SPORC as the interim target¶
Research Computing retired rc-onboard for research jobs on 2026-09-16, and fl-mlm has no
TIGRIS association: in ColdFront both fl-mlm and prdiscourse carry Needs Review, and
the RC documentation grants every project without a pending review a TIGRIS Slurm account.
sacctmgr show assoc lists fl-mlm on sporc only, and sbatch --test-only against the
tigris partition fails with Invalid account or account/partition combination.
fl-mlm is valid on SPORC today. Measured there, not taken from node tags:
| architecture | x86_64 Intel, RHEL 9.8 |
| GPU | A100 40 GB (job 21705076, nvidia-smi); one node of 4x H100 80 GB |
| driver | 610.43.02, CUDA 13.3, so the cu130 torch wheel applies; the cuda11 feature tag is stale |
| network | compute nodes reach PyPI, the PyTorch index, GitHub and Hugging Face (job 21705465) |
| submission | from the TIGRIS login with --clusters=sporc; $HOME is shared |
Not yet measured: whether the 7B training fits in 40 GB. No TIGRIS run logged peak GPU
memory, and the GH200's unified memory would not have surfaced an overflow. The first SPORC
run records it, along with seconds per step, since every --time in scripts/ was sized on a
GH200.
Measured 2026-10-03: see "The registered 7B on a 40 GB A100" below.
Constraint this places on results. A result set that mixes GH200 and A100 runs confounds
the contrast with hardware. Each job's output now records its cluster, job, GPU and peak memory,
so a mixed set is detectable from the files rather than from memory. Results written before
2026-09-17 carry no such record; all of them ran on TIGRIS GH200s under rc-onboard.