Feature summary
Add a public leaf recipe for GB300 + EKS + Ubuntu + training + Slurm, following the existing GB300 training and Slurm recipe patterns.
Problem and use case
AICR has a GB300 EKS Ubuntu training recipe and GB200 Slurm recipes, but no GB300 EKS Ubuntu training Slurm combination. Operators therefore cannot resolve this supported intent directly and must assemble an untracked configuration.
Proposed solution
Add a recipe that resolves from:
service: eks
accelerator: gb300
os: ubuntu
intent: training
platform: slurm
The recipe should:
- use AICR's current public component pins
- retain the standalone NFD architecture and keep GPU Operator's embedded NFD disabled
- define the Slurm, topology, storage, and accounting choices required for this combination
- keep provider infrastructure configuration outside the recipe
- compose existing bases and mixins where possible rather than duplicating values
Success criteria
Operational considerations
Any required scheduler or topology dependency fix must be consumed as a pinned, verified release. Recipe-time overrides cannot repair a dependency compiled into an upstream binary.
Feature summary
Add a public leaf recipe for GB300 + EKS + Ubuntu + training + Slurm, following the existing GB300 training and Slurm recipe patterns.
Problem and use case
AICR has a GB300 EKS Ubuntu training recipe and GB200 Slurm recipes, but no GB300 EKS Ubuntu training Slurm combination. Operators therefore cannot resolve this supported intent directly and must assemble an untracked configuration.
Proposed solution
Add a recipe that resolves from:
service: eksaccelerator: gb300os: ubuntuintent: trainingplatform: slurmThe recipe should:
Success criteria
Operational considerations
Any required scheduler or topology dependency fix must be consumed as a pinned, verified release. Recipe-time overrides cannot repair a dependency compiled into an upstream binary.