kali20gakki/msAgent / skills/quantizer/quantization-accuracy-tuning-orchestrator
quantization-accuracy-tuning-orchestrator
VerifiedEnd-to-end automated model quantization and accuracy tuning workflow. Use when user asks for automated model quantization and accuracy tuning, e.g. "自动量化", "量化调优", "一键量化", "精度调优", etc. 当用户给出多卡(≥2 卡)卡号时,量化阶段通过 EP 并行适配(msmodelslim-ep-parallel-adaptation)保证多卡 EP 并行,调优主流程仍在本 Skill。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 25.4 KB |
| references/background_information.md | markdown | 948 B |
| references/output_format.md | markdown | 5.0 KB |
| references/prepare_environment.md | markdown | 2.6 KB |
| references/prepare_model.md | markdown | 9.7 KB |
| references/quantization_tuning.md | markdown | 24.5 KB |
| references/routing.md | markdown | 2.6 KB |
| references/subagent_io_protocol.md | markdown | 10.3 KB |
| references/user_input.md | markdown | 5.9 KB |
| scripts/accuracy_append.py | script | 3.7 KB |
| scripts/accuracy_cleanup.py | script | 2.8 KB |
| scripts/accuracy_lookup.py | script | 3.3 KB |
| scripts/finalize_practice_repo.py | script | 3.7 KB |
| scripts/history_append.py | script | 3.1 KB |
| scripts/history_clear.py | script | 2.1 KB |
references/prepare_environment.md
View on GitHub# 环境准备
**Load when:** 进入量化配置调优前,需确认硬件、运行环境就绪。
## msmodelslim 安装(强制)
量化调优通过 **当前 shell 环境** 调用 `msmodelslim` CLI(`analyze` / `quant` 等),编排层脚本通过 Python 直接 `import msmodelslim`。
1. 确认已安装 Ascend 工具链与 CANN(NPU 场景)。
2. 安装 msmodelslim(版本以项目要求为准):
```bash
pip install msmodelslim
```
3. 验证:
```bash
python -c "import msmodelslim; print('ok')"
```
4. 若模型适配阶段执行过 `bash install.sh`,脚本会在**同一 Python 环境**中读取新 adapter,**无需**重启 MCP 服务;继续调优前确认 `import` 无报错即可。
## Ascend 运行环境变量
原先写在 `config.mcp.json` modelslim 条目中的环境变量,需在**执行量化/评测脚本的 shell** 中配置(或写入用户 `~/.bashrc` / 启动脚本)。典型项包括:
- `ASCEND_HOME_PATH` / `ASCEND_TOOLKIT_HOME` / `ASCEND_AICPU_PATH` / `ASCEND_OPP_PATH`
- `ATB_HOME_PATH` 及 ATB 相关 tuning 变量
- `LD_LIBRARY_PATH`(含 Ascend driver、ATB、opp 等路径)
- `PYTHONPATH`(含 ascend-toolkit `python/site-packages` 与 tbe 路径)
以本机 CANN 安装路径为准;不同机器路径可能不同,**禁止**假设固定目录存在。
## NPU 资源检查(强制)
环境检查时必须执行以下步骤:
1. 运行 `npu-smi info` 获取全部卡信息,确认**总卡数**(不可假设为 8 卡,不同机器卡数不同)
2. 逐卡检查 HBM 占用和进程列表,判断哪些卡空闲(注意:HBM 有少量占用且有进程运行才表示被占用;仅有少量 HBM 占用但无进程通常为驱动占用,可视为空闲)
3. 将空闲卡号列表回显给用户确认后,才确定为量化/评测设备
回显中必须体现:
- 机器总卡数
- 每张卡的 HBM 占用量和进程情况
- 最终选用的卡号列表及用途(量化卡号 / 评测卡号)
## 命令调用方式
- **敏感层分析 / 量化**:通过 `execute` 运行 `msmodelslim analyze`、`msmodelslim quant`(参数见各 Skill 文档)。
- **编排层 / 校验 / 评测**:通过 `execute` 运行 skill 目录下脚本,例如:
```bash
python skills/quantizer/quantization-accuracy-tuning-orchestrator/scripts/history_clear.py --save-path /path/to/workdir
```
编排脚本 **stdout** 输出单行 JSON(`{ok: ...}`);CLI 以 **exit code** 判定成败。
## 环境就绪确认
向用户回显:msmodelslim 可 import、Ascend 环境变量已配置、NPU 卡号已确认。获得用户认可后再进入模型准备阶段。
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |