inference
Load Donut checkpoints, build prompts, run single-image inference, convert JSON tokens, and launch the Gradio demo.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- inference
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Load Donut checkpoints, build prompts, run single-image inference, convert JSON tokens, and launch the Gradio demo.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/demo-and-prediction.md6.3 KB
- 参考文档references/troubleshooting.md4.5 KB
- 脚本scripts/launch_demo.py3.8 KB
- 脚本scripts/run_inference.py5.5 KB
03 SKILL.md 正文
Inference
Use this sub-skill for one-image Donut inference, prompt selection, JSON token round-trips, and the Gradio demo.
Route First
- Start with
references/demo-and-prediction.mdfor command recipes, prompt templates, image preprocessing behavior, output shapes, and token round-trips. - Use
scripts/run_inference.pyfor CLI inference and prompt-variant comparison. - Use
scripts/launch_demo.pyfor the Gradio demo launcher. - Use
references/troubleshooting.mdfor model loading, download, prompt-token, CPU/CUDA, and Gradio failures. - If you need dataset scoring, accuracy, or evaluation over a dataset, route to
../training/SKILL.md. - If the image came from synthetic generation and the user is actually asking about synthetic-data creation, route to
../synthdog/SKILL.mdinstead. - For the shared package API and cross-cutting troubleshooting, use the parent-tree references (
../../references/api-reference.mdand../../references/troubleshooting.md) once the root skill is in place.
Boundaries
- Include:
DonutConfig,DonutModel,from_pretrained,inference,json2token,token2json, prompt/task selection, image preprocessing behavior, CPU/CUDA inference behavior, demo launcher. - Exclude: training loops, Lightning trainer setup, dataset evaluation metrics, SynthDoG generation.
Quick Paths
python scripts/run_inference.py --helppython scripts/launch_demo.py --help
What to do
- Load a local checkpoint or Hub checkpoint with the Donut runtime API.
- Build the exact task prompt the checkpoint expects.
- Run one image at a time, or compare several prompt variants against the same image.
- Convert structured JSON to Donut tokens and back when you need to inspect prompt/output formatting.
- Start the Gradio demo with a local model and a sample image if available.
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).