线/profiling、融合算子改写、推理免拷贝、PTQ int8 量化、精度对齐与归档时调用。本文为总览与索引，细化策略见 references/。

MindSpore Lite云侧推理 Ascend 后端离线转换（ONNX → MindIR）与推理部署全流程。覆盖固定 shape、动态分档、纯动态 shape 的转换策略，以及 MindIR 推理验证与部署注意事项。

Build configuration, CMake options, cross-compilation and packaging. Use when building MindSpore Lite, configuring CMake, cross-compiling for ARM/iOS/MCU, packaging release archives, or troubleshooting build errors.

Model conversion pipeline, parser development, optimization passes and quantization. Use when converting models to .ms, writing parser code, implementing optimizer passes, or configuring quantization.

Debugging, unit testing, benchmarking and performance analysis. Use when running gtest, benchmark tools, profiling latency or accuracy, diagnosing operator precision issues, delegate fallback, or memory leaks.

Add or extend HyperParallel unit tests under tests/ut. Use when writing UT, increasing coverage, or mocking distributed/hardware. Not for multi-card ST (see testing.md / distributed-op-testing) or shard-op-specific UT constraints (distributed-op-testing rule).

GitCode fork workflow: commit, push, create/append PR, status, squash, lint/test. Triggers: 帮我提交, create PR, PR 状态, /commit, /create-pr, etc. Origin=fork, upstream=main. Details in references/.

Review HyperParallel changes for distributed correctness, stream sync, memory safety, cross-platform parity, and code quality. Use for /code-review, PR review, or when the user asks to review/check changes.

Implement distributed ops from a confirmed analysis plan; write UT/ST and run until executable tests pass. Goal mode — no step-by-step confirmation.

Drive a red MindSpore-family GitCode PR gate to green: diagnose, /check-pr → /retest, triage failures, patch or escalate, loop until both pr-check-pass and ci-pipeline-passed. Use for 门禁/CI/retest/ check-pr/autofix/把 PR 修绿, or any MindSpore GitCode PR gate…

Recommend DP/FSDP/TP/PP/EP/CP strategy mixes from model + hardware constraints, with memory/comm/compute/bubble estimates. Use for strategy planning, scale-out, OOM, or comparing configs. Not for implementing platform/op code (platform-dev / dist-op-dev) or…

Develop HyperParallel platform abstraction: new Platform APIs, torch/mindspore backends, FSDP/HSDP/Pipeline/Activation Checkpoint, DTensorBase, collectives. Use when changing platform/ or cross-backend features. Not for shard op YAML/impl (use dist-op-dev) or…

Distributed operator analysis. Analyzes operator interfaces provided by the user and outputs a standardized implementation plan. Requires human confirmation before development begins.