conversion-and-deployment
Plan PINTO_model_zoo model conversion, quantization, and backend deployment workflows from bundled catalog evidence and local script inspection.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- conversion-and-deployment
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Plan PINTO_model_zoo model conversion, quantization, and backend deployment workflows from bundled catalog evidence and local script inspection.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/conversion-recipes.md10.1 KB
- 参考文档references/deployment-backends.md5.6 KB
- 参考文档references/troubleshooting.md5.8 KB
- 脚本scripts/classify_conversion_script.py11.6 KB
03 SKILL.md 正文
conversion-and-deployment
Use this sub-skill when the user asks how to convert, quantize, or deploy a PINTO_model_zoo model artifact across ONNX, TensorFlow/SavedModel, TFLite, OpenVINO, CoreML, TensorFlow.js, TF-TRT, or EdgeTPU-oriented backends.
Typical triggers:
- "convert this ONNX model to TFLite/OpenVINO/CoreML/TFJS"
- "make an INT8/full-integer/FP16/TFLite/EdgeTPU artifact"
- "understand this quantization or conversion script"
- "deploy this model on Raspberry Pi, OpenVINO CPU, browser TFJS, TF-TRT, or EdgeTPU"
- "why do I need calibration data or a representative dataset?"
Route before planning
- If the user has not selected a model, route model search and format
availability to
model-catalogand the bundled catalog at../../references/model-catalog.json. - If a needed artifact is not present locally, route download/script safety and
license checks to
model-acquisitionbefore proposing commands that fetch or unpack files. - If the user wants to run inference, benchmark, or smoke-test an artifact,
route execution planning to
inference-demosafter this skill identifies the target backend and artifact constraints. - Stay in this skill for conversion-chain selection, quantization requirements, calibration gates, backend deployment preflight, and conversion-script diagnosis.
Operating workflow
- Pin the source and target. Record current artifact family, desired target family, precision goal, input shape/layout/color order, deployment hardware, and whether network/dataset/hardware use is approved.
- Prefer an existing zoo artifact. If the catalog already lists the target format flag for the selected model, advise using that artifact before rebuilding it. Treat catalog flags as availability evidence, not proof that a shallow checkout already contains the files.
- Inspect local scripts without executing them. Use
scripts/classify_conversion_script.pyon candidateconvert*.sh,*quant*.py,*tflite*.py,*openvino*.py,*coreml*.py, or similar files in the selected model directory. The helper reports conversion families, representative-dataset requirements, and dataset/network/hardware risks. - Choose a recipe. Use
references/conversion-recipes.mdfor format-family constraints, TFLite quantization modes, ONNX/TensorFlow/OpenVINO/CoreML/TFJS/ TF-TRT chains, and calibration stop conditions. - Preflight the backend. Use
references/deployment-backends.mdfor Raspberry Pi/TFLite, EdgeTPU, OpenVINO CPU, browser TFJS/WebGL, TF-TRT/GPU, and CoreML deployment gates. - Stop instead of guessing when required calibration data, network access, large datasets, proprietary hardware, GPU/EdgeTPU devices, or license approval is absent. Provide the smallest safe next check and mark unverified hardware claims explicitly.
Expected answer shape
For user-facing plans, return:
- selected source artifact and target artifact family;
- whether a catalog artifact already appears available;
- conversion chain with required tools/environments named only at the family level unless the selected folder supplies exact scripts;
- calibration or representative-dataset requirements;
- backend preflight and stop gates;
- smoke-test handoff to
inference-demoswhen execution is requested; - open risks from
references/troubleshooting.md.
Owned failure modes
This sub-skill owns troubleshooting for conversion, quantization, calibration,
and backend deployment errors. Use references/troubleshooting.md when symptoms
include missing converter imports, unsupported TFLite/EdgeTPU ops, representative
sample mismatch, OpenVINO XML/BIN mismatch, browser TFJS loading issues, TF-TRT
GPU/runtime mismatch, or hardware-only verification claims.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).