[draft] Xuejun/enable arch test

.devops/openvino.Dockerfile

@@ -1,12 +1,12 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2
-ARG OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857
+ARG OPENVINO_VERSION_MAJOR=2026.2.1
+ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.34.4
-ARG IGC_VERSION_FULL=2_2.34.4+21428
-ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
+ARG IGC_VERSION=v2.36.3
+ARG IGC_VERSION_FULL=2_2.36.3+21719
+ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases

.github/workflows/build-cache.yml

@@ -68,8 +68,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone
@@ -96,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/build-openvino.yml

@@ -39,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone
@@ -96,8 +96,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/build-self-hosted.yml

@@ -266,8 +266,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -446,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -506,8 +506,11 @@ jobs:
           cmake -B build/ReleaseOV -G Ninja \
             -DCMAKE_BUILD_TYPE=Release \
             -DGGML_OPENVINO=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
-          cmake --build build/ReleaseOV --config Release -j $(nproc)
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build/ReleaseOV --config Release --parallel
 
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
@@ -521,8 +524,26 @@ jobs:
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          cp LICENSE ./build/ReleaseOV/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/ReleaseOV/bin .
+          dest=./build/ReleaseOV/bin
+          OPENVINO_ROOT=./openvino_toolkit
+          ov_lib="$OPENVINO_ROOT/runtime/lib/intel64"
+
+          # Bundle OpenVINO runtime libs + TBB. Binaries built with RPATH=$ORIGIN
+          # load these siblings without setupvars.sh / LD_LIBRARY_PATH.
+          cp -P "$ov_lib"/libopenvino.so* \
+                "$ov_lib"/libopenvino_c.so* \
+                "$ov_lib"/libopenvino_*_plugin.so \
+                "$ov_lib"/libopenvino_intel_npu_compiler*.so \
+                "$OPENVINO_ROOT"/runtime/3rdparty/tbb/lib/*.so* \
+                "$dest"
+          cp -P /usr/lib/x86_64-linux-gnu/libOpenCL.so.1* "$dest" 2>/dev/null || true
+          cp "$ov_lib"/cache.json "$dest" 2>/dev/null || true
+
+          # OpenVINO licensing
+          cp -r "$OPENVINO_ROOT"/docs/licensing "$dest"/openvino-licensing
+
+          cp LICENSE "$dest"
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C "$dest" .
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -538,8 +559,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
 
     steps:
       - name: Set OpenVINO version output
@@ -607,7 +628,9 @@ jobs:
             -A x64 ^
             -DCMAKE_BUILD_TYPE=Release ^
             -DGGML_OPENVINO=ON ^
-            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake ^
+            ${{ env.CMAKE_ARGS }}
 
           cmake --build build\ReleaseOV --config Release -- /m
 
@@ -624,8 +647,29 @@ jobs:
         id: pack_artifacts
         shell: powershell
         run: |
-          Copy-Item LICENSE .\build\ReleaseOV\bin\
-          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip .\build\ReleaseOV\bin\*
+          # Locate the extracted OpenVINO toolkit root (same pattern as the Build step).
+          $OPENVINO_ROOT = (Get-ChildItem -Directory openvino_toolkit | Select-Object -First 1).FullName
+          if (-not $OPENVINO_ROOT) {
+            Write-Error "OpenVINO toolkit folder not found under .\openvino_toolkit"
+            exit 1
+          }
+
+          $dest = ".\build\ReleaseOV\bin\Release"
+
+          $ovBin = Join-Path $OPENVINO_ROOT 'runtime\bin\intel64\Release'
+          Copy-Item -Path (Join-Path $ovBin '*.dll')       -Destination $dest -Force
+          Copy-Item -Path (Join-Path $ovBin 'cache.json')  -Destination $dest -Force
+
+          $tbbBin = Join-Path $OPENVINO_ROOT 'runtime\3rdparty\tbb\bin'
+          Copy-Item -Path (Join-Path $tbbBin 'tbb*.dll') -Destination $dest -Force
+
+          # OpenVINO licensing
+          $licensingDest = Join-Path $dest 'openvino-licensing'
+          New-Item -ItemType Directory -Force -Path $licensingDest | Out-Null
+          Copy-Item -Path (Join-Path $OPENVINO_ROOT 'docs\licensing\*') -Destination $licensingDest -Recurse -Force
+
+          Copy-Item LICENSE $dest
+          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip $dest\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6

docs/backend/OPENVINO.md

@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail
 
-OPENVINO_VERSION_MAJOR="2026.2"
-OPENVINO_VERSION_FULL="2026.2.0.21903.52ddc073857"
+OPENVINO_VERSION_MAJOR="2026.2.1"
+OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
 
 </details>
 
@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================
 
-set "OPENVINO_VERSION_MAJOR=2026.2"
-set "OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857"
+set "OPENVINO_VERSION_MAJOR=2026.2.1"
+set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
 
 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.2` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
 
 </details>
 

ggml/include/ggml.h

@@ -695,7 +695,9 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        char padding[16];
+        // add a struct ggml_tensor * named org_src, initialized to NULL, for keeping track of original source tensors in case of in-place operations
+        struct ggml_tensor * org_src;
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);

ggml/src/ggml-backend.cpp

@@ -942,42 +942,61 @@ static char * fmt_size(size_t size) {
     return buffer;
 }
 
+static void ggml_backend_debug_print_tensor_info(const struct ggml_tensor * tensor, const char * backend_name) {
+    printf("%20.20s [%5.5s] shape=[%lld,%lld,%lld,%lld] stride=[%zu,%zu,%zu,%zu] offset=%zu",
+        tensor->name,
+        backend_name ? backend_name : "NULL",
+        (long long) tensor->ne[0], (long long) tensor->ne[1], (long long) tensor->ne[2], (long long) tensor->ne[3],
+        tensor->nb[0], tensor->nb[1], tensor->nb[2], tensor->nb[3],
+        tensor->view_offs);
+}
+
+static void ggml_backend_debug_print_tensor_srcs(const struct ggml_tensor * tensor, const char * backend_name) {
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = tensor->src[i];
+        if (src == NULL) {
+            continue;
+        }
+        printf(" ");
+        ggml_backend_debug_print_tensor_info(src, backend_name);
+    }
+}
+
 static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
+    for (int i = 0; i < 200; i++) {
         if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
             ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
-            GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs", cur_split, ggml_backend_name(split_backend),
+            printf("\n## SPLIT #%d: %s # %d inputs", cur_split, ggml_backend_name(split_backend),
                 sched->splits[cur_split].n_inputs);
             for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
                 if (j == 0) {
-                    GGML_LOG_DEBUG(": ");
+                    printf(": ");
                 }
-                GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
-                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+                printf("[%s] ", sched->splits[cur_split].inputs[j]->name);
+
             }
-            GGML_LOG_DEBUG("\n");
+            printf("\n");
             cur_split++;
         }
         struct ggml_tensor * node = graph->nodes[i];
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-        if (sched->debug > 1) {
-            ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d,c=%d:", i, ggml_op_desc(node), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
-                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)], node->flags & GGML_TENSOR_FLAG_COMPUTE ? 1 : 0);
+        if (2 > 1) {
+
+
+
+            printf("node #%3d (%10.10s): %20.20s shape=[%lld,%lld,%lld,%lld] offset=%zu:", i, ggml_op_name(node->op), node->name,
+                (long long) node->ne[0], (long long) node->ne[1], (long long) node->ne[2], (long long) node->ne[3],
+                node->view_offs);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
                     continue;
                 }
-                ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
-                GGML_LOG_DEBUG(" %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
-                    fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+                printf(" %20.20s shape=[%lld,%lld,%lld,%lld] offset=%zu", src->name,
+                    (long long) src->ne[0], (long long) src->ne[1], (long long) src->ne[2], (long long) src->ne[3],
+                    src->view_offs);
             }
-            GGML_LOG_DEBUG("\n");
+            printf("\n");
         }
     }
 }
@@ -1242,6 +1261,28 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         GGML_ASSERT(*cur_backend_id != -1);
     }
 
+    // OpenVINO currently uses ggml tensor names as graph indices. Some models (e.g. gpt-oss and
+    // llama4) can contain duplicate ggml tensor names, so we append node ids here to keep names
+    // unique. This is a temporary workaround and will be further optimized away in the future.
+    {
+        bool has_openvino_backend = false;
+        for (int i = 0; i < sched->n_backends; i++) {
+            if (strcmp(ggml_backend_name(sched->backends[i]), "OPENVINO") == 0) {
+                has_openvino_backend = true;
+                break;
+            }
+        }
+
+        if (has_openvino_backend) {
+            for (int i = 0; i < graph->n_nodes; i++) {
+                struct ggml_tensor * node = graph->nodes[i];
+                char                 new_name[128];
+                snprintf(new_name, sizeof(new_name), "%s#%d", node->name, i);
+                ggml_format_name(node, "%s", new_name);
+            }
+        }
+    }
+
     // pass 5: split graph, find tensors that need to be copied
     {
         int i_split = 0;
@@ -1360,6 +1401,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                                 ggml_set_input(tensor_copy);
                                 ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
                             }
+                            tensor_copy->org_src = src;
                             tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
@@ -1375,6 +1417,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         sched->n_splits = i_split + 1;
     }
 
+    // ggml_backend_sched_print_assignments(sched, graph);
     if (sched->debug) {
         ggml_backend_sched_print_assignments(sched, graph);
     }