sample_factory_appo_update
Compute reduced APPO update evidence with PPO clipping and V-trace-style value correction for asynchronous policy lag.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- sample_factory_appo_update
- description
- Compute reduced APPO update evidence with PPO clipping and V-trace-style value correction for asynchronous policy lag.
02 资源文件 2
- 脚本scripts/appo_update.py4.0 KB
- 脚本tests/test_appo_update.py1.3 KB
03 SKILL.md 正文
Sample Factory APPO Update
Use this skill when a recovery needs deterministic evidence that asynchronous trajectory data is corrected with the paper's APPO mechanisms: PPO clipping plus V-trace-style truncated-importance value targets.
Do not use the reduced scalar optimizer as proof of full neural-network training.
Inputs
- Rewards, discounts, values, and bootstrap value.
- Behavior and target log probabilities.
- PPO clip threshold.
- V-trace rho and c truncation thresholds.
- Optional scalar policy parameter and learning rate for a tiny optimizer step.
Outputs
- Importance ratios and clipped ratios.
- Corrected value targets.
- Policy loss and value loss.
- Optimizer trace with parameter and loss before/after.
Workflow
- Compute importance ratios from target minus behavior log probabilities.
- Clip ratios for PPO surrogate loss.
- Compute temporal-difference deltas and apply truncated rho/c weights.
- Build value targets by backward recursion from the bootstrap value.
- Compute policy and value losses.
- For reduced recovery, update a scalar parameter with a deterministic gradient step and record
params_beforeandparams_after.
Validation
Run:
python scripts/appo_update.py
python tests/test_appo_update.py
Limitations
This skill validates the algorithmic contract, not production-scale GPU training or recurrent policies.
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).