Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/scimate-ai/hpc-skills/hpc-gpu-stacknpx skills add SciMate-AI/HPC-Skills --skill hpc-gpu-stackgit clone --depth 1 https://github.com/SciMate-AI/HPC-SkillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/scimate-ai/hpc-skills/hpc-gpu-stack)<a href="https://agentmods.dev/skills/scimate-ai/hpc-skills/hpc-gpu-stack"><img src="https://agentmods.dev/badge/skills/scimate-ai/hpc-skills/hpc-gpu-stack.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00071 | $0.00835 |
| Opus 5 | $0.00036 | $0.00417 |
| Sonnet 5 | $0.00014 | $0.00167 |
| Haiku 4.5 | $0.00007 | $0.00084 |
Grade A, and why
hpc-gpu-stack scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 5d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 69 lines — stays where its author put it; the contents beside it link to each section on GitHub.
HPC GPU Stack
Treat GPU execution as one coherent stack: CUDA toolchain, host compiler, launcher, scheduler mapping, and device visibility must agree before kernel tuning matters.
Start
- Read
references/cuda-and-host-compiler-matrix.mdbefore choosingnvcc, host compiler, or a CUDA build baseline. - Read
references/gpu-aware-mpi-and-rank-mapping.mdwhen the workflow spans MPI ranks, one-rank-per-GPU layouts, or CUDA-aware MPI. - Read
references/device-visibility-and-scheduler-integration.mdwhen Slurm,CUDA_VISIBLE_DEVICES, MIG, or scheduler-provided GPU allocation is involved. - Read
references/memory-streams-and-overlap-playbook.mdwhen debugging device memory pressure, pinned-memory transfers, streams, or overlap assumptions. - Read
references/build-and-launch-workflow.mdwhen turning a CUDA code path into a reproducible compile-and-run workflow. - Read
references/runtime-debugging-and-profiling.mdwhen kernels fail at runtime, ranks see the wrong device, or performance is unexpectedly poor. - Read
references/error-recovery.mdwhen configure, compile, launch, or runtime CUDA behavior fails. - Read
references/error-pattern-dictionary.mdwhen a GPU failure needs a fast pattern match.
Work sequence
- Confirm the execution model first:
- single GPU
- one MPI rank per GPU
- hybrid MPI plus threads with explicit rank-to-GPU placement
- Keep CUDA toolkit, host compiler, and MPI stack mutually compatible.
- Let the scheduler expose the intended GPU allocation before forcing manual device selection.
- Get a minimal kernel and launch baseline working before tuning streams, overlap, or transport variables.
- Reproduce failures on one node and the smallest GPU count that still shows the issue before scaling out.
Guardrails
- Do not assume
nvccaccepts any host compiler visible inPATH. - Do not mix rank-to-GPU mapping logic from Open MPI, MPICH-family, and Slurm without checking which environment variables are actually set.
- Do not tune streams or overlap to compensate for a broken device-mapping or memory-capacity issue.
- Do not debug multi-node GPU failures before a single-node baseline is trustworthy.
What ships with it
13 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- agents/openai.yaml 253 B
- assets/templates/cuda_mpi_gpu_slurm.sh 463 B runs code
- assets/templates/cuda_single_gpu_slurm.sh 399 B runs code
- assets/templates/cuda_vector_add_minimal.cu 1.1 KB
- assets/templates/nvcc_build_example.sh 297 B runs code
- references/build-and-launch-workflow.md 807 B
- references/cuda-and-host-compiler-matrix.md 1.4 KB
- references/device-visibility-and-scheduler-integration.md 1.2 KB
- references/error-pattern-dictionary.md 1.2 KB
- references/error-recovery.md 756 B
- references/gpu-aware-mpi-and-rank-mapping.md 1.2 KB
- references/memory-streams-and-overlap-playbook.md 1016 B
- references/runtime-debugging-and-profiling.md 804 B
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 5d ago First seen · 69 lines · 71 tokens per session scan A f21a9ef9608b
hpc-gpu-stack is a skill published in the GitHub repository SciMate-AI/HPC-Skills (84 stars, last pushed 5mo ago), licensed MIT. It adds 71 tokens to every session and 835 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
latchbio-integration
Build, register, debug, and operate bioinformatics workflows on Latch using the Python SDK, CLI, Latch Data and Registry, Nextflow, Snakemake, programmatic execution, and Latch MCP. Use when authoring or deploying Latch workflows, configuring resources or interfaces, moving data, integrating Registry, or launching and…
modal-compute
Run explicitly chosen research benchmark or replication jobs on Modal's serverless infrastructure. Use when a Feynman research workflow needs burst remote GPU compute and the Modal CLI is available.
remote-compute-ssh
Evaluate and use SSH Remote Compute before choosing where to run GPU, high-memory, parallel, batch, model-inference, bioinformatics, or other long-running scientific work; supports short remote commands and asynchronous jobs with automatic harvest and analysis.
dnanexus-integration
DNAnexus cloud genomics platform. Build apps/applets, manage data (upload/download), dxpy Python SDK, run workflows, FASTQ/BAM/VCF, for genomics pipeline development and execution.
latchbio-integration
Latch platform for bioinformatics workflows. Build pipelines with Latch SDK, @workflow/@task decorators, deploy serverless workflows, LatchFile/LatchDir, Nextflow/Snakemake integration.
modal-serverless-gpu
Run approved CPU or GPU work through OpenScience computejob on the user's configured Modal account. Use for isolated scientific scripts, dependency provisioning, durable outputs, logs, status, cancellation, and recovery. Never invoke the Modal SDK or CLI directly.