sie_id: convaiinnovations/laya description: "English typed-decision classifier (Laya): answers noul, choice and score questions, one row per question. 512-token rows." release_date: "2026-09-23" hf_id: convaiinnovations/laya hf_revision: aa8c91ca088ec597df95a0d1c76b3063cb2ae5e8 inputs: text: true image: false audio: false video: false tasks: encode: null score: null extract: {} max_sequence_length: 512 profiles: default: max_batch_tokens: 16384 compute_precision: bfloat16 adapter_path: sie_server.adapters.laya.adapter:LayaAdapter adapter_options: loadtime: {} runtime: {} # One state with a few questions is a small forward that kernel launches, # not the GPU, bound. This profile replays its encoder as a CUDA graph # (request it as `MODEL:graphs`). It is not the default because a graph # moves probabilities by more than the 0.02 the server README allows a # default speed-up, though by less than batching already does: see # "Laya CUDA graphs" there. graphs: extends: default adapter_options: loadtime: cuda_graphs: bucketed runtime: {}