Shared LLMFlux directory on Delta

[lmarini]

So that others can help mantain it.

We can use the shared software directory for LLMHub to share the files.

Service user should be able to deploy this for other users.

• Test deploy scripts with this directory
• How do we install modules?

[dadamsncsa]

Install location for the shared LLMFlux deployment: /sw/llmhub/llmflux/ (under the svcdeltallmhub-owned LLMHub software root).

Layout:

• conda-env/ — lightweight host env for the llmflux CLI (pure Python; the PyTorch/vLLM/Ollama stack runs inside the container)
• containers/<version>/llm_processor.sif
• modules/llmflux/<version>.lua + .version

The deployment is owned by svcdeltallmhub and published by impersonating it (the delta_sudo_svcdeltallmhub sudo group), so any team member can maintain it. The publish scripts live in a shared, group-maintainable directory (/projects/bfmz/svcdeltallmhub/llmflux-deploy/); see #101.

Test deploy scripts with this directory — the publish tooling is in place and validated in dry-run; the live end-to-end deploy is the remaining step.

How do we install modules? — the module tree is /sw/llmhub/llmflux/modules. We plan to request it be added to the global MODULEPATH so module load llmflux/<version> works with no setup; in the interim, module use /sw/llmhub/llmflux/modules.

[dadamsncsa]

Refinement: each version now gets a fresh per-version conda env (/sw/llmhub/llmflux/envs/<version>) plus a new module file that becomes the default; older versions keep working against their own env. Each env also carries the standard Jupyter + data-science packages and a setup-kernel command (registers the env as a Jupyter kernel), so it doubles as a general-purpose env.