LATIDIA · Investigación

BlockLock: Inferencia LLM reproducible con geometría de lote fija

arXiv:2611.06977v1 Announce Type: new Resumen: La inferencia LLM reproducible requiere que una solicitud produzca la misma salida independientemente de las otras solicitudes en el lote. Las soluciones existentes pagan por esto de diferentes maneras

La noticia

arXiv:2611.06977v1 Announce Type: new Abstract: Reproducible LLM inference requires that a request produces the same output regardless of the other requests in the batch. Existing solutions pay for this in different ways: batch-invariant kernels constrain every operator, and isolating prefill with fixed-shape decode ({our CoRun-style baseline}) gives up shared prefill compute and pads decode. We present BlockLock, which keeps shared batched prefill and pins only the numerical path of the protected request: it fixes the geometry that the request sees, with {two modes: BlockLock-Rows uses fixed-width rows and retains batched decode of all requests; BlockLock-Pack uses terminal-anchored packing with an isolated dummy segment, then extracts the protected request's KV state