[greyoak111]

组件 / Component

• id: ml-tflite
• 上游 / Upstream: python-tflite-runtime 2.21.0 (Termux main), TensorFlow add.bin test model
• 许可证 / License: Apache-2.0
• 主页 / Homepage: https://www.tensorflow.org/lite

What

TensorFlow Lite CPU inference on the device:

• ml-tflite info — runtime, numpy and bundled model
• ml-tflite inspect model.tflite — input/output tensor names, shapes and dtypes
• Python API: tflite_runtime.interpreter.Interpreter

The component is self-contained: numpy is bundled, so it has no component dependencies. The bundled usr/share/ml-tflite/add.bin is TensorFlow's 544-byte Apache-2.0 test model, used only by the offline selftest and examples. User models are supplied separately.

体积与共用包 / Size and sharing

• 下载 12.2 MB · 装后 38.5 MB · 4 个包（tflite-runtime、numpy、libopenblas + 544 字节测试模型）
• 与 sci-python、pandas 共用 python-numpy、libopenblas（同一 Termux 版本；有意做成自包含，不 requires sci-python）/ shares python-numpy and libopenblas with sci-python and pandas (same Termux version; self-contained on purpose)

Tested

===== 总结 ===== tflite=2.21.0 numpy=… model=PASS inference=PASS

The selftest loads the bundled model, fills the input tensors with zeros, invokes the interpreter and checks every output. Local audit against devkit-2026.10.28: 0 errors; 🔴 for human review because the component fetches an upstream model from GitHub (sha256 verified at build time). The model's source and Apache-2.0 notice are in usr/share/ml-tflite/NOTICE. The staged package passed locally.

检查 / Checklist

• node tools/check-recipes.mjs 通过 / passes
• ci-build.mjs --dry 依赖都能解 / resolves (also built for devkit-2026.10.28)
• 读过组件编写规则 / read the authoring rules
• 描述与 skill 里的版本、能力以实际打进包的为准 / description and skill match what is packaged
• 看过审计报告 / read the audit report (shared packages and risk levels above)
• 第一次运行要下载的：skill 里写明大小与来源、先问用户 / first-run downloads disclosed and ask first (N/A — the component does not download at runtime; user models are provided by the user)
• SKILL.md 不会让 Claude 读取或打印令牌、密码 / never reads or prints secrets
• 对外发布、删除类操作写明先问用户 / asks before public or destructive actions (N/A)
• needs 如实填写 / filled honestly (shizuku, display, gpu all false)
• 不涉及反代、共享或转售账户、破甲、批量操控账号 / not for reverse proxies, shared accounts, jailbreaking or mass-controlled accounts

真机测试 / Device test

• node tools/verify-device.mjs 全过，记录另开 PR 进 verified.json / not run yet; can follow up
• 设备 / Device:
• cccodroid 版本 / version:

🤖 Generated with Claude Code

[github-actions]

组件审计报告

底座 devkit-2026.10.28（3768afcf79a7）· 规则见 tools/audit.mjs 开头。🟢 看报告即可 · 🟡 要读脚本和 skill · 🔴 要逐行审

🔴 ml-tflite 2.21.0-1-r1

下载 12.2 MB · 装后 38.5 MB · 962 个文件 · 4 个包 （第一次发布）

• ⚠️ postinst 里会 pip install：python-tflite-runtime —— 装组件时会去 PyPI 现场拉依赖（版本不定、离线装坏、卸载删不掉）；把缺的纯 Python 依赖写进配方的 pip
• ⚠️ 底座与仓库版本不同（组件按底座的版本用）: ca-certificates 1:2026.10.20-1→1:2026.11.05-1, libexpat 2.9.0→2.10.0, libsqlite 3.54.1→3.54.1-1, python 3.14.8→3.14.10, python-pip 26.2.2-2→26.3, zlib 1.3.2→1.4
• ✅ 不会盖掉底座的文件
• ✅ 21 个 ELF 检查过
• ✅ 每个 ELF 要的库都找得到
• ✅ bin 入口与 selftest 命令都在

人要看的：

• 🟡 SKILL.md 联网下载：L20 如果报 '找不到 numpy'，就是组件没装完整——重装组件，不要 'pip install numpy' 另装一份。；L72 - **numpy 缺失**：组件没装完整，重装本组件；不要 'pip install numpy' 另装一份。
• 🔴 fetch 了仓库外的文件：add.bin ← raw.githubusercontent.com（哈希打包时已核对）
• 🟡 配方自带文件 2 个：usr/bin/ml-tflite、usr/share/ml-tflite/NOTICE
• 🟡 装完要跑包的 postinst：python-numpy、python-tflite-runtime

和别的组件共用的包：octave：libopenblas；sci-python：libopenblas, python-numpy

包清单

包	版本	下载
libopenblas	0.3.34	2.5 MB
python-numpy	2.5.4	3.5 MB
python-tflite-runtime	2.21.0-1	1.5 MB
add.bin	undefined	0.0 MB

CI 实跑 selftest（arm64 + termux-docker）

组件	结果	说明
ml-tflite	✅	退出码 0；最后一行：===== 总结 ===== tflite=2.21.0 numpy=2.5.4 model=PASS inference=PASS

环境同 App 里 Claude 的 Bash 工具（PREFIX / HOME / TMPDIR / termux-exec），没有 GPU、屏幕、Shizuku。通过只说明「装得上、跑得起来」，安卓特有的问题仍要真机。