Transformer Train Step

transformer_train_step applies controlled forward, backward, gradient, and optimizer-style arithmetic pressure.

pantheon --test transformer_train_step --duration 60 --gpu 0 --mem 50 --verify --profile

Area	What it tests
Compute	Forward, gradient, and update-style arithmetic.
Memory	Synthetic activation and state working-set reads.
Telemetry	SM/cache/DRAM activity, power, clocks, and thermals.
Result	Synthetic training steps per second.

It is not a convergence or time-to-train benchmark. Validate training quality, collectives, optimizer selection, and framework behavior separately.