model-usage
Initialize OpenFlamingo models, load checkpoints, prepare multimodal prompts, run forward/generate, and debug media caching constraints.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- model-usage
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Initialize OpenFlamingo models, load checkpoints, prepare multimodal prompts, run forward/generate, and debug media caching constraints.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md8.5 KB
- 参考文档references/generation-workflows.md9.2 KB
- 参考文档references/troubleshooting.md9.3 KB
- 脚本scripts/validate_generation_inputs.py7.4 KB
03 SKILL.md 正文
OpenFlamingo Model Usage
Use this sub-skill when a task requires OpenFlamingo model construction, checkpoint loading, prompt/media preprocessing, generation, log-likelihood style forward() calls, media caching, or architecture debugging.
Do not start full model downloads, benchmark datasets, training, or evaluation unless the caller explicitly provides the required local files, network permission, hardware budget, and time budget. The safe default is to validate inputs and prepare code that works with a caller-supplied local cache or checkpoint.
Route by task
- Need API signatures or object relationships: read references/api-reference.md.
- Need an offline/local-cache generation recipe: read references/generation-workflows.md.
- Need to fix token, tensor, checkpoint, cache, decoder-layer, or import errors: read references/troubleshooting.md.
- Need a deterministic preflight check without downloads: run
scripts/validate_generation_inputs.py.
Critical invariants
- Public imports are
from open_flamingo import create_model_and_transforms, Flamingo. - The package/distribution identity is
open_flamingo2.0.1. vision_xforFlamingo.forward()andFlamingo.generate()isB x T_img x F x C x H x W; current OpenFlamingo supportsF=1only.- Text prompts must use the exact special tokens
<image>and<|endofchunk|>. Each image/media item should have a matching<image>token in the prompt for the same example. - For generation, set
tokenizer.padding_side = "left"before tokenization and passattention_mask. generate(num_beams=N)internally repeatsvision_xalong batch dimension whenN > 1; do not pre-repeat the media batch for beam search.cache_media()is for repeatedforward()/classification-style scoring, not forgenerate(). After cached calls, calluncache_media().
Safe input validation
From this sub-skill directory, validate a three-image captioning prompt without importing OpenFlamingo or downloading anything:
python scripts/validate_generation_inputs.py \
--batch-size 1 \
--num-media 3 \
--num-frames 1 \
--channels 3 \
--height 224 \
--width 224 \
--prompt '<image>An image of two cats.<|endofchunk|><image>An image of a sink.<|endofchunk|><image>An image of'
If the script exits nonzero, correct the reported shape/token issue before attempting model execution.
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).