Merge commit '52d42686560a9e8f441f9b9780c8890c37d2802d' into concedo_experimental

# Conflicts:
#	.github/actions/windows-setup-cuda/action.yml
#	.github/workflows/build-riscv.yml
#	.github/workflows/build-sanitize.yml
#	.github/workflows/build-self-hosted.yml
#	.github/workflows/release.yml
#	.github/workflows/server-sanitize.yml
#	docs/build-s390x.md
#	docs/build.md
#	ggml/CMakeLists.txt
#	ggml/cmake/common.cmake
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml-cpu/repack.cpp
#	ggml/src/ggml-cuda/CMakeLists.txt
#	ggml/src/ggml-cuda/fattn.cu
#	ggml/src/ggml-hexagon/ggml-hexagon.cpp
#	ggml/src/ggml-hexagon/htp/hvx-sin-cos.h
#	ggml/src/ggml-hexagon/htp/rope-ops.c
#	ggml/src/ggml-hip/CMakeLists.txt
#	ggml/src/ggml-musa/CMakeLists.txt
#	ggml/src/ggml-vulkan/ggml-vulkan.cpp
#	pyproject.toml
#	requirements/requirements-convert_legacy_llama.txt
#	requirements/requirements-gguf_editor_gui.txt
#	requirements/requirements-server-bench.txt
#	requirements/requirements-tool_bench.txt
#	scripts/make-release-checks.sh
#	tests/test-backend-ops.cpp
#	tests/test-jinja.cpp
#	tools/cli/README.md
#	tools/completion/README.md
#	tools/llama-bench/llama-bench.cpp
#	tools/server/README.md

common/speculative.cpp

@@ -1094,8 +1094,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
         // Target prefill may contain token IDs or multimodal embeddings. Both
         // produce the target-layer features used to seed the draft KV cache, so
-        // skipping the embedding batches leaves a hole in the draft's cache and
-        // the next injection fails to initialize.
+        // embeddings are injected too, except the pinned ones skipped below.
         // TODO: revisit after https://github.com/ggml-org/llama.cpp/pull/24669 is merged
         const bool has_tokens     = batch_in.token != nullptr;
         const bool has_embeddings = batch_in.embd  != nullptr;
@@ -1131,6 +1130,13 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             }
             const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1;
 
+            // an M-RoPE image pins all its rows to one position, so a windowed draft
+            // cache cannot free cells for it - skip it, the draft can jump over the gap
+            const bool pos_pinned = batch_in.pos[i_batch_beg[seq_id]] == batch_in.pos[i_batch_end[seq_id]];
+            if (has_embeddings && n_rows > 1 && pos_pinned) {
+                continue;
+            }
+
             for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
                 const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);