Skip to content

preflight: rocm7-0-34-4-dynres for the v0.34.4 fold, measured on gfx1151 - #378

Merged
glennneuber merged 1 commit into
task/upstream-sync-0.34.4from
preflight/rocm7-0-34-4
Sep 24, 2026
Merged

glennneuber merged 1 commit into
task/upstream-sync-0.34.4from
preflight/rocm7-0-34-4

Conversation

@glennneuber

Copy link
Copy Markdown

This PR adds gate 5's rocm7 profile for the v0.34.4 fold. It targets #375's branch, not main. I measured and ran it on gfx1151 only, and it touches no other platform's profile.

What

rocm7-0-34-4-dynres in expectations.toml. Until now, no rocm7 profile admitted a version after 0.34.1, so preflight on the ROCm host could not resolve the fold at all (exit 2).

  • Pin. llama_cpp_build = "161755f29" (b11081), read through the container route preflight: read llama-server's sha from its version line, not by position #376 fixed.
  • Toolchain. toolchain_build = "rocm-7.2.4". The image is built with Dockerfile.rocm on rocm/dev-ubuntu-24.04:7.2.4-complete (ADR 0042).
  • Patch set. 001 002 004 005 801 802 903. There is no 906: upstream's own HIP revert shipped by b10969, and 906 was retired with it.
  • Version pattern. ^0\.34\.[34]-dynres-\d+-g… matches the fold's gate images before v0.34.4-dynres is cut and after. A 0.34.3 build on b10969 also matches, and then fails payload_pin, which is the right answer for it.
  • Ladders. measure_ladder.py measured them on the fold image, and I pasted its output whole (ADR 0012 rule 8). They equal rocm-0-34-1-dynres's exactly: nemotron [266, 266, 578, 2306, 3270], gemma4 [1102]×5, qwen35 [1034, 1034, 1034, 2306, 4082]. That is expected, because tools/mtmd changed between b10969 and b11081 by one error-handling hunk only.

First run: VERDICT PASS

preflight.py --platform rocm7 --quality against maxusai-ollama:0.34.3-dynres-5-g29ae523-rocm7-gfx1151 (built from 29ae52351), on a container of its own, gave PASS=20 SKIP=12:

  • payload_pin, toolchain_pin, image_tag and go_patch_marker;
  • the three token ladders 5/5, the three payload proofs, and the text baselines 19/19/13;
  • the pinned budgets: nemotron 3328 → 3270 and gemma4 560 → 529;
  • think_format on all three archs through the fold's single pass: valid JSON after thinking, in 1401, 127 and 273 tokens;
  • endpoint_exclusive: no contention.

All the skips predate this profile. No rocm7 profile has recorded a poison-probe, aspect or extraction-quality expectation. The run record is committed as runs/preflight-rocm7-0344-fold-g29ae523.json. It is force-added past the runs/ ignore, as the other cited records are.

Verification

  • python3 test_verdicts.py passes: 196 tests OK, 6 skipped.
  • tomllib loads the file.
  • To reproduce on gfx1151, start the fold image on a free port and run:
    preflight.py --host <it> --platform rocm7 --container <it> --quality
    

amd-server/rocm-gfx1151

🤖 Generated with Claude Code

The v0.34.4 fold's gfx1151 profile. No rocm7 profile admitted a version after 0.34.1, so
preflight on this host could not resolve the fold at all.

- Ladders: measure_ladder.py on the fold image
  (maxusai-ollama:0.34.3-dynres-5-g29ae523-rocm7-gfx1151, Dockerfile.rocm on Ubuntu 7.2.4).
  Its output is pasted whole, and every ladder equals rocm-0-34-1-dynres's.
- Pin: llama_cpp_build = 161755f29 (b11081), read through the container route of #376.
- Patchset: 001 002 004 005 801 802 903. There is no 906, retired at b10969.
- version_pattern: covers the fold's pre-tag and post-tag stamps.

First run: VERDICT PASS, PASS=20 SKIP=12. That includes the pinned budgets and think_format on
all three arches through the single pass. The run record is committed next to the profile
(force-added past the runs/ ignore, as the other cited run records are).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@glennneuber
glennneuber merged commit e914470 into task/upstream-sync-0.34.4 Sep 24, 2026
2 checks passed
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