compat 906: carry upstream's revert of prop.integrated on HIP — fixes the gfx1151 vision regression

[glennneuber]

Fixes the vision regression that made ollama 0.34.x unusable on gfx1151, by carrying an upstream revert that landed 78 minutes after our payload was cut.

The defect

Upstream c7d8722922a (llama.cpp#24233) set info.devices[id].integrated = prop.integrated on HIP builds. On an integrated/UMA device that changes which buffer types the backend claims, and llama.cpp#28211 root-causes the consequence as an MMQ tile-barrier race on gfx115x producing wrong output with no crash and no warning, once a decode exceeds n_ubatch.

Upstream reverted it in #28604, merged 2026-09-08T14:19:53Z. Our payload b10864 is 5d806aa25, 2026-09-08T13:01:03Z. We are on the wrong side of it by 78 minutes.

Two independent upstream reports describe our exact symptom on our exact silicon: #27419 ("upper image content is lost while Vulkan/CPU are correct" — Ryzen AI Max+ 395 / Radeon 8060S / gfx1151 / Qwen3.8-27B) and #28211.

Before and after

Five-model campaign against the 2026-09-17 baseline (#303), same runner, same digests, same environment, one power envelope. Scene bbox IoU:

model	0.32.1 (b9888)	0.34.1 (b10864)	0.34.1 + 906
qwen3.8:27b-q4_K_M	0.991	0.065	1.000
gemma4:31b-it-q4_K_M	0.961	0.922 (3/6 objects)	0.960
nemotron3:33b-q4_K_M	0.857	0.161	0.862
qwen3.6:35b-a3b-q4_k_m	0.953	0.273	0.972
gemma4:26b-a4b-it-q4_K_M	0.973	0.966	0.975

All five return to 6/6 · 6/6 · 6/6 boxes/labels/colours and 5/5 · 5/5 · ✅ invoice extraction. Unpatched 0.34.1 scored 0/5 on invoice for three of them.

This satisfies gate clause 4 — qwen35moe (qwen3.6:35b-a3b) shows 0 degenerate and scores above the baseline.

Why it tracked the image-token budget

gemma4:31b, budget pinned per cell with run_budget_sweep.sh, name_bbox_mean_iou:

budget	0.32.1	0.34.1
280	0.612	0.691
560	0.713	0.714
1120	0.728	0.000

1120 image tokens push the decode past n_ubatch; 280 and 560 do not. Since the fork pins --image-max-tokens 1120 on every gemma4 launch (#309), every gemma4 vision request on an unpatched 0.34.x build was hitting the broken path.

Refuted along the way, each by experiment

Recorded so nobody re-runs them:

• fp16 accumulation in the vision tower — GGML_CUDA_CUBLAS_COMPUTE_TYPE f32 / f16 / unset all equally broken, env var verified to reach the llama-server process
• multi-sub-batch image decode — at num_batch 2048 the image decodes in a single decoding image batch 1/1, n_tokens_batch = 2040, still broken
• the b10864 resize-algo flip (#27594) — reverted as a compat patch, verified present in-binary by output-hash diff at identical prompt_eval/eval, fixed nothing. That patch (905) is dropped in this branch rather than left as dead weight.
• the gemma4 budget bump (#28335) — creates the rung that exposes the bug, is not the bug
• GGML_CUDA_NO_PINNED=1 — not a workaround; the runner aborts on load (signal: aborted, core dumped)
• direct I/O, GPU layer placement, host drift, and our own compat series — stock ollama/ollama:0.34.1-rocm reproduces the regression identically

prompt_eval_count is identical between builds at every budget rung, so preprocessing and token counts never moved — the corruption is downstream of the image path entirely.

Verification

Full series 001/002/004/005/801/903/906 applies to a clean b10864 with plain git apply, and the outcome is checked rather than just the exit code: after applying, grep 'integrated = prop.integrated' returns nothing. Build is 2m38s on the ccache mount.

Caveats

• gemma4:31b name_bbox returns to 0.663–0.72 against the baseline's 0.728 — restored, but repeats would be good, especially given the known run-to-run spread on 0.34.x (docs(vision): SOLVED — the 0.34.0 gemma4 vision shift is upstream MLX #3912 #310).
• qwen3.8 multi-image loses q4_bbox_hit where the baseline had it; the anchored variant passes. One cell, unrepeated.
• Drop this patch once the payload advances past d4389a4dd92.

Cross-platform verification requested in #313 — in particular whether this defect class reaches CUDA unified-memory devices, which bears on gate issue ollama#17475.

🤖 Generated with Claude Code

---

Recreated from #314, which GitHub closed when the head branch was renamed from fix/restore-bilinear-resize (a leftover from a refuted hypothesis) to fix/hip-integrated-flag-906. Same commits, same content; #314 carries the earlier discussion.