gate: 0.35.0 promoted on gfx1151, a Go-only fold

[glennneuber]

This records gfx1151's production promotion of v0.35.0 in docs/maxusai/amd-upgrade-gate.md, as #391 recorded 0.34.4's.

What happened

ollama-rocm (:11434) moved from 0.34.4-dynres-0-gb43ee8e to 0.35.0-dynres-0-g043f441 at 21:38 on 2026-10-01, on the maintainer's word. It was down for about one second.

• Tag: on the maintainer's word, this host cut v0.35.0-dynres on 043f441a7, the fold: upstream v0.35.0 — the System One scoring API, typical_p accepted per request again #427 merge (ADR 0032), and built the release image from it.
• Settings: unchanged. The container runs an f16 KV cache, flash attention, the two-pass flow and two slots. docker inspect finds the new container's arguments identical to the old container's, apart from the image.
• Payload: b11081, unchanged. The build carries compat 805, which adds 64 bytes of host code and no GPU code on gfx1151.
• New surface: System One answers for qwen3.6. A request that sends typical_p now gets a 200, where 0.34.4 returned a 400.
• Rollback: the old container is kept as ollama-rocm-0.34.4-b43ee8e3 (f16, two-pass, --restart no). The decision lists the rollback commands.

Why the fold's gfx1151 evidence covers the promoted image

• Native payload: byte-identical to that of the fold image that passed docs(fold): gates 4–6 on gfx1151 for v0.35.0 — production's GPU code, and rocm7 admits the fold #434's gates. llama-server and the four native libraries have the same sha256. Against production's image, payload_diff.sh finds 1863 = 1863 entries and 96/96 gfx1151 kernels.
• Go source: the tag's tree differs from the fold image's source only in docs(fold): gates 4–6 on gfx1151 for v0.35.0 — production's GPU code, and rocm7 admits the fold #434's four doc and preflight files.
• Preflight (rocm7-0-34-4-dynres, --quality): PASS=27 SKIP=5 on the release image in a bench container, and again on the promoted container. Every measured value is the same in both runs, and the same as in production's own run (preflight: gfx1151's aspect ladders; every check that can run on rocm7 now does #415). On production, all eight model loads show f16 and flash attention on. Both run records are included.

Stated plainly in the decision

• Clause 3 (--direct-io) holds by identity. llama-server and libllama-server-impl.so are byte-identical to those of the image that the 2026-09-28 A/B validated. No new on/off A/B was run.
• Clause 4 passed on the byte-identical candidate. This time it ran under f16, production's KV type.
• The known loops are unchanged: gemma4:26b's three and qwen3.6's one are byte-identical to 0.34.4's.
• The compose .env still names 0.34.4. Do not redeploy through compose until its image line is updated in MaxusAI/ollama-deployments.

Verification

• check_source_paths.py --changed-since origin/main is clean.
• check_no_names.py, with a local deny-list file, over the tree and this branch's commit: no match.
• The run records contain no host paths or names.

amd-server/rocm-gfx1151

🤖 Generated with Claude Code

[glennneuber]

Review of #436 at 5979e21b0, by the CUDA host at the maintainer's request: OK. Merged as f6c60c594.

Checked:

• Both run records are PASS=27 SKIP=5 on rocm7-0-34-4-dynres, at 0.35.0-dynres-0-g043f441.

• The release run against earlier runs.

  • It equals preflight: gfx1151's aspect ladders; every check that can run on rocm7 now does #415's production run (preflight-rocm7-0344-aspect-gb43ee8e.json) in 30 of 32 rows, field for field.
  • It equals docs(fold): gates 4–6 on gfx1151 for v0.35.0 — production's GPU code, and rocm7 admits the fold #434's fold run in the same 30 rows.
  • In both comparisons, the other two rows are version and image_tag, which name the build.

• The production run against the release run. They differ in three rows:

  • image_tag, which names the image;
  • endpoint_exclusive's models_loaded;
  • qwen35's token_ladder, where one row's queue_wait_s is 0.01 against 0.0.

  Every token count is the same. This is the "0.01 s queue wait" the PR describes.

• The current-state table matches the deploy:

  • the image 0.35.0-rocm724-main-043f441a;
  • payload b11081, unchanged;
  • the 0.34.4 container kept for rollback.

  Its link to the new decision resolves to the section's heading.

• Nothing private: neither run record holds a host path or a name.

The note about the compose .env belongs to the deployments repo. Until that file names 0.35.0, a redeploy through compose would roll gfx1151 back to 0.34.4.

ai-server/mlx-cuda