[dzannotti]

Problem

The upstream sync in #28 was landed as a squashed PR because divergence required manual
conflict resolution. The resolution silently dropped two upstream commits entirely:

• 38a5b42d9 — HIP: Enable AllReduce for ROCm (ggml-org#27825)
• 9f31776c3 — opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP (ggml-org#27637)

Both were dropped completely — every file each commit touched is unmodified from the
pre-sync merge-base — so there is no half-applied state and no build breakage today. The
observable effect is that ROCm AllReduce init returns nullptr, the dispatcher in
ggml-cuda.cu treats that as "init failed", and multi-GPU ROCm silently falls back to the
meta backend's generic AllReduce. The OpenCL MoE batch-size heuristic is simply absent.

How this was found

Of the 251 files touched by the 72 upstream commits in the sync window, 228 are byte-identical
to upstream. For the remainder, any file that upstream modified but that we never modified
since the merge-base is by definition a dropped change:

MB=$(git merge-base HEAD upstream/master)
git diff --name-only $MB upstream/master | while read -r f; do
    git diff --quiet HEAD upstream/master -- "$f" && continue
    n=$(git diff --numstat $MB HEAD -- "$f" | awk '{print $1+$2}')
    [ -z "$n" ] && echo "DROPPED: $f"
done

Change

Both commits cherry-picked with -x. Both applied with zero conflicts. Re-running the audit
against this branch returns empty — all 72 upstream commits are now accounted for.

Testing

Not built locally. These touch only HIP and OpenCL backend paths, which do not compile in a
default CPU/CUDA build, so standard CI will not exercise them.

Needs GGML_HIP=ON and GGML_OPENCL=ON builds before merge.

Follow-up

Once this lands, master still reports "72 commits behind" because the squash recorded no
upstream ancestry. That is fixed separately with git merge -s ours upstream/master, which
keeps our tree byte-identical and only records the parent link. See the SYNC.md PR.