evaluation
Run and troubleshoot VLMEvalKit evaluations, result reuse, async API mode, and run summaries.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- evaluation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Run and troubleshoot VLMEvalKit evaluations, result reuse, async API mode, and run summaries.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/cli-reference.md9.2 KB
- 参考文档references/results-and-status.md8.1 KB
- 参考文档references/troubleshooting.md9.5 KB
- 参考文档references/workflows.md8.5 KB
- 脚本scripts/run_torchrun.sh3.2 KB
- 脚本scripts/scan_api_failures.py11.2 KB
- 脚本scripts/summarize_runs.py11.9 KB
03 SKILL.md 正文
evaluation
Use this sub-skill when a task is to run, resume, inspect, summarize, or troubleshoot VLMEvalKit evaluations through run.py or vlmutil.
Route first
- Build safe
run.pyandvlmutilcommands with CLI reference. - Choose an execution pattern for image, video, multi-turn, local/API, inference-only, eval-only, async API, and reuse runs with workflows.
- Interpret outputs,
status.json, symlinked latest files, checkpoints, summaries, and failure scans with results and status. - Diagnose common failures with troubleshooting.
- Use scripts/run_torchrun.sh only when GPUs are intentionally visible and
torchrunis desired. - Use scripts/scan_api_failures.py to inspect prediction/evaluation files without depending on the source checkout.
- Use scripts/summarize_runs.py to summarize one or more VLMEvalKit
status.jsonrun directories.
Handle here
python run.py --data ... --model ...andpython run.py --config ...command construction.--mode all|infer|eval,--work-dir,--api-nproc,--retry,--keep-failed,--reuse, and--reuse-auxdecisions.- OpenAI-compatible evaluation commands using
--base-url,--key,--custom-prompt,--video-llm,--local-media,--stream,--api-mode, and--monitor-interval. - Judge configuration with
--judge,--judge-args,--judge-base-url,--judge-key,--judge-api-nproc,--judge-retry, and--judge-timeout. - Result reuse, missing/failed prediction triage, run summaries, and utility/result scans.
- Environment knobs that affect evaluation outputs:
PRED_FORMAT,EVAL_FORMAT,SPLIT_THINK,SKIP_ERR,MMEVAL_ROOT,EVAL_PROXY,LOCAL_LLM,FWD_API,LMUData, andVLMEVALKIT_USE_MODELSCOPE.
Route elsewhere
- Implementing or changing model/API wrappers, prompt adapters, LiteLLM/LMDeploy provider classes, or
supported_VLMentries: use ../model-development/SKILL.md. - Implementing or changing benchmarks, dataset converters, TSV schemas,
build_prompt, orevaluate: use ../benchmark-authoring/SKILL.md.
Verification boundary
Skill construction used README.md, docs/en/Quickstart.md, docs/en/ConfigSystem.md, docs/en/EvalByLMDeploy.md, run.py, vlmeval/inference.py, vlmeval/inference_mt.py, vlmeval/inference_video.py, vlmeval/inference_api.py, vlmeval/tools.py, vlmeval/smp/file.py, vlmeval/smp/status_report.py, tests/test_inference_api.py, and selected source scripts as evidence. Inspection verified package import/CLI surfaces and lightweight native API-pipeline checks. It did not verify live provider API calls, dataset downloads, Gradio services, or large GPU model evaluations.
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).