architectures-training
Plan and sanity-check UniLM architecture research, long-context/distributed training, decoding acceleration, PFPO/ReSA math evaluation, and credential-safe offline workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- architectures-training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Plan and sanity-check UniLM architecture research, long-context/distributed training, decoding acceleration, PFPO/ReSA math evaluation, and credential-safe offline workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/distributed-training.md6.2 KB
- 参考文档references/troubleshooting.md6.1 KB
- 参考文档references/workflows.md12.7 KB
- 脚本scripts/check_training_plan.py14.8 KB
03 SKILL.md 正文
Architectures and Training
Use this sub-skill for low-level UniLM umbrella research tasks that involve Diff-Transformer, YOCO, DeepNet, LongNet/LongViT, RetNet, X-MoE, BitNet, aggressive decoding, PFPO, ReSA, or distributed launch planning. It is a planning and safety skill: it helps construct, review, and troubleshoot commands without running expensive training, long-context evaluation, downloads, credentialed callers, or GPU-only kernels by default.
Do not use this sub-skill for ordinary UniLM or s2s fine-tuning; route those to ../language-seq2seq/SKILL.md. Route E5/SimLM retrieval training to ../embeddings-retrieval/SKILL.md, and route Kosmos, TextDiffuser, audio, document, or multimodal inference/training to the relevant multimodal or vision sub-skill.
Start Here
- Classify the request: YOCO long-context, architecture module explanation, decoding acceleration, PFPO/ReSA math evaluation, or distributed training review.
- Read
references/workflows.mdfor concrete command templates, required inputs, data assumptions, and architecture-specific routing. - Read
references/distributed-training.mdbefore suggestingtorchrun, DeepSpeed, fairseq, FSDP, multi-node, long-context, flash-attention, or custom-kernel runs. - Read
references/troubleshooting.mdwhen diagnosing missing checkpoints/data/configs, credentialed callers, old vendored fairseq expectations, JSONL mismatches, OOM, or benchmark-scale safety. - Run
scripts/check_training_plan.py --helpand then a mode-specific dry run to validate required fields and produce a safe command template before telling a user to launch anything.
Bundled Helper
scripts/check_training_plan.pyvalidates command plans foryoco-needle,yoco-task,diff-transformer,gad-decoding,pfpo-offline-eval, andresa-math-eval; run it whenever a request asks for a command, launch review, or dry-run alternative.
Safety Defaults
- Treat all native YOCO training/evaluation, PFPO training, GAD/IAD decoding, and ReSA evaluation as expensive unless the user explicitly confirms resources, data, checkpoints, and runtime compatibility.
- Never run PFPO
openai_api_caller_v1.pyorservice_api_caller_v1.pywithout explicit credential/service confirmation; prefer offline JSON/JSONL validation and local result aggregation. - Do not vendor or depend on the original repo checkout. The references in this sub-skill summarize the source evidence and the helper emits templates with placeholder paths only.
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).