reduced_recovery_protocol
Build and validate soft-mode reduced recovery artifacts for residual EBM text generation without reading the original repository.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- reduced_recovery_protocol
- description
- Build and validate soft-mode reduced recovery artifacts for residual EBM text generation without reading the original repository.
02 资源文件 2
03 SKILL.md 正文
Reduced Recovery Protocol
Use this skill when full residual EBM reproduction is blocked and the run is explicitly in soft mode. The skill defines the artifact contract for a reduced experiment that still exercises fixed-LM proposals, residual energy scoring, conditional NCE optimization, and importance reweighting.
Do not use this skill in hard mode as acceptance evidence. Do not read the original source repository during recovery.
Inputs
- Attempt directory.
- Runtime handoff with full-runtime blockers and allowed sources.
- Module plan and generated skill paths.
- A reduced data item containing prefix, positive continuation, and fixed-LM-style negatives.
Outputs
- Experiment plan explaining full and reduced targets.
- Executable recovery command.
- Source manifest with forbidden-source checks.
- Recovery result with numeric proxy metric and mechanism checks.
- Training trace with
params_before,params_after,loss_before, andloss_after. - Generated-skill invocation evidence.
Workflow
- Confirm recovery mode is soft and the full target is blocked by runtime or data constraints.
- Write the experiment plan before running the experiment.
- Build a data item from allowed paper/runtime sources, not from the original repository.
- Invoke the generated residual scoring, NCE training, and importance sampling skills.
- Save command logs and source manifest.
- Run the Distiller recovery experiment validator.
- Treat the proxy as acceptable only if mechanism checks and validation pass.
Validation
Run:
python scripts/check_recovery_artifacts.py --attempt-dir <attempt_dir>
python /share/project/yuyang/workspace/Paper2Skills/Distiller/skills/recover-paper/scripts/validate_recovery_experiment.py <attempt_dir>
Limitations
This protocol proves mechanism fidelity under reduced conditions. It does not claim the CC-News or Toronto Book Corpus perplexity result unless the full model, dataset, and sample-count protocol are actually run.
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).