Use BF16 state at every CUDA batch and stabilize physical-ring attention
#1
by hkust-aaron - opened
Exactly the six files already published in GitHub commit 3f9ff63cb3657ea829ca15eb8deb28d00974fcbd. Formal GPU acceptance 422927/v17 passed two independent reference/release pairs at B1 and B512, including repeatable tokens, actual BF16 storage, independent prefill, cache bytes and peak memory. The reference uses the same corrected attention backend; this is not bitwise equivalence to historical cuDNN outputs. Warmed 4K+2K output TPS medians: B1 230.385 to 210.726 (-8.53%); B512 8270.370 to 8217.190 (-0.64%); peak allocation unchanged. Historical tables remain labeled and unchanged. Source hashes and acceptance artifacts are documented in SOURCE.json.
rockmatrix-ai changed pull request status to merged