vision-language-tasks
Guides OFA captioning, VQA, RefCOCO, SNLI-VE, OCR, and ImageNet finetuning and evaluation workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- vision-language-tasks
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides OFA captioning, VQA, RefCOCO, SNLI-VE, OCR, and ImageNet finetuning and evaluation workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/task-reference.md2.1 KB
- 参考文档references/troubleshooting.md1.4 KB
- 参考文档references/workflows.md1.6 KB
- 脚本scripts/coco_caption_eval.py2.3 KB
03 SKILL.md 正文
vision-language-tasks
Use this sub-skill when the user wants to run or adapt OFA's main multimodal understanding workflows: captioning, VQA, visual grounding, visual entailment, OCR, or ImageNet classification.
Trigger phrases
- "Run OFA captioning"
- "VQA beam search or all-candidate eval"
- "RefCOCO training or evaluation"
- "SNLI-VE / visual entailment"
- "OCR or Chinese OCR"
- "ImageNet classification with OFA"
What this sub-skill owns
- caption finetuning and COCO evaluation,
- VQA constrained / all-candidate / beam / zero-shot routes,
- RefCOCO, RefCOCO+, and RefCOCOg workflows,
- SNLI-VE workflows,
- OCR and ImageNet workflows,
- task-specific selected-column and checkpoint guidance.
What it excludes
- text-to-image generation ->
image-generation, - Gigaword and GLUE ->
language-tasks, - pretraining ->
pretraining, - generic launch mechanics ->
setup-and-command-building.
Read these files
- references/workflows.md for the end-to-end workflow shapes.
- references/task-reference.md for task-specific columns, checkpoints, and metrics.
- references/troubleshooting.md for common workflow failures.
- scripts/coco_caption_eval.py for COCO caption metric evaluation.
Typical workflow
- Identify the task family and split.
- Confirm the correct
selected_colsand checkpoint. - Validate the input file with
data-formatswhen the row shape is uncertain. - Render the command or read the recipe for the exact launch shape.
- Use the adapted metric helper if the workflow needs offline scoring.
Notes
- The task family usually determines the selected columns.
- Many commands differ only by the task name and the checkpoint path.
- VQA has multiple evaluation modes; choose beam or all-candidate explicitly.
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).