Commit History

Today

  1. VRAM Match v7.5: display reserve measured (1.01-1.16 GiB bracket, reserve moves to 1.16), headless gate stays 0.96 with the measured range stated, every row labelled measured-check or no-measured-check, and a correction to my own v7.4 free-space figure (0.66 -> 0.50)

    8fbb027

2 days ago

  1. VRAM Match v7.4: input embedding stays off the card (llama.cpp dev_input=CPU), q8_0 is 1.0625 B/value, recurrent state for hybrid models, measured 96% headless gate, display reserve as ~1.1 GiB, cliff-shaped speed decay, and the 27B IQ4_XS row that closes to 0.05 GiB

    cbf3e64

3 days ago

  1. VRAM Match v7.3: third 5080 point (27B IQ4_XS at 75.6%), the measured spill cliff (~300 MB -> 6x slower), headless gate called out as ~0.2 GiB conservative at the edge

    4975f8d

4 days ago

  1. README: v7.2 per-era speed anchors, the 3x break stated, and the unexplained Pascal gap left as a span

    f80676f
  2. VRAM Match v7.2: efficiency anchor per card era (modern 62-78% of roofline, from a published RTX 5080 run set)

    Echo
    3c8a996
  3. VRAM Match v7.1: one-click receipt on every row (terms labelled, snapshot date, feedback address)

    5f0d903

5 days ago

  1. VRAM Match v7.0: gate on usable VRAM, reserve grows with context

    f933c43

14 days ago

  1. Link measured-anchors dataset

    513ac77
  2. VRAM Match v6.0 as a static Space

    66a1f37