Peak bytes of the SDPA-fallback scores tensor live during one prefill
chunk of chunk queries against a ctx-token context: the last chunk
sees the longest keys, and the lazy graph keeps one layer's scores live
at a time. Zero when every attention layer's head dim fuses. Sliding
layers see at most window + L keys. Admission-only (fit and the
in-flight memory guard); nothing here changes what is computed.
Peak bytes of the SDPA-fallback scores tensor live during one prefill chunk of
chunkqueries against actx-token context: the last chunk sees the longest keys, and the lazy graph keeps one layer's scores live at a time. Zero when every attention layer's head dim fuses. Sliding layers see at mostwindow + Lkeys. Admission-only (fit and the in-flight memory guard); nothing here changes what is computed.