[glennneuber]

In short

• 0.35.0 changes nothing on the H100.
  • GGUF output is identical to 0.34.4's in all 84 campaign arms and all 3,000 OCRBench answers.
  • MLX moves no more than 0.34.4 moves against itself.
  • OCRBench scores repeat, except qwen3.6 at 877 against 878.
• Keep 908. On Hopper too, llama.cpp's own tiling leaves more gemma4:26b-a4b think-on cases in loops that never
  end: 4 of 27 without 908, 1 with it. The sm_120 host gave 6 and 1. Both arms reproduced byte for byte on a second
  run.
• gemma4:31b-it-q4_K_M's prefill drift is not the build, the prompt cache or the serving path. Every
  configuration measures the same prefill today, so 2026-09-30's figure stays unexplained.
• The prompt cache costs time outside the prefill timer: about 4 % per campaign cell, and 21 % per OCRBench item
  on gemma4:31b-it-q4_K_M.
• MLX-CUDA clamps gemma4's image budgets to 70–1120, read from behaviour. budgets_observed stays false.

What's in it

• New:
  • docs/maxusai/vision-campaign-2026-10-04-h100-sm90.md;
  • its bundle (35 cells), host profile and budget readings under vision-suite/bench-runs/.
• OCRBench:
  • eight new arms (bench-runs/ocrbench/ext_h0350-*) and their host profile;
  • a 2026-10-04 section in ocrbench-h100.md;
  • a note in ocrbench.md.
• Records:
  • The v0.35.0 fold record gains the H100's gates 4–6 and its deploy. Its preflight-profile table now reflects preflight: CUDA profiles admit 0.35.0; payload pins on native Linux #444.
  • The 908 register row gets the sm_90 measurement.
  • llama-server-prompt-cache.md and the 2026-09-30 campaign document point to the new evidence.
  • [expect.mlx-cuda.gemma4] gets a dated comment. No value changes.

Checks

• Every table renders verbatim from committed data, with summarize_engine_compare.py --bundle and
  summarize_extbench.py. The derived counts come from snippets in the documents that read only committed files.
• Local runs pass: test_verdicts.py (238 OK), test_bundle_campaign.py, test_summarizers.py,
  test_extbench.py and check_source_paths.py. The no-names check needs CI's deny-list.
• The data was scanned for hostnames, home paths, IDs and emails. The only hit is an email address printed in an
  OCRBench test image, already in the 2026-10-01 files.

Not in it

• No code, and no new scripts.
• The runs' logs, raw responses and thinking traces stay outside the repo. The VM they ran on is deleted on
  2026-10-07.

🤖 Generated with Claude Code