Study MI300A partition-mode effects on training

The MI300A exposes memory/compute partitioning modes that change how memory bandwidth and capacity are presented to processes. These may interact with unified-memory behavior, DataLoader host-memory pressure, and per-rank throughput in ways discrete GPUs don't exhibit.

Possible study: enumerate available partition modes on the target platform, run a fixed training configuration under each, and report throughput, memory behavior, and any failure modes.

Start by identifying the target MI300A platform and the fixed training configuration to use for the study. Enumerate its available partition modes, run the same training workload under each, and report throughput, memory behavior, DataLoader host-memory pressure, per-rank effects, and failure modes.

Written by the indexing model from the issue text.