Large models on multiple GPUs

	Recommended use
Start with	Step 3.7 Flash on the qualified 2× RTX PRO 6000 Blackwell PP-2 path
Topology	Pipeline parallelism for the published Step configuration
Validate	Placement, cross-device transport, admission, and end-to-end output on the exact topology
Read next	Step model card, Serving: PP-2

Pipeline stages, tensor parallelism, expert parallelism, and replicas are different execution shapes. Use the name and gate for the topology actually running; do not call PP or replicas TP.