SonicSampler unifies LLM sampling and speculative verification into one Triton kernel.

Ponnusamy, Sahni, Wang, and Tri Dao attack a real serving problem: existing sampling implementations accelerate only parts of the logit-processing/token-selection/verification pipeline, need multiple kernel launches, or assume every request in a batch samples identically, which breaks CUDA Graph execution for dynamic workloads (arXiv 2607.20475). Their kernel suite handles grammar-constrained decoding, penalties, filtering, and speculative verification in a single batched kernel while staying CUDA Graph compatible, with hierarchical two-stage top-k giving up to 10x, and up to 16x across heterogeneous workloads. Drop-in win if you run mixed sampling configs behind one endpoint.