model-export
Preflight and run WeNet checkpoint export to TorchScript, ONNX CPU or GPU, IPEX, Horizon BPU, and deployment-ready model artifacts.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- model-export
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Preflight and run WeNet checkpoint export to TorchScript, ONNX CPU or GPU, IPEX, Horizon BPU, and deployment-ready model artifacts.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/export-reference.md3.6 KB
- 参考文档references/troubleshooting.md2.8 KB
- 脚本scripts/check_export_inputs.py3.3 KB
03 SKILL.md 正文
WeNet Model Export
Use this sub-skill when the user has a trained WeNet checkpoint and wants TorchScript/JIT, quantized JIT, ONNX, IPEX, Horizon BPU, or other deployment artifacts.
Start here
-
Confirm the experiment has a compatible
train.yamland checkpoint. If the user provides a model directory, preflight it first:python sub-skills/model-export/scripts/check_export_inputs.py \ --model-dir exp/conformer --mode jit -
Read references/export-reference.md for export modes, command templates, required files, streaming chunk choices, output artifacts, and optional dependencies.
-
Read references/troubleshooting.md when files are missing, ONNX dependencies are absent, streaming parameters are inconsistent, CUDA providers are unavailable, quantized export fails, or a vendor toolchain is missing.
Route by task
- Use ../training-and-decoding/SKILL.md when the user still needs to train, average, or locate checkpoints.
- Use ../runtime-deployment/SKILL.md after export to choose C++/mobile/web/server runtime targets.
- Use ../package-transcription/SKILL.md
when a local model directory is intended only for package
load_model().
Key decisions
- JIT export is the common CPU-safe deployment artifact for the C++ libtorch runtime.
- ONNX CPU export creates separate encoder, CTC, and decoder graphs and checks them with ONNX Runtime CPU when dependencies are installed.
- ONNX GPU, TensorRT, IPEX, Horizon BPU, and vendor exports require matching optional packages and toolchains. Treat parser acceptance as weaker evidence than a completed export and runtime check.
- Streaming export must use chunk settings compatible with the intended runtime
mode; non-streaming
-1/-1export should not be reused as a streaming model.
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).