multitask-modeling
Use DeepCTR-Torch SharedBottom, ESMM, MMOE, and PLE multi-task models with aligned labels, losses, metrics, and predictions.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- multitask-modeling
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use DeepCTR-Torch SharedBottom, ESMM, MMOE, and PLE multi-task models with aligned labels, losses, metrics, and predictions.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md7.0 KB
- 参考文档references/mtl-models-and-training.md6.6 KB
- 参考文档references/troubleshooting.md6.0 KB
- 脚本scripts/mmoe_multitask_smoke.py5.6 KB
03 SKILL.md 正文
multitask-modeling
Use this sub-skill when the task needs one DeepCTR-Torch model to optimize more than one target with the implemented multi-task classes: SharedBottom, ESMM, MMOE, or PLE.
Route map
- Build
SparseFeat,DenseFeat,VarLenSparseFeat, feature names, andmodel_inputdictionaries with../feature-column-inputs/SKILL.mdfirst. - Use
references/mtl-models-and-training.mdfor model selection, label matrix construction, compile/fit/predict flow, ByteRec-style two-target recipes, and metric handling. - Use
references/api-reference.mdfor exact constructor signatures, supported task types, loss strings, output shape, and parameter notes. - Use
references/troubleshooting.mdbefore changing model code when target shape, task order, loss list, ESMM semantics, flattened metrics, or tiny batch behavior looks wrong. - Run
scripts/mmoe_multitask_smoke.py --helpor the default smoke command to verify that a local installation can train an MMOE on inline two-label data and return predictions shaped(n_samples, 2). - Route single-output CTR/ranking models, single-target target arrays, persistence, and callbacks to
../single-task-modeling/SKILL.mdand shared training references.
Operating rules
- Multi-task labels are a two-dimensional NumPy-like matrix shaped
(n_samples, num_tasks). Columnimust be the label fortask_names[i]. model.predict(model_input, batch_size)returns a NumPy array shaped(n_samples, num_tasks). Prediction columnibelongs totask_names[i]; evaluate each task column explicitly.task_typeslength must equallen(task_names). Supported values are"binary"and"regression"forSharedBottom,MMOE, andPLE;ESMMaccepts two binary tasks only.- Compile multi-task models with a loss list of the same length and order as
task_names, for example['binary_crossentropy', 'binary_crossentropy']or['binary_crossentropy', 'mse']. - Do not claim support for arbitrary custom task graphs: this package exposes the four fixed multi-task architectures above.
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).