p3o_surrogate_loss
Build decomposed P3O surrogate loss terms with clipped replay gradients and behavior-to-target KL regularization.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- p3o_surrogate_loss
- description
- Build decomposed P3O surrogate loss terms with clipped replay gradients and behavior-to-target KL regularization.
02 资源文件 2
- 脚本scripts/p3o_loss.py1.6 KB
- 脚本tests/test_p3o_loss.py357 B
03 SKILL.md 正文
P3O Surrogate Loss
Use this skill when a recovery or implementation must preserve the P3O objective from Equation 14. It is not a full RL framework; it computes deterministic scalar components that can be embedded in a learner.
Inputs
- On-policy score-gradient proxies and advantages.
- Off-policy score-gradient proxies, advantages, and importance ratios.
- Target and behavior action distributions for KL estimation.
clip_thresholdandkl_coefficient, usually from the ESS scheduler.
Outputs
on_policy,off_policy,kl_penalty, andobjectivescalar values.- Clipped ratios used for replay data.
Workflow
- Multiply on-policy advantages by score-gradient proxies.
- Clip each replay importance ratio at
cand multiply by replay advantage and score proxy. - Compute
KL(beta || pi_theta)over behavior and target distributions. - Combine terms as
on + off - lambda * KL.
Source Boundary
Use this skill with the paper, module documents, generated artifacts, and ordinary package documentation. Do not read or depend on the original P3O repository.
Validation
Run python scripts/<script>.py --self-test or python -m pytest tests from the skill directory. The bundled tests use only the Python standard library.
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).