generation
Routes GPT2-Chinese prompt-based text generation and batch sample export.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- generation
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes GPT2-Chinese prompt-based text generation and batch sample export.
- disable-model-invocation
- true
02 资源文件 1
- 参考文档references/troubleshooting.md2.2 KB
03 SKILL.md 正文
Generation
Use this sub-skill when the task is about generating text from a checkpoint with generate.py or generate_texts.py.
Read first
../references/workflows.mdfor the end-to-end generation flow.../references/cli-reference.mdfor the exact flags and output paths.../references/model-overview.mdwhen you need to match a checkpoint to a vocabulary bundle.../references/troubleshooting.mdand this sub-skill's troubleshooting file for loop, prompt, and sample-saving issues.../../scripts/check_install.pywhen you want to confirm the environment can instantiate the model and sample a token sequence.
What belongs here
- Prompted generation from a checkpoint directory.
generate.pyfor interactive or single-prompt generation.generate_texts.pyfor batch generation by title list.- Sampling controls such as
--prefix,--length,--temperature,--topk,--topp, and--repetition_penalty. - Sample saving via
--save_samplesor--save_path.
What does not belong here
- Corpus preprocessing and checkpoint creation belong in training.
- Tokenizer or vocabulary creation belongs in tokenization.
- Generic sampling theory without this repo's CLI behavior belongs elsewhere.
How to route a generation request
- Pick the output shape.
- One prompt, one stream of text: use
generate.py. - Multiple titles or articles on disk: use
generate_texts.py.
- One prompt, one stream of text: use
- Choose the checkpoint and vocab pair.
model_pathormodel_pathshould point at a saved checkpoint directory.tokenizer_pathmust match the config and checkpoint vocab size.
- Choose the prompt form.
- The usual checkpoints expect a
[CLS]-prefixed prompt. - The prompt is tokenized before sampling, so keep the tokenizer mode in mind.
- The usual checkpoints expect a
- Choose the decoding behavior.
--fast_patternis the preferred smoke path.--topk,--topp, and--temperatureshape the sampling distribution.
- Choose the save target.
--save_sampleswrites asamples.txtfile.generate_texts.pywrites one text file per generated article.
Common decision points
- If the user wants a quick demo, keep
--batch_size 1and use a short length. - If the user wants file-based batch output, use the title list interface instead of copy-pasting one prompt at a time.
- If the user is asking why the same sample repeats, check
nsamplesversusbatch_sizefirst. - If the task is really about tokenization quality or vocabulary choice, route back to the tokenization sub-skill.
Output expectations
- Printed samples on stdout for the interactive path.
samples.txtwhen--save_samplesis used.<title-index>-<article-index>.txtfiles whengenerate_texts.pyis used.
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).