virtual_update_interference
Rank replay candidates by loss increase under a virtual incoming-batch update for MIR selection.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- virtual_update_interference
- description
- Rank replay candidates by loss increase under a virtual incoming-batch update for MIR selection.
02 资源文件 2
- 脚本scripts/mir_scoring.py3.5 KB
- 脚本tests/test_mir_scoring.py1.3 KB
03 SKILL.md 正文
Virtual Update Interference
Use this skill when a recovery harness or implementation needs the core MIR selector. It computes a one-step virtual update from the incoming batch, evaluates candidate memory examples before and after that virtual update, and selects the examples with the greatest predicted loss increase.
Inputs
- Current linear model parameters or an adapter exposing equivalent loss and gradient behavior.
- Incoming batch examples with features and binary labels.
- Candidate memory examples with features, labels, and optional best historical losses.
- Learning rate, replay budget, and scoring variant
smi_1orsmi_2.
Outputs
- Virtual parameters.
- Candidate score records with current loss, virtual loss, and MIR score.
- Top-k selected replay candidates ordered by descending interference.
Workflow
- Clone the current parameters so the real model is not mutated.
- Compute the average logistic-loss gradient on the incoming batch.
- Apply one virtual SGD step to produce
theta_v. - Score each candidate with
loss(theta_v) - loss(theta)forsmi_1. - For
smi_2, subtractmin(current_loss, best_loss). - Sort by descending score, tie-breaking by stable example id.
Validation
Run the included tests or validate_skill_tree.py --run-tests. They verify non-mutation, ranking of a harmed candidate, empty candidate behavior, and smi_2 historical-loss handling.
Limitations
The bundled script is a deterministic standard-library logistic-regression implementation for small recovery experiments. Larger neural systems should preserve the same contract with framework-specific models.
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).