SynthStrip node missing mem_gb causes OOM with concurrent anatomical runs

[m9h]

Summary

The SynthStrip nipype node in mriqc/workflows/shared.py does not declare mem_gb, so it defaults to 0.25 GB. MRIQC's memory-aware scheduler then launches multiple SynthStrip instances concurrently, each loading the full PyTorch model (~4-6 GB). This causes OOM kills on systems with ample memory (e.g. 24 GB allocated for a single participant).

Reproducer

Run MRIQC on a multi-session BIDS subject with several T1w/T2w volumes (e.g., 4x T1w + 2x T2w):

mriqc /data /out participant \
    --participant-label SUB01 \
    --nprocs 8 \
    --mem-gb 22 \
    --no-sub

The scheduler dispatches 3-4 concurrent synthstrip nodes. Each subprocess loads synthstrip.1.pt via torch.load(), consuming ~4-6 GB. Total memory spikes to 12-20 GB for brain extraction alone, triggering OOM (exit code 137):

[Node] Error on "mriqc_wf.anatMRIQC.synthstrip_wf.synthstrip"
RuntimeError: subprocess exited with code 137.
slurmstepd: error: Detected 1 oom_kill event in StepId=3.batch.

Root cause

In mriqc/workflows/shared.py, the synthstrip node only sets num_threads:

synthstrip = pe.Node(
    SynthStrip(num_threads=omp_nthreads),
    name='synthstrip',
    num_threads=omp_nthreads,
)

Other compute-heavy nodes in the same codebase correctly declare their memory — e.g., SpatialNormalization has mem_gb=3, segmentation has mem_gb=5 — but SynthStrip was missed when the thread-limiting fix landed in #1101.

Proposed fix

Add mem_gb to the node so the scheduler properly gates concurrent instances:

synthstrip = pe.Node(
    SynthStrip(num_threads=omp_nthreads),
    name='synthstrip',
    num_threads=omp_nthreads,
    mem_gb=6,
)

Prior work

This is a continuation of #1004, which was closed by #1101. That PR correctly added thread limiting (-n flag / torch.set_num_threads()), but the memory scheduling side was not addressed.

Environment

• MRIQC 24.0.2 (via Neurodesk neurocontainer mriqc_24.0.2_20241108)
• Slurm-managed workstation, 16 CPUs, 32 GB RAM, 24 GB allocated
• BIDS dataset: 4x T1w + 2x T2w for a single participant across multiple sessions