yahavb

Benchmark and profile Rolling Forcing (Wan2.1-T2V-1.3B text-to-video, TP4×SP2, 8 NeuronCores) on AWS Trn2 via a Kubernetes Job, and emit a grounded report from `kubectl logs`. The job runs the 3-run methodology (warmup → clean fps → profiled), assembles a neuron-explorer.harmony.a2z.com upload bundle, profiles EVERY executed NEFF for an engine-utilization + MFU table with a priority-ranked optimization list, and persists everything (incl. generated mp4 frames) to one timestamped run dir on the S3-backed PVC: /var/mdl/rolling-forcing/runs/<ddmmyyHHMMSS>/. Use when the user says "benchmark rolling forcing on trn2", "profile RF", "generate the RF neuron report", "get RF fps / MFU", "build the harmony trace bundle for RF", or "compare rolling forcing vs streamdiffusion fps". The companion skill `neuron-3run-benchmark` is the general version; this one is wired to RF's actual entry point (e2e_pipeline.py), topology, and outputs.