Commit History

  1. Disable 4b (RAM), serve 0.8b when model=4b requested

    AustraliaSilver
    d03b9b6
  2. Add /api/sysinfo (CPU model, cgroup limits, mem)

    AustraliaSilver
    f730aca

  1. Web UI: model cards + live NLI playground with probability bars

    42898ca
  2. 0.8b fp32 (bf16 emulation slow on CPU), 4b stays bf16

    c19fe0d
  3. nyu-mll/glue instead of glue (loading script blocked)

    9aaace9
  4. openjev NLI cross-encoder bench (0.8B + 4B, MNLI eval)

    07b8f23

  1. Dynamic int8 quantization on encoder (LAYA_QUANTIZE)

    62be1c0
  2. Replace kev with Laya typed-decisions server (TypeSafe /v1/systemone shape)

    88a68ec