Skip to content

preflight: gfx1151's aspect ladders; every check that can run on rocm7 now does - #415

Merged
glennneuber merged 1 commit into
mainfrom
preflight/rocm7-aspect-ladder
Sep 29, 2026
Merged

glennneuber merged 1 commit into
mainfrom
preflight/rocm7-aspect-ladder

Conversation

@glennneuber

Copy link
Copy Markdown

The rocm7-0-34-4-dynres profile recorded no aspect ladder, so check_aspect_ladder read SKIP on all three arches. That was the last check skipped for want of an expectation on gfx1151. This records the three ladders, at the maintainer's word.

Measured

Measured 2026-09-29 on the promoted image 0.34.4-dynres-0-gb43ee8e, in a bench container with production's environment. The measurement makes the check's own probe calls: image_prefix (B8), then visual_tokens at 4:3, 4:1 and 1:1, the axis the 16:9 ladder cannot move. Each aspect_ladder line is the generator's output (aspect_measure.py in the run directory), pasted whole (ADR 0012 rule 8).

arch model 1024x768 1920x480 768x768 why
nemotron_h_omni nemotron3:33b-q4_K_M 770 902 578 native grid at stride 32
gemma4 gemma4:31b-it-q4_K_M 1066 1058 1091 fills the 1120-token budget at stride 48; equal to mlx-metal's row
qwen35 qwen3.8:27b-q4_K_M 1038 1026 1026 raised to the 1024-token minimum at stride 32

Each value is its grid plus the two marker tokens. The comments in expectations.toml give the grids.

Verified

A preflight with the new entries ran on the same image and environment, with --quality: VERDICT PASS, PASS=27 SKIP=5, against 24 and 8 before.

  • The aspect ladder passes 3/3 on every arch. Quality (1.00 on every measure) and the fp16 canary pass as before.
  • The five that still skip are the four Metal and MLX checks and qwen35's pinned budget. None of them applies on gfx1151, so every check that can run on rocm7 now does.
  • The run record is committed as runs/preflight-rocm7-0344-aspect-gb43ee8e.json. The README's matrix columns do not change. Its source comment now names this run as gfx1151's newest, and the regenerated table equals the committed one byte for byte.
  • The gate's baseline section records the step, and the profile's own note lists what was recorded after its first run.

Checks

  • test_verdicts.py: OK.
  • check_source_paths.py --changed-since origin/main is clean.
  • The run record contains no host paths or names, and the name scan of the diff finds nothing.

amd-server/rocm-gfx1151

🤖 Generated with Claude Code

…eck that can run here now does

The rocm7 profile recorded no aspect ladder, so check_aspect_ladder read SKIP on all three arches. This records
them, measured 2026-09-29 on the promoted image 0.34.4-dynres-0-gb43ee8e in production's environment by the probe
calls the check makes (image_prefix, then visual_tokens at 4:3, 4:1 and 1:1); each line is the generator's
output (aspect_measure.py in the run directory), pasted whole.

- nemotron_h_omni 770 / 902 / 578: the native grid at stride 32, plus the two markers.
- gemma4 1066 / 1058 / 1091: the 1120-token budget fill at stride 48; equal to mlx-metal's gemma4 row.
- qwen35 1038 / 1026 / 1026: every ratio raised to the 1024-token minimum at stride 32.

A preflight with them, on the same image and environment, passed: VERDICT PASS, PASS=27 SKIP=5 (was 24/8), the
aspect ladder 3/3 on every arch. The five skips left are the four Metal and MLX checks and qwen35's pinned budget,
none of which applies on gfx1151. The run record is committed; the README's matrix, whose columns it does not
change, now names it as gfx1151's newest run, and the gate records the step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant