Speculative Decode

speculative_decode stresses alternating draft and verification-style work to exercise a different compute and cache mix from ordinary autoregressive decode.

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

Area	What it tests
Execution	Repeated draft/verify-style work phases.
Memory	Synthetic token-state and cache reads.
Telemetry	Cache, DRAM, power, clock, and thermal response.
Result	Synthetic verified tokens per second.

It does not implement a particular draft model, acceptance policy, or serving runtime; use it to compare underlying hardware behavior.