gsdm_recovery_harness
Run a bounded BCMF proxy recovery that exercises generated Graphically Structured Diffusion Model skills.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- gsdm_recovery_harness
- description
- Run a bounded BCMF proxy recovery that exercises generated Graphically Structured Diffusion Model skills.
02 资源文件 2
- 脚本scripts/run_gsdm_recovery.py11.3 KB
- 脚本tests/test_run_gsdm_recovery.py1.3 KB
03 SKILL.md 正文
GSDM Recovery Harness
Use this skill during soft-mode recovery for Graphically Structured Diffusion Models when full paper-scale training is blocked by wall-clock or GPU cost. It assembles the generated graph, permutation, and objective skills into one executable BCMF proxy experiment.
Do not use this skill to claim full reproduction of the paper's RMSE curves. It produces mechanism-faithful reduced evidence only.
Inputs
attempt_dir: current Distiller attempt directory.skill_root: directory containing the generated GSDM skills.environment/runtime_handoff.json: runtime preflight fromprepare-recovery-environment.- BCMF dimensions and optimizer settings.
Outputs
recovery/logs/generated_data_item.jsonrecovery/logs/training_trace.jsonrecovery/logs/generated_skill_invocations.jsonrecovery/source_manifest.jsonrecovery/recovery_result.json
Workflow
- Build a deterministic tiny BCMF item.
- Invoke the graph skill to produce the structured attention mask.
- Invoke the permutation skill to verify a full plate-index swap preserves that mask.
- Invoke the objective skill to encode mixed continuous/binary variables and compute the masked denoising loss.
- Run a scalar denoiser optimizer step against the masked
x0objective. - Write source-boundary, command, and mechanism-check evidence.
CLI
python scripts/run_gsdm_recovery.py --attempt-dir /path/to/attempt --skill-root /path/to/generated/skills --learning-rate 0.05
Validation
python -m pytest tests
The tests run the scalar step in memory and assert that loss decreases and parameters change. The learning rate is exposed so refinement cycles can stress optimizer stability without editing the script.
Limitations
The proxy uses a small scalar denoiser rather than the full transformer. It is acceptable only under soft recovery mode after the runtime handoff records full-scale blockers.
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).