llama-cpp-bin 11236.0.0__tar.gz → 11243.0.0__tar.gz

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama-cpp-bin
-Version: 11236.0.0
+Version: 11243.0.0
 Summary: llama.cpp server binary built from source
 Author-email: Vladyslav Tkachenko <vladlearns@gmail.com>
 License: AGPL-3.0-or-later

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/.devops/intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2026.1.1-devel-ubuntu24.04
 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
@@ -19,7 +19,7 @@ RUN npm ci
 COPY tools/ui/ ./
 RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
 
-FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS build
+FROM docker.io/intel/oneapi-toolkit:$ONEAPI_VERSION AS build
 
 ARG GGML_SYCL_F16=ON
 ARG LEVEL_ZERO_VERSION=1.28.2
@@ -59,7 +59,7 @@ RUN mkdir -p /app/full \
     && cp requirements.txt /app/full \
     && cp .devops/tools.sh /app/full/tools.sh
 
-FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS base
+FROM docker.io/intel/oneapi-toolkit:$ONEAPI_VERSION AS base
 
 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/.github/workflows/build-sycl.yml

@@ -48,7 +48,7 @@ jobs:
 
     env:
       ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      ONEAPI_INSTALLER_VERSION: "2026.1"
       LEVEL_ZERO_VERSION: "1.33.1"
       LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
 
@@ -63,8 +63,8 @@ jobs:
         shell: bash
         run: |
           cd /tmp
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/5996e26b-f48a-42b1-8db0-b002ad0bd8d7/intel-oneapi-toolkit-2026.1.1.33_offline.sh -O intel-oneapi-toolkit_offline.sh
+          sudo bash intel-oneapi-toolkit_offline.sh -s -a --silent --eula accept
 
       - name: Install Level Zero SDK
         shell: bash
@@ -129,11 +129,11 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/0cb67a0d-67f6-410b-868b-f4a0a17ff0cf/intel-oneapi-toolkit-2026.1.1.32_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      ONEAPI_INSTALLER_VERSION: "2026.1"
     steps:
       - name: Clone
         id: checkout

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/.github/workflows/release.yml

@@ -1344,11 +1344,11 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/0cb67a0d-67f6-410b-868b-f4a0a17ff0cf/intel-oneapi-toolkit-2026.1.1.32_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.28.2/level-zero-win-sdk-1.28.2.zip
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+      ONEAPI_INSTALLER_VERSION: "2026.1"
 
     steps:
       - name: Clone
@@ -1391,9 +1391,11 @@ jobs:
         run: |
           echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
 
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.6.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_def.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_avx2.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.3.dll" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
@@ -1408,13 +1410,11 @@ jobs:
             echo "Level Zero loader DLL not found in oneAPI or SDK; relying on system driver/runtime"
           fi
 
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl9.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl-ls.exe" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-fallback-bfloat16.spv" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libsycl-native-bfloat16.spv" ./build/bin
 
           cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin
@@ -1454,8 +1454,8 @@ jobs:
 
     env:
       ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2025.3.3"
-      LEVEL_ZERO_VERSION: "1.28.2"
+      ONEAPI_INSTALLER_VERSION: "2026.1"
+      LEVEL_ZERO_VERSION: "1.33.1"
       LEVEL_ZERO_UBUNTU_VERSION: "u24.04"
 
     steps:
@@ -1469,16 +1469,16 @@ jobs:
         shell: bash
         run: |
           cd /tmp
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
-          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/5996e26b-f48a-42b1-8db0-b002ad0bd8d7/intel-oneapi-toolkit-2026.1.1.33_offline.sh -O intel-oneapi-toolkit_offline.sh
+          sudo bash intel-oneapi-toolkit_offline.sh -s -a --silent --eula accept
 
       - name: Install Level Zero SDK
         shell: bash
         run: |
           cd /tmp
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb
-          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
-          sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze1_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze1.deb
+          wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze-dev_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze-dev.deb
+          sudo apt-get install -y ./libze1.deb ./libze-dev.deb
 
       - name: Download UI build
         uses: actions/download-artifact@v7

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/docs/backend/SYCL.md

@@ -52,6 +52,10 @@ The packages for FP32 and FP16 would have different accuracy and performance on
 
 ## News
 
+- 2026.09
+  - Update the CI build environment for oneAPI 2026.1 (unified oneAPI Toolkit). oneDNN is removed from the Deep Learning Essentials package in 2026.0, so the CI now uses the oneAPI Toolkit installer which still includes oneDNN.
+  - oneAPI 2026.1 improves the SYCL build performance: measured with the same code on Arc B570, prompt processing 1331 vs 434 t/s (3.1x) vs the 2025.3-based release build.
+
 - 2026.04-05
   - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q6_K, Q8_0.
   - Fused MoE.
@@ -257,7 +261,7 @@ Platform #0: Intel(R) OpenCL HD Graphics
  `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```
 
-2. **Install Intel® oneAPI Base toolkit**
+2. **Install Intel® oneAPI Toolkit**
 
 SYCL backend depends on:
   - Intel® oneAPI DPC++/C++ compiler/running-time.
@@ -267,11 +271,11 @@ SYCL backend depends on:
 
 - **For Intel GPU**
 
-All above are included in both **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** packages.
+With the 2026.0 release, the Intel® oneAPI Base toolkit and the HPC toolkit are combined into the **Intel® oneAPI Toolkit**, and **oneDNN is removed from the Intel® Deep Learning Essentials** package (oneDNN is distributed separately since then). The **Intel® oneAPI Toolkit** includes oneDNN until 2027.0.
 
-It's recommended to install **Intel® Deep Learning Essentials** which only provides the necessary libraries with less size.
+It's recommended to install the **Intel® oneAPI Toolkit**.
 
-The **Intel® oneAPI Base toolkit** and **Intel® Deep Learning Essentials** can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+The **Intel® oneAPI Toolkit** can be obtained from the official [Intel® oneAPI Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit-download.html) page.
 
 Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.
 
@@ -281,6 +285,7 @@ Upon a successful installation, SYCL is enabled for the available Intel devices,
 
 |Verified release|
 |-|
+|2026.1 |
 |2025.3.3 |
 |2025.2.1|
 |2025.1|

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -9,6 +9,7 @@
 #include "ggml-quants.h"
 #include "ggml.h"
 
+#include <algorithm>
 #include <atomic>
 #include <cerrno>
 #include <climits>
@@ -960,6 +961,24 @@ static bool has_view_op_input(const ggml_tensor * op) {
     return false;
 }
 
+// OV slices whole elements per axis, so each stride must be a multiple of the next smaller one
+// (e.g. a batch stride of m*nb[1] + pad bytes cannot be expressed and would be read wrongly).
+static bool has_strides_on_element_grid(const ggml_tensor * t) {
+    std::vector<size_t> strides;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (t->ne[i] > 1) {
+            strides.push_back(t->nb[i]);
+        }
+    }
+    std::sort(strides.begin(), strides.end());
+    for (size_t i = 1; i < strides.size(); i++) {
+        if (strides[i - 1] == 0 || strides[i] % strides[i - 1] != 0) {
+            return false;
+        }
+    }
+    return true;
+}
+
 static bool has_non_contiguous_view_input(const ggml_tensor * op) {
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         if (op->src[i] == nullptr) {
@@ -1540,6 +1559,9 @@ static ggml_openvino_op_support ggml_backend_openvino_device_supports_op_impl(gg
         if (supported_types.find(src->type) == supported_types.end()) {
             return {false, "src[" + std::to_string(i) + "] type " + std::string(ggml_type_name(src->type)) + " is not supported"};
         }
+        if (!has_strides_on_element_grid(src)) {
+            return {false, "src[" + std::to_string(i) + "] strides are not multiples of each other"};
+        }
         const bool is_supported_3d_moe_expert =
             op->op == GGML_OP_MUL_MAT_ID && i == 0 && (src->type == GGML_TYPE_MXFP4 || src->ne[3] == 1);
         if (ggml_is_quantized(src->type) && src->ne[2] != 1 && !is_supported_3d_moe_expert) {

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan-types.h

@@ -51,6 +51,8 @@ typedef struct VkPhysicalDeviceCooperativeMatrixDecodeVectorFeaturesNV {
 
 #include <cmath>
 
+#include <functional>
+
 #include <iomanip>
 
 #include <iostream>

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/scripts/snapdragon/ggml-hexagon-trace.py

@@ -514,7 +514,7 @@ def generate_perfetto_trace(filtered_ops, trace_events, output_path):
         tm = time_mappers[dev]
         e['ts_ns'] = tm.cycle_to_ns(e['start_cyc'])
         dur_ns = tm.dur_cycles_to_ns(e['start_cyc'], e['end_cyc'] - e['start_cyc'])
-        e['dur_ns'] = max(dur_ns, 100)
+        e['dur_ns'] = max(dur_ns, 1)
 
     # Allocate slots (sub-tracks) to prevent overlaps on same virtual track
     active_slots = defaultdict(list)

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/src/models/dflash.cpp

@@ -446,19 +446,16 @@ static ggml_tensor * build_dflash2_conv(
 
     ggml_tensor * weight_all = ggml_add(ctx0, coeff_all, base_side);
 
+    // taps at or past block_size only read the left padding and add nothing
+    const int64_t n_taps = std::min(kernel_size, block_size);
+
     ggml_tensor * result = nullptr;
-    for (int64_t tap = 0; tap < kernel_size; ++tap) {
+    for (int64_t tap = 0; tap < n_taps; ++tap) {
         ggml_tensor * values = blocks;
         if (tap > 0) {
-            ggml_tensor * zeros = ggml_fill(ctx0,
-                    ggml_new_tensor_3d(ctx0, hidden->type, hidden_size, std::min(tap, block_size), n_blocks), 0.0f);
-            if (tap < block_size) {
-                ggml_tensor * previous = ggml_view_3d(ctx0, blocks, hidden_size, block_size - tap, n_blocks,
-                        blocks->nb[1], blocks->nb[2], 0);
-                values = ggml_concat(ctx0, zeros, previous, 1);
-            } else {
-                values = zeros;
-            }
+            ggml_tensor * previous = ggml_view_3d(ctx0, blocks, hidden_size, block_size - tap, n_blocks,
+                    blocks->nb[1], blocks->nb[2], 0);
+            values = ggml_pad_ext(ctx0, previous, 0, 0, tap, 0, 0, 0, 0, 0);
         }
         values = ggml_reshape_2d(ctx0, values, hidden_size, n_tokens);
 

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/mtmd/models/conformer.cpp

@@ -124,8 +124,7 @@ ggml_cgraph * clip_graph_conformer::build() {
                 const auto pos_len = matrix_bd->ne[0];
                 const auto q_len   = matrix_bd->ne[1];
                 const auto h       = matrix_bd->ne[2];
-                matrix_bd          = ggml_pad(ctx0, matrix_bd, 1, 0, 0, 0);
-                matrix_bd          = ggml_roll(ctx0, matrix_bd, 1, 0, 0, 0);
+                matrix_bd          = ggml_pad_ext(ctx0, matrix_bd, 1, 0, 0, 0, 0, 0, 0, 0);
                 matrix_bd          = ggml_reshape_3d(ctx0, matrix_bd, q_len, pos_len + 1, h);
                 matrix_bd          = ggml_view_3d(ctx0, matrix_bd, q_len, pos_len, h, matrix_bd->nb[1],
                                                         matrix_bd->nb[2], matrix_bd->nb[0] * q_len);

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/mtmd/models/gemma4a.cpp

@@ -117,14 +117,12 @@ ggml_cgraph * clip_graph_gemma4a::build() {
             Qcur = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 0, 3, 1, 2)); // [D, C, B, H]
 
             // K/V block context extraction via overlapping view:
-            // Pad to S*B elements, roll right by P to create left-padding,
+            // Left pad by P and right pad to S*B elements,
             // then view with stride C in the block dimension (overlapping windows).
             auto extract_blocks = [&](ggml_tensor * t) -> ggml_tensor * {
-                // [D, H, N] -> pad to S*B -> roll right by P -> cont (materialize)
+                // [D, H, N] -> left pad by P, right pad to S*B
                 const int64_t pad_kv = S * B - n_pos;
-                t = ggml_pad(ctx0, t, 0, 0, pad_kv, 0);     // [D, H, S*B]
-                t = ggml_roll(ctx0, t, 0, 0, P, 0);          // left-pad by P
-                t = ggml_cont(ctx0, t);                       // materialize roll (removes view offset)
+                t = ggml_pad_ext(ctx0, t, 0, 0, 0, 0, P, pad_kv - P, 0, 0); // [D, H, S*B]
                 // Overlapping view: stride for B dim is C positions, not S
                 // ne = [D, H, S, B], data_size = D*H*S*B*sizeof = source_nbytes (exact fit)
                 // nb1=D*sizeof, nb2=D*H*sizeof, nb3=C*D*H*sizeof (overlap: C < S)
@@ -219,9 +217,8 @@ ggml_cgraph * clip_graph_gemma4a::build() {
                 x = ggml_cont(ctx0, ggml_transpose(ctx0, x));
             }
 
-            // Causal depthwise Conv1D via ggml_ssm_conv (pad+roll for left-only padding).
-            x = ggml_pad(ctx0, x, 4, 0, 0, 0);
-            x = ggml_roll(ctx0, x, 4, 0, 0, 0);
+            // Causal depthwise Conv1D via ggml_ssm_conv, left padded only.
+            x = ggml_pad_ext(ctx0, x, 4, 0, 0, 0, 0, 0, 0, 0);
             x = ggml_ssm_conv(ctx0, x, layer.conv_dw_w);
             if (layer.conv_dw_b) {
                 x = ggml_add(ctx0, x, layer.conv_dw_b);

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/mtmd/models/granite-speech.cpp

@@ -143,9 +143,7 @@ ggml_cgraph * clip_graph_granite_speech::build() {
             }
             cb(x, "conv_glu", il);
 
-            x = ggml_pad(ctx0, x, conv_pad, 0, 0, 0);
-            x = ggml_roll(ctx0, x, conv_pad, 0, 0, 0);
-            x = ggml_pad(ctx0, x, conv_pad, 0, 0, 0);
+            x = ggml_pad_ext(ctx0, x, conv_pad, conv_pad, 0, 0, 0, 0, 0, 0);
             x = ggml_ssm_conv(ctx0, x, layer.conv_dw_w);
             cb(x, "conv_dw", il);
 

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/mtmd/models/parakeet.cpp

@@ -287,8 +287,7 @@ ggml_cgraph * clip_graph_parakeet::build() {
                     const auto n_frame    = rel_pos_scores->ne[1];
                     const auto n_head     = rel_pos_scores->ne[2];
 
-                    rel_pos_scores = ggml_pad(ctx0, rel_pos_scores, 1, 0, 0, 0);
-                    rel_pos_scores = ggml_roll(ctx0, rel_pos_scores, 1, 0, 0, 0);
+                    rel_pos_scores = ggml_pad_ext(ctx0, rel_pos_scores, 1, 0, 0, 0, 0, 0, 0, 0);
 
                     rel_pos_scores = ggml_reshape_3d(ctx0, rel_pos_scores, n_frame, pos_window + 1, n_head);
                     rel_pos_scores = ggml_cont(ctx0, rel_pos_scores);
@@ -366,9 +365,7 @@ ggml_cgraph * clip_graph_parakeet::build() {
 
             // use ggml_ssm_conv for f32 precision
             const int dw_pad = (hparams.audio_conv_kernel_size - 1) / 2;
-            cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
-            cur = ggml_roll(ctx0, cur, dw_pad, 0, 0, 0);
-            cur = ggml_pad(ctx0, cur, dw_pad, 0, 0, 0);
+            cur = ggml_pad_ext(ctx0, cur, dw_pad, dw_pad, 0, 0, 0, 0, 0, 0);
             ggml_format_name(cur, "enc_%d_conv_dw_pad", il);
 
             cur = ggml_ssm_conv(ctx0, cur, layer.conv_dw_w);

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/mtmd/mtmd-helper-common.h

@@ -87,7 +87,7 @@ struct decode_embd_batch {
             : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd), n_tokens(n_tokens), embd(embd) {
         GGML_ASSERT(n_tokens > 0 && n_pos_per_embd > 0 && n_mmproj_embd > 0);
         pos   .resize((size_t) n_tokens * (size_t) n_pos_per_embd);
-        logits.resize(n_tokens, 0);
+        logits.resize(n_tokens);
     }
 
     void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/server/README.md

@@ -1509,6 +1509,13 @@ This endpoint requires that the model uses a pooling different than type `none`.
 
 See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-reference/embeddings).
 
+For multimodal models (loaded with `--mmproj`), each element of `input` can also be an object with a `content` array, using the same parts as `/v1/chat/completions`:
+- `{ "type": "text", "text": "..." }`: text is added to the prompt as-is
+- `{ "type": "image_url", "image_url": { "url": "..." } }`: remote URL, base64 data URI, or local file (`file://`, requires `--media-path`)
+- `{ "type": "input_audio", "input_audio": { "data": "..." } }` and `{ "type": "input_video", "input_video": { "url": "..." } }`: same as `/v1/chat/completions`, requires a model with audio or video support
+
+Each object gives one embedding. This input shape is not part of the OpenAI Embeddings API; it follows the shape used by providers like OpenRouter for vision embedding models.
+
 *Examples:*
 
 - input as string
@@ -1537,6 +1544,26 @@ See [OpenAI Embeddings API documentation](https://platform.openai.com/docs/api-r
   }'
   ```
 
+- `input` as multimodal content
+
+  ```shell
+  curl http://localhost:8080/v1/embeddings \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer no-key" \
+  -d '{
+          "input": [
+              { "content": [
+                  { "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,/9j/4AAQSkZJRg..." } },
+                  { "type": "text", "text": "Describe this image" }
+              ] },
+              { "content": [
+                  { "type": "text", "text": "hello" }
+              ] }
+          ],
+          "encoding_format": "float"
+  }'
+  ```
+
 ### POST `/v1/responses/input_tokens`: Token Counting
 
 Similar to [Response input token counts API](https://developers.openai.com/api/reference/python/resources/responses/subresources/input_tokens/methods/count).

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/server/server-common.cpp

@@ -970,7 +970,7 @@ server_tokens process_mtmd_prompt(
 }
 
 /**
- * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * tokenize a single input "prompt" object
  * use tokenize_input_prompts() if the input could be an array.
  * this supports these cases:
  * - "prompt": "string"
@@ -978,7 +978,7 @@ server_tokens process_mtmd_prompt(
  * - "prompt": [12, 34, "string", 56, 78]
  * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] }
  */
-static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special, const mtmd_helper_init_opt & init_opt) {
+server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special, const mtmd_helper_init_opt & init_opt) {
     constexpr char JSON_STRING_PROMPT_KEY[] = "prompt_string";
     constexpr char JSON_MTMD_DATA_KEY[] = "multimodal_data";
     const bool has_mtmd = mctx != nullptr;
@@ -1147,6 +1147,79 @@ static void handle_media(
     }
 }
 
+// load media files from an OAI content array, then replace each media part with a media marker text part
+static void oaicompat_content_load_media(json & content, const server_chat_params & opt, std::vector<raw_buffer> & out_files) {
+    for (auto & p : content) {
+        std::string type = json_value(p, "type", std::string());
+        if (type == "image_url") {
+            if (!opt.allow_image) {
+                throw std::runtime_error("image input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
+            }
+
+            json image_url = json_value(p, "image_url", json::object());
+            std::string url = json_value(image_url, "url", std::string());
+            handle_media(out_files, url, opt.media_path);
+
+            p["type"] = "media_marker";
+            p["text"] = get_media_marker();
+            p.erase("image_url");
+
+        } else if (type == "input_audio") {
+            if (!opt.allow_audio) {
+                throw std::runtime_error("audio input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
+            }
+
+            // note: don't need to validate "format", it's redundant
+            json input_audio = json_value(p, "input_audio", json::object());
+            std::string url  = json_value(input_audio, "data",
+                                    json_value(input_audio, "url", std::string()));
+            handle_media(out_files, url, opt.media_path);
+
+            p["type"] = "media_marker";
+            p["text"] = get_media_marker();
+            p.erase("input_audio");
+
+        } else if (type == "input_video" || type == "video_url") {
+            if (!opt.allow_video) {
+                throw std::runtime_error("video input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
+            }
+
+            // accept the OpenAI-style "video_url" key as an alias of "input_video"
+            json input_video = json_value(p, type, json::object());
+            std::string url  = json_value(input_video, "data",
+                                    json_value(input_video, "url", std::string()));
+            handle_media(out_files, url, opt.media_path);
+
+            p["type"] = "media_marker";
+            p["text"] = get_media_marker();
+            p.erase("input_video");
+            p.erase("video_url");
+
+        } else if (type != "text") {
+            throw std::invalid_argument("unsupported content[].type");
+        }
+    }
+}
+
+server_tokens tokenize_oai_content_array(const llama_vocab * vocab, mtmd_context * mctx, const server_chat_params & opt, json content, bool add_special, bool parse_special, const mtmd_helper_init_opt & init_opt) {
+    if (!content.is_array()) {
+        throw std::invalid_argument("\"content\" must be an array");
+    }
+
+    std::vector<raw_buffer> files;
+    oaicompat_content_load_media(content, opt, files);
+
+    std::string prompt;
+    for (const auto & p : content) {
+        prompt += json_value(p, "text", std::string());
+    }
+
+    if (files.empty()) {
+        return server_tokens(common_tokenize(vocab, prompt, add_special, parse_special), false);
+    }
+    return process_mtmd_prompt(mctx, prompt, files, init_opt);
+}
+
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
@@ -1233,56 +1306,7 @@ json oaicompat_chat_params_parse(
             throw std::invalid_argument("Expected 'content' to be a string or an array");
         }
 
-        for (auto & p : content) {
-            std::string type = json_value(p, "type", std::string());
-            if (type == "image_url") {
-                if (!opt.allow_image) {
-                    throw std::runtime_error("image input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
-                }
-
-                json image_url = json_value(p, "image_url", json::object());
-                std::string url = json_value(image_url, "url", std::string());
-                handle_media(out_files, url, opt.media_path);
-
-                p["type"] = "media_marker";
-                p["text"] = get_media_marker();
-                p.erase("image_url");
-
-            } else if (type == "input_audio") {
-                if (!opt.allow_audio) {
-                    throw std::runtime_error("audio input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
-                }
-
-                // note: don't need to validate "format", it's redundant
-                json input_audio = json_value(p, "input_audio", json::object());
-                std::string url  = json_value(input_audio, "data",
-                                        json_value(input_audio, "url", std::string()));
-                handle_media(out_files, url, opt.media_path);
-
-                p["type"] = "media_marker";
-                p["text"] = get_media_marker();
-                p.erase("input_audio");
-
-            } else if (type == "input_video" || type == "video_url") {
-                if (!opt.allow_video) {
-                    throw std::runtime_error("video input is not supported - hint: if this is unexpected, you may need to provide the mmproj");
-                }
-
-                // accept the OpenAI-style "video_url" key as an alias of "input_video"
-                json input_video = json_value(p, type, json::object());
-                std::string url  = json_value(input_video, "data",
-                                        json_value(input_video, "url", std::string()));
-                handle_media(out_files, url, opt.media_path);
-
-                p["type"] = "media_marker";
-                p["text"] = get_media_marker();
-                p.erase("input_video");
-                p.erase("video_url");
-
-            } else if (type != "text") {
-                throw std::invalid_argument("unsupported content[].type");
-            }
-        }
+        oaicompat_content_load_media(content, opt, out_files);
     }
 
     auto caps = common_chat_templates_get_caps(opt.tmpls.get());

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/server/server-common.h

@@ -300,6 +300,15 @@ std::vector<server_tokens> tokenize_input_prompts(
                                         bool parse_special,
                                         const mtmd_helper_init_opt & init_opt);
 
+// tokenize a single prompt, see tokenize_input_prompts() for the supported shapes
+server_tokens tokenize_input_subprompt(
+                                        const llama_vocab * vocab,
+                                        mtmd_context * mctx,
+                                        const json & json_prompt,
+                                        bool add_special,
+                                        bool parse_special,
+                                        const mtmd_helper_init_opt & init_opt);
+
 //
 // OAI utils
 //
@@ -330,6 +339,16 @@ json oaicompat_chat_params_parse(
     const server_chat_params & opt,
     std::vector<raw_buffer> & out_files);
 
+// used by /embeddings endpoint, content has the same format as a chat message content array
+server_tokens tokenize_oai_content_array(
+    const llama_vocab * vocab,
+    mtmd_context * mctx,
+    const server_chat_params & opt,
+    json content,
+    bool add_special,
+    bool parse_special,
+    const mtmd_helper_init_opt & init_opt);
+
 // TODO: move it to server-task.cpp
 json format_embeddings_response_oaicompat(
     const json & request,

{llama_cpp_bin-11236.0.0 → llama_cpp_bin-11243.0.0}/llama.cpp/tools/server/server-context.cpp

@@ -3193,7 +3193,9 @@ private:
                                 return;
                             }
 
-                            if (slot.task->params.cache_prompt) {
+                            const bool is_stateless_task = slot.task->type == SERVER_TASK_TYPE_EMBEDDING || slot.task->type == SERVER_TASK_TYPE_RERANK;
+
+                            if (slot.task->params.cache_prompt && !is_stateless_task) {
                                 // reuse any previously computed tokens that are common with the new prompt
                                 n_past = slot.prompt.tokens.get_common_prefix(input_tokens);
 
@@ -5403,7 +5405,27 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
         }
     }
 
-    auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true, ctx_server.init_opt);
+    // same shapes as tokenize_input_prompts(), plus OAI content: { "content": [ { "type": "text"|"image_url"|"input_audio"|"input_video", ... } ] }
+    auto tokenize_entry = [&](const json & p) {
+        if (p.is_object() && p.contains("content")) {
+            return tokenize_oai_content_array(ctx_server.vocab, ctx_server.mctx, meta->chat_params, p.at("content"), true, true, ctx_server.init_opt);
+        }
+        return tokenize_input_subprompt(ctx_server.vocab, ctx_server.mctx, p, true, true, ctx_server.init_opt);
+    };
+
+    std::vector<server_tokens> tokenized_prompts;
+    if (prompt.is_array() && !json_is_array_and_contains_numbers(prompt)) {
+        for (const auto & p : prompt) {
+            tokenized_prompts.push_back(tokenize_entry(p));
+        }
+    } else {
+        tokenized_prompts.push_back(tokenize_entry(prompt));
+    }
+    if (tokenized_prompts.empty()) {
+        res->error(format_error_response("\"input\" must not be empty", ERROR_TYPE_INVALID_REQUEST));
+        return res;
+    }
+
     for (const auto & tokens : tokenized_prompts) {
         // this check is necessary for models that do not add BOS token to the input
         if (tokens.empty()) {