Slurm's jobs in llmgw's accounting: projects as accounts, jobs priced
CI / test (push) Failing after 7m24s

With slurm.users llmgw, each project is a Slurm account, its organization's
and team's above it, with the project's members and its team's and
organization's on it; slurm.conf holds a job to its user's accounts. The
controller's agent keeps them with sacctmgr, only what it made, and reports
the jobs that ended in the last day; the leader records each once, its
GPU-hours (core-hours without cards) priced at auth.gpu_hour_cost into its
project's usage as model slurm/PARTITION.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-11-15 11:54:01 +01:00
1 parent 8547528fb7
commit fcda4d3f85
22 files changed
+1390 -76

No files matched your search

+11
View File
@@ -217,6 +217,17 @@ agents install); root and the machines' own users log in as before. A site whose
directory (sssd against Active Directory or LDAP) passes `--slurm-users
site`, or later `llmgw ctl slurm users off`.
With llmgw's users, Slurm's accounts are llmgw's projects: an account for
each organization, team and project, nested (`acme`, `acme-ml`,
`acme-ml-training`), and on each project's account its members and its
team's and organization's. A job runs on one of its user's accounts and no
other (`sbatch --account=acme-ml-training`; without it, the user's first),
and when it ends its cards' hours (a job without cards, its cores') are
priced at `auth.gpu_hour_cost` (`core_hour_cost`) into the project's usage,
where it shows as model `slurm/PARTITION` beside the project's requests.
Accounts a Slurm administrator made are left alone. A job's cost does not
yet count against a budget, which binds API keys only.
Jobs may run in containers: `llmgw ctl slurm containers on` has the agents
install enroot and pyxis on every Slurm machine (NVIDIA's releases, at
versions llmgw pins and checks; pyxis is built against each machine's own