physicsnemo
Route PhysicsNeMo users to the right model, data, scaling, mesh, diffusion, active-learning, or deployment workflow.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- physicsnemo
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route PhysicsNeMo users to the right model, data, scaling, mesh, diffusion, active-learning, or deployment workflow.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/install-and-environment.md1.7 KB
- 参考文档references/optional-dependencies.md1.9 KB
- 参考文档references/package-api-map.md1.8 KB
- 参考文档references/repo-provenance.md1.0 KB
- 参考文档references/repo-routing-metadata.json525 B
- 参考文档references/troubleshooting.md2.0 KB
- 脚本scripts/physicsnemo_environment_smoke.py2.2 KB
03 SKILL.md 正文
PhysicsNeMo
Use this repo skill for NVIDIA PhysicsNeMo: a scientific-ML framework for physics-aware models, data pipelines, mesh processing, distributed/domain-parallel execution, diffusion workflows, active learning, and ONNX export.
This skill is a router, not a full manual. Start here to choose the right sub-skill, then follow the bundled reference files and scripts for the detailed workflow.
Quick start
- Public install:
pip install nvidia-physicsnemo - If you need a specific backend or optional workflow, read optional dependencies first.
- Tiny smoke: run
python scripts/physicsnemo_environment_smoke.pyafter install. - For contributor/editable use in a local checkout, keep the public package name
nvidia-physicsnemoand do not depend on this repository path at runtime.
What to route where
- Choose a model family or example starting point → model-selection
- Build readers, datasets, transforms, or
DataLoaderpipelines → datapipes - Use
DistributedManager, DDP/FSDP2,ShardTensor, or mesh-parallel launchers → distributed-and-domain-parallel - Create, validate, repair, sample, serialize, or analyze
Mesh/DomainMeshobjects → mesh-and-geometry - Work with diffusion samplers, preconditioners, guidance, multi-diffusion, or generative recipes → diffusion-and-generative
- Operate active-learning loops or export to ONNX → active-learning-and-deployment
Root import and package shape
physicsnemo.modelsroot exports onlyDiT,DoMINO, andFullyConnected.- Most model families are imported from their family subpackages, such as
physicsnemo.models.fno,physicsnemo.models.meshgraphnet, orphysicsnemo.models.graphcast. physicsnemo.meshroot exports the core object model; validation, I/O, sampling, repair, and geometry helpers live in submodules.physicsnemo.active_learningroot exports the driver/config/registry surface; protocol interfaces live in itsprotocolssubmodule.
If a user asks for an exact class name, backend option, or import path, use package API map and the relevant sub-skill reference rather than guessing from memory.
Read these first when needed
- Install and environment — install commands, backend notes, and the package smoke check.
- Optional dependencies — extras such as
cu12,cu13,gnns,mesh-extras,datapipes-extras,sym,natten, and Transformer Engine. - Troubleshooting — import mistakes, missing extras, CUDA/backend issues, data/weight access, and long-example overrun prevention.
- Repository provenance — source commit, package version, dirty-state summary, and evidence paths.
Selection hints
When the task is ambiguous, resolve these axes before diving deeper:
- Data shape: regular grid, weather lat-lon or HEALPix, unstructured mesh/graph, point cloud, particle graph, or coordinate/tabular inputs.
- Task type: surrogate/regression, autoregressive forecasting, physics-informed training, diffusion/generative, active learning, deployment, or scaling.
- Constraints: CUDA availability, optional extras, external datasets, pretrained weights, latency, and whether the user wants a tiny smoke or a full example.
If the route still depends on the exact model family, open model-selection first.
Minimal verification workflow
- Install the package and any needed extras.
- Run
python scripts/physicsnemo_environment_smoke.py. - Route to the appropriate sub-skill.
- Use that sub-skill’s bundled reference files for the deep workflow.
Notes
- Do not send users to original repo paths at runtime; only use bundled references and scripts.
- Do not treat large example training jobs as smoke tests.
- Keep experimental APIs labeled experimental when a workflow explicitly depends on them.
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).