rail_representation_loss
Compute RAIL-KD layerwise and concatenated intermediate representation losses from pooled hidden states.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- rail_representation_loss
- description
- Compute RAIL-KD layerwise and concatenated intermediate representation losses from pooled hidden states.
02 资源文件 2
- 脚本scripts/rail_loss.py4.0 KB
- 脚本tests/test_rail_loss.py1.2 KB
03 SKILL.md 正文
RAIL Representation Loss
Use this skill when a distillation experiment must implement the RAIL-KD intermediate representation objective. It is appropriate after a teacher-layer mapping has selected exactly one teacher layer for each student layer. Do not use it for attention distribution matching, MiniLM relation losses, or layer-search logic.
Inputs
- Teacher hidden states: nested numeric lists shaped
[m][tokens][teacher_dim]. - Student hidden states: nested numeric lists shaped
[m][tokens][student_dim]. - Projection matrices for teacher and student vectors.
- Variant:
layerwisefor RAIL-KDl orconcatenatedfor RAIL-KDc. - Optional layer weights for layerwise mode.
Outputs
- Scalar intermediate loss.
- Diagnostics containing pooled vectors, normalized vectors, and per-layer distances when applicable.
Workflow
- Mean-pool each selected hidden-state layer across tokens.
- Project teacher and student vectors to the same dimension.
- L2-normalize projected vectors with a small epsilon guard.
- For layerwise mode, compute weighted squared distances and sum them.
- For concatenated mode, concatenate pooled vectors first, project once, normalize, then compute one squared distance.
- Return diagnostics for recovery mechanism checks.
Validation
Run:
python scripts/rail_loss.py --demo
python -m pytest tests
Limitations
This skill uses small pure-Python numeric lists for portability. Production transformer code should pass actual tensors through equivalent pooling, projection, normalization, and distance operations while preserving the same contracts.
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).