Claude Skills para Treinamento LoRA e SLURM
weipengxumpi/claude-skills
Automação de treinamento LoRA e workflows SLURM em HPC
Ver no GitHubConjunto de skills customizados para Claude Code que automatizam workflows de treinamento Wan2.2-S2V LoRA e gerenciamento de jobs SLURM em clusters HPC. Inclui recursos para retomada de checkpoints, avaliação de modelos, análise de filas e transferência de dados entre nós.
Instalação
npx convez-ai weipengxumpi/claude-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Retomar treinamentos LoRA de checkpoints intermediários com configuração automática de scripts SLURM
- Gerar comandos de inferência e avaliação com parâmetros corretos de modelo e ambiente
- Analisar tempos de espera e execução de jobs SLURM com relatórios detalhados
- Sincronizar dados entre instâncias cloud e clusters LEONARDO via SSH com rsync paralelo
Veredito

Skills muito especializadas para ecossistema HPC/SLURM. Excelente automação de tarefas repetitivas em treinamento de modelos, mas com curva de aprendizado alta e aplicabilidade limitada a quem trabalha com clusters de computação científica.
README
A tradução para o português deste README ainda está na fila.
Claude Code Custom Skills
Custom Claude Code skills for working with Wan2.2-S2V LoRA training and SLURM on an HPC cluster.
Skills
| Skill | What it does |
|---|---|
continue-training | Resume a Wan2.2-S2V LoRA run from its latest (or chosen) checkpoint. Generates a new ..._leo_<N+1>.sh SLURM script that resumes via --lora_checkpoint with a configurable --skip_frames, following the project's _cont<STEP>_SF<SF> convention. |
eval-training | Generate the livedealer_infer.py evaluation command for a training run (script or SLURM job id). Emits the inference snippet on the eyes-only test set with the correct WAN_* env flags, lora_path/step, width/height, and pose/object inputs. |
slurm-wait-analysis | Query sacct for your jobs, compute queue wait time and run time, and write a readable Markdown/HTML report with a per-job table, summary cards, and averages. Supports filtering by node count and minimum run time. |
transfer-lambda-to-leonardo | Copy/sync data from a Lambda cloud instance (or any non-LEONARDO Linux box) to LEONARDO $WORK. Solves the no-smallstep-cert auth problem via SSH agent forwarding, creates the destination dir (remote rsync is 3.1.3, no --mkpath), and runs a resumable rsync to the datamover with a dmover1-4 parallel-split option for many small files. |
Installation
Each skill is a directory containing a SKILL.md (with name/description
frontmatter) plus its supporting scripts. To use them in a project, copy or
symlink the skill directories into that project's .claude/skills/:
git clone https://github.com/<your-username>/claude-skills.git
ln -s "$(pwd)/claude-skills/continue-training" /path/to/project/.claude/skills/continue-training
ln -s "$(pwd)/claude-skills/eval-training" /path/to/project/.claude/skills/eval-training
ln -s "$(pwd)/claude-skills/slurm-wait-analysis" /path/to/project/.claude/skills/slurm-wait-analysis
Or install them user-wide under ~/.claude/skills/.
Claude Code discovers each skill from its SKILL.md and invokes it when your
request matches the skill's description.