The TurboQuant conversion of one bf16 entry: encode the source's live window (consuming the source), then the shared bookkeeping. The converted cache attends with the source's TurboQuant read (turboQuantKernels); fusedDecode is its decode policy.
turboQuantKernels
fusedDecode
The TurboQuant conversion of one bf16 entry: encode the source's live window (consuming the source), then the shared bookkeeping. The converted cache attends with the source's TurboQuant read (
turboQuantKernels);fusedDecodeis its decode policy.