KwongFuk
- 16 skills
- 0 followers
- 8 hours ago last updated
- ▌ Path Storage Policy · kwongfukEnforce a general path planning and storage policy: keep /home for code, configs, manifests, and standard lightweight experiment results, while placing caches, environments, datasets, model weights, checkpoints, intermediate artifacts, and any large files under /scratch. Use when deciding where to create files, store outputs, download assets, configure caches, or lay out experiment directories.
- ▌ Scratch Trash Guard · kwongfukGuard deletion and cleanup requests with a mandatory choice between permanent deletion and moving to /scratch/gguo/trash, plus a required user-entered confirmation code before execution. Use when the user asks to delete, remove, clean up, purge, discard, or archive files/directories, or when a task would normally use rm or otherwise destroy files.
- ▌ Repro Record Workflow · kwongfuk bundleRun research/paper/code reproduction tasks with strict execution logging. Use when requests involve reproducing results, debugging training/evaluation pipelines, Slurm/HPC experiments, dataset/model setup, or iterative small-task execution where every subtask must be written to a repository-level CODEX_RECORD.md.
- ▌ Palmetto Slurm Workflow · kwongfuk bundleRun Clemson Palmetto HPC workflows with Slurm and remote development tooling. Use when requests mention Palmetto, Slurm (`sbatch`/`salloc`/`srun`), login-node vs compute-node usage, conda/module environment setup, file sync between laptop and cluster, PyCharm remote interpreter setup, or Jupyter SSH tunneling.
- ▌ Palmetto Slurm Fit Check · kwongfuk bundleCheck Palmetto Slurm CPU/GPU/memory fit before submitting or editing GPU jobs. Use when a request involves sbatch scripts, GPU training jobs, H100/A100/H200 nodes, or deciding whether a current CPU request is too high for available node fragmentation.
- ▌ Palmetto Slurm Safe Guard · kwongfuk bundleEnforce safe Slurm job templates on Palmetto: keep /home for code and lightweight standard results, keep caches/checkpoints/models/data and other large files on /scratch, plus preflight safety checks, periodic low-noise GPU/process monitoring, resumable/atomic result saving, and signal-safe shutdown handling. Use whenever writing or updating sbatch scripts or long-running translation/training/inference jobs.
- ▌ Slurm Auto Retry Watchdog · kwongfuk bundleSubmit a Slurm sbatch job through a watchdog that monitors sacct/squeue, classifies common failures from job state and stdout/stderr logs, and automatically resubmits retriable failures up to a fixed attempt limit. Use when the user wants automatic retry for Slurm jobs, wants jobs to recover from timeout/OOM/node failures, or wants a reusable trigger/workflow around sbatch submission.
- ▌ Palmetto Gpu Usage Advisor · kwongfuk bundleInspect Palmetto GPU usage and recommend which card type and GPU count to request right now. Use when the user asks about current GPU occupancy, cluster card availability, or which GPU type/count is the best choice before submitting a Palmetto job.
- ▌ Exact Training Resume Guard · kwongfukEnforce exact-resume support for long-running training jobs. Use when writing or updating any training script, launcher, sbatch file, DeepSpeed/Accelerate/TRL training entrypoint, or checkpoint policy where future runs must resume from the last step with optimizer, scheduler, RNG, and framework state preserved after timeout, preemption, or manual interruption.
- ▌ Palmetto Conda Scratch Guard · kwongfuk bundleEnforce Palmetto storage policy: keep /home for code, configs, and standard small experiment results, while placing Conda envs, caches, datasets, models, containers, checkpoints, and other large files under /scratch; run preflight checks for login-vs-compute node, GPU visibility, and CUDA tools/runtime before installs or training. Use when creating/updating envs, pulling HF models/datasets, preparing Apptainer images, debugging missing GPU/CUDA, or validating cluster node context.
- ▌ Palmetto Cpu Request Advisor · kwongfuk bundleEstimate and right-size Palmetto CPU requests before submitting CPU or GPU jobs. Use when choosing Slurm cpus-per-task/ntasks, interpreting low CPU utilization, or avoiding over-allocation for serial jobs, dataloaders, and GPU training runs.
- ▌ Conversation History Recorder · kwongfukRecord structured conversation history for projects under active workspaces. Use when a task touches a project, when user instructions must be remembered exactly, or when Codex should preserve project problems, feedback, and user choices across turns without logging tools or code-level implementation details.
- ▌ Codex Custom Provider Pack · kwongfuk bundleConfigure Codex against the `https://api.mol.us.ci/v1` custom provider, keep a table of hosted and local models, and use helper scripts to scan `/scratch/gguo/.hf_cache/hub`, smoke-test API access, or generate JSONL samples for evaluation.
- ▌ Palmetto Apptainer Libcuda Fix · kwongfukFix Palmetto Slurm jobs where Apptainer or vLLM workers fail with `libcuda.so cannot found` even though `/lib64/libcuda.so.1` exists on the host. Use when logs show CUDA driver discovery failures inside containerized GPU jobs and you need a safe repair plus smoke validation before resubmitting the full run.
- ▌ Slurm Self Healing Watchdog · kwongfuk bundleSubmit a Slurm sbatch job through a watchdog that waits for terminal state, diagnoses failures from sacct/squeue plus stdout/stderr, invokes Codex to apply a minimal repair when needed, and resubmits up to 5 times. Use when the user wants a trigger-like loop that can monitor stopped jobs, analyze why they failed, patch the repo or sbatch script, and retry automatically.
- ▌ Conversation History Recorder 2 · kwongfukRecord structured conversation history for any project under /home/gguo/code. Use when a task touches a workspace project, when user instructions must be remembered exactly, or when Codex should preserve project problems, feedback, and user choices across turns without logging tools or code-level implementation details.