Fix srun failure in the killarney imagenet job script

[lebrice]

Problem

On Killarney, the imagenet example job dies at its first srun:

srun: fatal: SLURM_MEM_PER_CPU, SLURM_MEM_PER_GPU, and SLURM_MEM_PER_NODE are mutually exclusive.

Killarney's lua job submit plugin adds a default --mem (8G) even when --mem-per-gpu is passed. The job then has both SLURM_MEM_PER_NODE=8192 and SLURM_MEM_PER_GPU=65536 set, and every srun inside it inherits both.

Fix

Request --mem=64G (per node) instead of --mem-per-gpu=64G in scripts/job_killarney.sh. For the default 1-GPU job the amount of memory is the same.

Verification

Test jobs on Killarney with the same resource flags:

• --mem-per-gpu=64G (job 5684573): the job gets SLURM_MEM_PER_GPU=65536 and SLURM_MEM_PER_NODE=8192, and srun true fails with the error above.
• --mem=64G (job 5684592): the job gets only SLURM_MEM_PER_NODE=65536, and srun true succeeds.

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.10%. Comparing base (5c25d22) to head (b64c21e).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #222   +/-   ##
=======================================
  Coverage   72.10%   72.10%           
=======================================
  Files          22       22           
  Lines        2370     2370           
=======================================
  Hits         1709     1709           
  Misses        661      661           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.