diff --git a/llama/compat/906-revert-hip-integrated-flag.patch b/llama/compat/906-revert-hip-integrated-flag.patch
new file mode 100644
index 00000000000..b39097fe8fb
--- /dev/null
+++ b/llama/compat/906-revert-hip-integrated-flag.patch
@@ -0,0 +1,50 @@
+MAXUSAI 906: carry upstream's revert of "restore prop.integrated on HIP builds".
+
+Upstream llama.cpp PR #28604 (d4389a4dd92, 2026-09-08T14:19:53Z) reverts
+c7d8722922a (#24233), which set info.devices[id].integrated = prop.integrated on
+HIP builds. On an integrated/UMA device that flag changes which buffer types the
+HIP backend claims, and upstream issue #28211 root-causes the result as an MMQ
+tile-barrier race on gfx115x that produces WRONG OUTPUT with no crash and no
+warning, once a decode exceeds n_ubatch.
+
+WE MISSED THE FIX BY 78 MINUTES. Our payload b10864 is 5d806aa25, 2026-09-08
+13:01:03Z. The revert landed at 14:19:53Z the same day, so ollama v0.34.1 has the
+defect and not the fix.
+
+MEASURED HERE (gfx1151, Radeon 8060S, ROCm 7.2.x), gemma4:31b-it-q4_K_M with the
+image-token budget pinned per cell, name_bbox_mean_iou:
+
+    budget   0.32.1/b9888   0.34.1/b10864
+       280       0.612          0.691
+       560       0.713          0.714
+      1120       0.728          0.000   <-- total loss
+
+Only the top rung fails, and only on the newer payload. prompt_eval_count is
+identical between builds on every test, so the grid and token count are unchanged
+and the pixels reaching the model are the same -- consistent with corruption
+downstream of preprocessing rather than in it. Related upstream reports on this
+exact silicon: #27419 (upper image content lost on HIP, Vulkan and CPU correct)
+and #28211 (wrong logits past n_ubatch).
+
+GGML_CUDA_NO_PINNED=1 is NOT a workaround: it reaches the runner and the runner
+aborts (signal: aborted, core dumped) on model load.
+
+Drop this patch when the payload advances past d4389a4dd92.
+
+Files: ggml/src/ggml-cuda/ggml-cuda.cu
+diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
+index 0e6601f..38bd4c9 100644
+--- a/ggml/src/ggml-cuda/ggml-cuda.cu
++++ b/ggml/src/ggml-cuda/ggml-cuda.cu
+@@ -304,11 +304,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
+ 
+         info.default_tensor_split[id] = total_vram;
+         total_vram += device_vram;
+-#if defined(GGML_USE_HIP)
+-        info.devices[id].integrated = prop.integrated;
+-#else
+         info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
+-#endif
+         info.devices[id].nsm        = prop.multiProcessorCount;
+         info.devices[id].smpb       = prop.sharedMemPerBlock;
+         info.devices[id].warp_size  = prop.warpSize;