Sched: Reintroduce less synchronizations between token, with fixed pipeline parallelism.#20793

ggerganov merged 6 commits into

ggml-org:masterggml-org/llama.cpp:masterfrom

aendk:akieslinger/rework-reduce-per-token-syncsaendk/llama.cpp:akieslinger/rework-reduce-per-token-syncsCopy head branch name to clipboard