random_layer_mapping
Sample and audit RAIL-KD random teacher-to-student intermediate layer mappings for bounded knowledge distillation experiments.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- random_layer_mapping
- description
- Sample and audit RAIL-KD random teacher-to-student intermediate layer mappings for bounded knowledge distillation experiments.
02 资源文件 2
- 脚本scripts/rail_mapping.py2.5 KB
- 脚本tests/test_rail_mapping.py1015 B
03 SKILL.md 正文
Random Layer Mapping
Use this skill when implementing or checking the RAIL-KD layer-selection mechanism. It is appropriate for transformer distillation or proxy experiments that need to select m teacher intermediate layers from an n-layer teacher at each epoch and align them with m student layers. Do not use it for deterministic PKD/CKD mapping search or attention-over-all-layers methods.
Inputs
teacher_layer_count: positive integern.student_layer_count: positive integerm, withm <= n.seedand optionalepoch: reproducibility controls.sort_indices: keep true for the paper-faithful monotonic alignment.
Outputs
- A list of distinct zero-based teacher layer indices.
- Student-to-teacher mapping pairs.
- Optional coverage counts over repeated epochs.
Workflow
- Validate layer counts before sampling.
- Use seeded random sampling without replacement to choose exactly
mteacher layers. - Sort the selected teacher layers before pairing with student layers unless a diagnostic explicitly needs unsorted samples.
- Pair student layers
0..m-1with the selected teacher indices. - For multi-epoch experiments, record mappings and coverage so analysis can verify that randomization was exercised.
Validation
Run:
python scripts/rail_mapping.py --teacher-layers 12 --student-layers 6 --epochs 4 --seed 13
python -m pytest tests
The tests are standard-library compatible through the Distiller simple test runner.
Limitations
The script only handles mapping selection and audit statistics. It does not compute hidden-state losses, logits KD, or optimizer updates. Indexing is zero-based; convert explicitly if a downstream framework reports one-based layer numbers.
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).