Commit History

  1. Validate before the GPU call, so an invalid request costs no GPU time

    Verified signature
    7c16176
  2. Discriminate questions by type so validation errors name the actual defect

    Verified signature
    8ef3e6c
  3. Mark context-overflow errors as invalid requests, like validation errors

    Verified signature
    1cac566
  4. Drop flash-linear-attention (25 s Triton compile per cold worker for 5 ms warm gain); GPU duration 20 s from measurement

    Verified signature
    1c06a4c
  5. Add flash-linear-attention for the Gated DeltaNet kernels

    Verified signature
    1960443
  6. Kev-9B demo + systemone API (vendored kev code from jaredpalmer/kev)

    Verified signature
    1f7aab1
  7. initial commit

    Verified signature
    cd834fc