菜单
Plan and review GPU/scientific-compute optimization across PyTorch, CUDA-aware batching, memory pressure, mixed precision, profiling, and reproducibility. Use when workflows are slow, GPU-bound, memory-bound, or need scalable training/inference plans.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | gpu-compute-optimization |
| description | Plan and review GPU/scientific-compute optimization across PyTorch, CUDA-aware batching, memory pressure, mixed precision, profiling, and reproducibility. Use when workflows are slow, GPU-bound, memory-bound, or need scalable training/inference plans. |
| user-invocable | true |
| license | Apache-2.0 |
Source adaptation: BioRouter-original synthesis inspired by remaining Scientific Agent Skills and Auto-Empirical Research Skills capability clusters, rewritten under Apache-2.0. This is not copied upstream text; it is a native BioRouter routing and verification checklist.
Use this skill when the user asks about GPU-bound training, inference, profiling, mixed precision, batching, or memory optimization. Prefer the most specific BioRouter extension, BioSkill, or language style skill whenever a concrete platform, package, or file type appears.
Before producing output, capture:
Use these routes when applicable:
bio-skills category.causal-identification-gates, statistical-modeling-bayes, or econometrics-toolkit.citation-temporal-integrity and claim-evidence-integrity.replication-package-audit before treating the result as final.Do not present final claims until these are true:
Return:
Plan and review computational fluid dynamics workflows with meshes, boundary conditions, turbulence models, OpenFOAM/FluidSim-style solvers, validation cases, convergence checks, and reproducible simulation artifacts. Use when users mention CFD, fluid dynamics, Navier-Stokes, meshing, turbulence, OpenFOAM, or flow simulation.
Run evaluator-driven hypothesis trees for improving research artifacts, prompts, analyses, and designs. Use when a task needs branch-and-prune exploration, rubric scoring, iterative refinement, or explicit comparison of competing scientific/workflow alternatives.
Detect repeated user workflows and draft native BioRouter skills, hooks, or checklist updates from observed work patterns. Use when users ask to turn repeated work, screen-history patterns, project routines, or recurring guardrails into reusable BioRouter skills.
Plan and audit BIDS neuroimaging datasets, derivatives, validation, metadata, and reproducible preprocessing. Use when users mention BIDS, fMRI, MRI, EEG/MEG/iEEG, DICOM conversion, dataset_description.json, participants.tsv, or neuroimaging derivatives.
Plan and review clinical AI modeling workflows with EHR cohorts, PyHealth-style tasks, validation, bias checks, calibration, safety boundaries, and no-PHI handling. Use when users ask about clinical prediction models, risk scores, treatment-response ML, medical reports, or clinical model evaluation.
Plan and review general time-series forecasting with classical models, ML, foundation models, backtesting, leakage checks, and uncertainty. Use for non-omics forecasting, sensor streams, operations, finance, clinical vitals, demand, or irregular temporal data.