Abstract
Systems and methods generate multi-codebook semantic identifiers for generative retrieval using parallel discrete flow matching, latent-compressed encoder conditioning, and fused GPU execution. A user context is encoded to an encoder representation E, which is compressed to a latent-compressed encoder state cKV = E * Wdown with dc << D. N candidate discrete states over K codebooks are initialized from noise and iteratively denoised for T steps, updating all K codebook positions in parallel according to per-codebook schedules sigma_k(t) = t^{alpha_k} with learnable alpha_k. Cross-attention conditioning is computed over cKV using absorbed projection operations. Flow updates and candidate scoring execute within a single fused GPU kernel launch that may be persistent, retaining cKV in on-chip cache across steps and selecting top-n outputs.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Integrated System for Parallel Codebook Generation via Discrete Flow Matching with Latent-Compressed Encoder States and Fused GPU Kernel Execution", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10943