ascend-robotfeat: submit initial agent skill knowledge of task workflow guide

文件	最后提交记录	最后更新时间
knowledge	feat: submit initial agent skill knowledge of task workflow guide Co-authored-by: hhhzhuyizhi<zhuyanru2@huawei.com> # message auto-generated for no-merge-commit merge: !2579 merge master into master feat: submit initial agent skill knowledge of task workflow guide Created-by: hhhzhuyizhi Commit-by: hhhzhuyizhi Merged-by: ascend-robot Description: ## What this PR does / why we need it? Add new agent skill knowledge to guide right task work flow. ## Does this PR introduce any user-facing change? Not involved. ## How was this patch tested? Not involved. See merge request: Ascend/MindSpeed-MM!2579	6 天前
skills	refactor: refactor agent skill readme Co-authored-by: hhhzhuyizhi<zhuyanru2@huawei.com> # message auto-generated for no-merge-commit merge: !2574 merge master into master refactor: refactor agent skill readme Created-by: hhhzhuyizhi Commit-by: hhhzhuyizhi Merged-by: ascend-robot Description: ## What this PR does / why we need it? Refactor agent skills readme, move skill plan to main readme. ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2574	6 天前
README.md	refactor: refactor agent skill readme Co-authored-by: hhhzhuyizhi<zhuyanru2@huawei.com> # message auto-generated for no-merge-commit merge: !2574 merge master into master refactor: refactor agent skill readme Created-by: hhhzhuyizhi Commit-by: hhhzhuyizhi Merged-by: ascend-robot Description: ## What this PR does / why we need it? Refactor agent skills readme, move skill plan to main readme. ## Does this PR introduce any user-facing change? Please describe whether the PR will result in any user-facing usage changes. If there is related documentation, please specify its path. ## How was this patch tested? Please explain how to verify the correctness and effectiveness of this feature, as well as its usage constraints and limitations. See merge request: Ascend/MindSpeed-MM!2574	6 天前
setup_agent.sh	feat: add AI agent scaffold Co-authored-by: hanyyy<hanyue42@huawei.com> # message auto-generated for no-merge-commit merge: !2547 merge mm-skill into master feat: add AI agent scaffold Created-by: vasileone Commit-by: hanyyy Merged-by: ascend-robot Description: ## What this PR does / why we need it? 本 PR 新增 MindSpeed-MM 面向 AI coding agents 的 `.agents` 配置骨架。主要改动包括： - 新增 `.agents/README.md`，作为仓库级 AI coding agents 共享入口。 - 新增 `.agents/knowledge/architecture.md`，说明 MindSpeed-MM 当前的architecture及两套主要训练后端： - 新增 `.agents/skills/README.md`，定义 skill 的规划表和落地。 - 新增 `.agents/setup_agent.sh`，支持将 `.agents` 内容链接到 Codex、Claude、Cursor、Trae 等本地 agent 目录。 ## Does this PR introduce any user-facing change? 否，新增内容主要面向开发者和 AI coding agents。 ## How was this patch tested? 本 PR 为文档和配置骨架变更，未涉及运行时代码修改。 See merge request: Ascend/MindSpeed-MM!2547	8 天前

MindSpeed-MM Agent Configuration

This directory contains shared guidance for AI coding agents working on MindSpeed-MM.

The .agents directory is the single source for reusable agent-facing context. Tool-specific directories such as .codex/, .claude/, .cursor/, or .trae/ can be generated locally from this shared source when needed.

MindSpeed-MM follows the Agent Skills convention for skill layout.

Directory Layout

Path	Purpose
`skills/`	Skill index and implementation conventions.
`knowledge/`	Shared knowledge context for agents.
`setup_agent.sh`	Optional helper for linking `.agents` into local tool-specific directories.

Usage

Link this shared configuration into a local agent directory:

bash .agents/setup_agent.sh codex
bash .agents/setup_agent.sh claude
bash .agents/setup_agent.sh cursor
bash .agents/setup_agent.sh trae

The script also accepts a custom agent name and creates .<agent-name>/ as a local adapter directory. Generated adapter directories are added to .git/info/exclude.

Architecture Summary

MindSpeed-MM supports two main training backend paths. Agents should identify the active backend before changing model code, data code, checkpoint conversion, examples, or tests.

Backend	Primary Entries	Description
MindSpeed Core / Megatron	`mindspeed_mm/training.py`, `mindspeed_mm/pretrain_.py`, `examples//*.sh`	Megatron-style flow using model/data/forward providers and hybrid parallelism.
FSDP2	`mindspeed_mm/fsdp/train/trainer.py`, `mindspeed_mm/config/config_manager.py`, `mindspeed_mm/fsdp/utils/register.py`	YAML-driven flow using plugin registration, `ModelHub`, FSDP2 data builders, and parallel plans.

See knowledge/architecture.md for the agent-facing architecture overview.

Skill Plan

Skill	Domain	Status	Priority	Description
mindspeed-mm-fsdp2-model-only-vlm-migration	Integration	Planned	P0	指导新模型接入 FSDP2 后端，覆盖参考样例、注册、配置、数据字段和端到端验收，当前阶段仅支持vlm迁移。
performance-analysis-report	Optimization	Planned	P0	将 profiling 结果和训练日志整理为瓶颈分析报告与优化建议。
fsdp2-dataset-migration	Integration	Planned	P0	指导新数据集接入 FSDP2 数据链路，覆盖 dataset type、collator 和 batch key。
flops-mfu-analysis	Optimization	Planned	P0	基于模型配置、输入形状和运行指标估算 FLOPs 与 MFU。
fused-operator-optimization	Optimization	Planned	P0	规划 RMSNorm、EP-BMM、ROPE 等融合算子替换及精度性能验证。
npu-environment-setup	Integration	Planned	P1	梳理指定模型在 Ascend/NPU 环境下的依赖、环境变量、安装顺序和最小验证方式。
best-configuration-recommendation	Optimization	Planned	P1	结合模型规模和并行策略，推荐可解释的训练配置组合（EP、TP、CP、FSDP）。
transformers-alignment-gate	Verification	Planned	P1	为 Transformers 版本升级提供对齐检查。
checkpoint-conversion-routing	Integration	Planned	P1	根据源格式、目标格式和模型类型选择合适的权重转换路径并检查关键参数。
minimal-doc-sync	Collaboration	Planned	P2	根据代码变更识别 README、特性文档或 example 文档中的最小同步范围。
pr-description-generation	Collaboration	Planned	P2	根据 diff、测试结果、风险和用户影响生成 PR 描述与评审申请内容。
unit-test-authoring	Verification	Planned	P2	辅助编写符合仓库风格的单元测试

See skills/README.md for the full skill index.