torchdrug
Route TorchDrug graph learning, drug discovery, molecular ML, protein, knowledge graph, data, model, and training engine workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- torchdrug
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route TorchDrug graph learning, drug discovery, molecular ML, protein, knowledge graph, data, model, and training engine workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/repo-provenance.md1.6 KB
- 参考文档references/repo-routing-metadata.json496 B
- 参考文档references/troubleshooting.md3.7 KB
- 脚本scripts/check_torchdrug_env.py1.7 KB
03 SKILL.md 正文
TorchDrug
Use this repo skill when a task involves TorchDrug, TorchProtein, graph neural networks for drug discovery, molecular property prediction, molecular generation, retrosynthesis, protein workflows, knowledge graph completion, or TorchDrug's core.Engine training stack.
Start Here
- Read references/repo-provenance.md before deciding whether this skill is current for a TorchDrug checkout or package version.
- Read references/troubleshooting.md for install/import, PyTorch/PyG compiled packages, RDKit, data downloads, GPU,
mps, and package-version pitfalls. - Run
python scripts/check_torchdrug_env.pyin the user's TorchDrug environment for a no-download import and capability check.
Install And Verify
TorchDrug 0.2.1 supports Python >=3.7,<3.11 and PyTorch >=1.8.0. Prefer a fresh environment with Python 3.10 or earlier:
python -m pip install torchdrug
python - <<'PY'
import torchdrug
from torchdrug import data
print(torchdrug.__version__)
graph = data.Graph([[0, 1], [1, 0]], num_node=2)
print(graph.num_node, graph.num_edge)
PY
When installing from pip, install PyTorch first, then choose torch-scatter and torch-cluster wheels that match the PyTorch and CUDA/CPU build. Conda users can prefer the documented conda channels because they solve PyTorch, PyG, RDKit, and compiled packages together.
Route By Task
| User task | Read |
|---|---|
| Graph, molecule, protein, dataset, SMILES, sequence, packing, masking, splitting, collator, or data validation tasks | sub-skills/graph-data/SKILL.md |
Choosing or customizing graph layers, representation models, readouts, samplers, variadic tensor utilities, or MessagePassingBase subclasses |
sub-skills/layers-and-extensions/SKILL.md |
Building core.Engine loops, checkpointing, config serialization, logging, CPU/GPU settings, task contracts, or training harnesses |
sub-skills/training-engine/SKILL.md |
| Molecular property prediction, molecular pretraining, generation with GCPN/GraphAF, or USPTO50k retrosynthesis | sub-skills/molecular-workflows/SKILL.md |
| Knowledge graph triples, FB15k/WN18/YAGO/Hetionet, TransE/RotatE/NeuralLP/KBGAT, negative sampling, or filtered ranking | sub-skills/knowledge-graphs/SKILL.md |
| Protein sequences/structures, contact prediction, GearNet, ESM, protein property/function, or protein-protein interaction workflows | sub-skills/protein-workflows/SKILL.md |
Common Workflows
- For a tiny in-memory smoke test, route to
graph-dataand run itsscripts/smoke_graph_data.pyhelper. - For a new molecular prediction project, combine
graph-datafor the molecule dataset,molecular-workflowsfor task/model choices, andtraining-engineforcore.Enginesave/load. - For custom graph neural layers, use
layers-and-extensionsfirst, then return totraining-engineto wire the module into a task. - For protein GearNet or ESM tasks, use
protein-workflowsfor data/model/task choices andlayers-and-extensionsonly for graph construction internals. - For knowledge graph reasoning, use
knowledge-graphsfor model/task/dataset decisions andtraining-enginefor optimizer, evaluation, and checkpoint details.
Safety Boundaries
- Do not run dataset downloads, long training, retrosynthesis generation, benchmark sweeps, or ESM weight downloads unless the user allows network, storage, and compute cost.
- Do not assume
gpus=[0]is valid. Usegpus=Nonefor CPU-only work and checktorch.cuda.is_available()before GPU examples. - Do not use Apple
mps; TorchDrug documentation says Apple silicon can run on CPU butmpsis unsupported. - Do not tell users to open original TorchDrug docs, tests, or examples as runtime requirements. This skill bundles the needed routes, references, and safe helpers.
Repository Scope
This skill covers public TorchDrug package usage and repo-derived workflows: source APIs under torchdrug/, install metadata, Sphinx docs, tutorials, benchmarks, and representative tests. It intentionally does not cover maintainer release automation, full benchmark reproduction, or external dataset/model-weight acquisition beyond documented prerequisites.
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).