perf(decode): replayable transducer CUDA graphs (per-step + batched)#46

Open

sims1253 wants to merge 3 commits into

mudler:masterfrom

sims1253:perf/replayable-decode-graph

Commits on Jun 29, 2026

perf(decode): replayable per-step transducer graphs (GPU CUDA-graph capture)
sims1253
committed
perf(decode): coalesce per-step graph inputs into one host->device copy
sims1253
committed
perf(decode): replayable BATCHED transducer graphs (per-N CUDA-graph capture)
sims1253
committed