add geglu activation function (#14074)

Co-authored-by: dinhhuy <huy.dinh@brains-tech.co.jp>

.github/labeler.yml

@@ -86,3 +86,10 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
+
+Ascend NPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-cann.h
+            - ggml/src/ggml-cann/**
+            - docs/backend/CANN.md

.github/workflows/build-linux-cross.yml

@@ -231,3 +231,116 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-cpu-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-vulkan-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu \
+                  libvulkan-dev:loong64
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -839,12 +839,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-2022-cmake-cuda:
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -878,7 +878,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/release.yml

@@ -131,8 +131,9 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
+          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
+          # - build: 'arm64'
+          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -159,6 +160,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -207,6 +211,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -373,11 +380,11 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -405,7 +412,7 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/server.yml

@@ -180,7 +180,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -159,6 +159,11 @@ if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 #
 # build the library
 #

README.md

@@ -3,6 +3,7 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
@@ -28,6 +29,30 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
+## Quick start
+
+Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
+
+- Install `llama.cpp` using [brew, nix or winget](docs/install.md)
+- Run with Docker - see our [Docker documentation](docs/docker.md)
+- Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
+- Build from source by cloning this repository - check out [our build guide](docs/build.md)
+
+Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
+
+Example command:
+
+```sh
+# Use a local model file
+llama-cli -m my_model.gguf
+
+# Or download and run a model directly from Hugging Face
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+
+# Launch OpenAI-compatible API server
+llama-server -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -230,6 +255,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
+
 ## Supported backends
 
 | Backend | Target devices |
@@ -246,16 +272,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
-## Building the project
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
-
-- Clone this repository and build locally, see [how to build](docs/build.md)
-- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
-- Use a Docker image, see [documentation for Docker](docs/docker.md)
-- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
-
 ## Obtaining and quantizing models
 
 The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
@@ -263,7 +279,11 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+
+```sh
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
 
 By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
 

ci/run.sh

@@ -46,7 +46,20 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
+
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
+        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
+        else
+            echo "Warning: Using fallback CUDA architectures"
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
+        fi
+    else
+        echo "Error: nvidia-smi not found, cannot build with CUDA"
+        exit 1
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

common/arg.cpp

@@ -2869,6 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
             else { throw std::invalid_argument("invalid value"); }
         }

common/chat.cpp

@@ -82,10 +82,10 @@ json common_chat_msg::to_json_oaicompat() const
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg) {
     std::vector<common_chat_msg_diff> diffs;
-    // if (previous_msg.reasoning_content != current.reasoning_content) {
-    //     auto & diff = diffs.emplace_back();
-    //     diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, current.reasoning_content);
-    // }
+    if (previous_msg.reasoning_content != new_msg.reasoning_content) {
+        auto & diff = diffs.emplace_back();
+        diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, new_msg.reasoning_content);
+    }
     if (previous_msg.content != new_msg.content) {
         auto & diff = diffs.emplace_back();
         diff.content_delta = string_diff(previous_msg.content, new_msg.content);
@@ -385,9 +385,9 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
 
 template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
-    // if (!diff.reasoning_content_delta.empty()) {
-    //     delta["reasoning_content"] = msg.reasoning_content;
-    // }
+    if (!diff.reasoning_content_delta.empty()) {
+        delta["reasoning_content"] = diff.reasoning_content_delta;
+    }
     if (!diff.content_delta.empty()) {
         delta["content"] = diff.content_delta;
     }
@@ -598,6 +598,7 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
+        case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }

common/chat.h

@@ -70,7 +70,7 @@ struct common_chat_msg {
 };
 
 struct common_chat_msg_diff {
-    // std::string reasoning_content_delta;
+    std::string reasoning_content_delta;
     std::string content_delta;
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;

common/common.cpp

@@ -934,7 +934,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -1041,7 +1041,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_self_clear(lctx);
+        llama_memory_clear(llama_get_memory(lctx), true);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);

common/common.h

@@ -215,7 +215,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {

common/speculative.cpp

@@ -144,6 +144,8 @@ llama_tokens common_speculative_gen_draft(
     auto & smpl   = spec->smpl;
     auto & prompt = spec->prompt;
 
+    auto * mem = llama_get_memory(ctx);
+
     int reuse_i = 0;
     int reuse_n = 0;
 
@@ -173,7 +175,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(mem, false);
 
         prompt.clear();
     } else {
@@ -192,14 +194,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem, 0, 0, reuse_i);
+            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
+            llama_memory_seq_rm (mem, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -3709,8 +3709,7 @@ class BertModel(TextModel):
         self._try_set_pooling_type()
 
         if self.cls_out_labels:
-            key_name = gguf.Keys.Classifier.OUTPUT_LABELS.format(arch = gguf.MODEL_ARCH_NAMES[self.model_arch])
-            self.gguf_writer.add_array(key_name, [v for k, v in sorted(self.cls_out_labels.items())])
+            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3814,7 +3813,7 @@ class BertModel(TextModel):
             remove_whitespaces = tokenizer.clean_up_tokenization_spaces
             precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])
 
-            vocab_size = self.hparams.get("vocab_size", tokenizer.vocab_size)
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)
         else:
             sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
             sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
@@ -3827,7 +3826,7 @@ class BertModel(TextModel):
             tokenizer = SentencePieceProcessor()
             tokenizer.LoadFromFile(str(tokenizer_path))
 
-            vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size())
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
@@ -3857,33 +3856,26 @@ class BertModel(TextModel):
             unk_token = tokenizer_config_json.get("unk_token")
             unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))
 
-            for token_id in range(vocab_size):
+            for token_id in range(tokenizer.vocab_size):
                 piece = tokenizer._convert_id_to_token(token_id)
-                text = piece.encode("utf-8")
-                score = tokenizer_json["model"]["vocab"][token_id][1]
+                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:
+                    text = piece.encode("utf-8")
+                    score = tokenizer_json["model"]["vocab"][token_id][1]
 
-                toktype = SentencePieceTokenTypes.NORMAL
-                if token_id == unk_token_id:
-                    toktype = SentencePieceTokenTypes.UNKNOWN
-                elif token_id in tokenizer.all_special_ids:
-                    toktype = SentencePieceTokenTypes.CONTROL
-                elif token_id in added_vocab.values():
-                    toktype = SentencePieceTokenTypes.USER_DEFINED
-                # No reliable way to detect this, but jina doesn't have any
-                # elif tokenizer.IsByte(token_id):
-                #     toktype = SentencePieceTokenTypes.BYTE
+                    toktype = SentencePieceTokenTypes.NORMAL
+                    if token_id == unk_token_id:
+                        toktype = SentencePieceTokenTypes.UNKNOWN
+                    elif token_id in tokenizer.all_special_ids:
+                        toktype = SentencePieceTokenTypes.CONTROL
+                    elif token_id in added_vocab.values():
+                        toktype = SentencePieceTokenTypes.USER_DEFINED
+                    # No reliable way to detect this, but jina doesn't have any
+                    # elif tokenizer.IsByte(token_id):
+                    #     toktype = SentencePieceTokenTypes.BYTE
 
-                tokens[token_id] = text
-                scores[token_id] = score
-                toktypes[token_id] = toktype
-
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
-            for i in range(1, pad_count + 1):
-                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
+                    tokens[token_id] = text
+                    scores[token_id] = score
+                    toktypes[token_id] = toktype
 
         if isinstance(tokenizer, SentencePieceProcessor):
             # realign tokens (see HF tokenizer code)
@@ -3896,6 +3888,12 @@ class BertModel(TextModel):
                 SentencePieceTokenTypes.UNKNOWN,
             ] + toktypes[3:-1]
 
+            if self.model_arch == gguf.MODEL_ARCH.NOMIC_BERT_MOE:
+                # Add mask token missing from sentencepiece.bpe.model
+                tokens[250001] = b'<mask>'
+                scores[250001] = 0.0
+                toktypes[250001] = SentencePieceTokenTypes.CONTROL
+
         self.gguf_writer.add_tokenizer_model("t5")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)

docs/build.md

@@ -1,5 +1,9 @@
 # Build llama.cpp locally
 
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
 **To get the Code:**
 
 ```bash

docs/install.md

@@ -1,28 +1,42 @@
 # Install pre-built version of llama.cpp
 
-## Homebrew
+| Install via | Windows | Mac | Linux |
+|-------------|---------|-----|-------|
+| Winget      | ✅      |      |      |
+| Homebrew    |         | ✅   | ✅   |
+| MacPorts    |         | ✅   |      |
+| Nix         |         | ✅   | ✅   |
 
-On Mac and Linux, the homebrew package manager can be used via
+## Winget (Windows)
+
+```sh
+winget install llama.cpp
+```
+
+The package is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/issues/8188
+
+## Homebrew (Mac and Linux)
 
 ```sh
 brew install llama.cpp
 ```
+
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
-## MacPorts
+## MacPorts (Mac)
 
 ```sh
 sudo port install llama.cpp
 ```
-see also: https://ports.macports.org/port/llama.cpp/details/
 
-## Nix
+See also: https://ports.macports.org/port/llama.cpp/details/
 
-On Mac and Linux, the Nix package manager can be used via
+## Nix (Mac and Linux)
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
+
 For flake enabled installs.
 
 Or
@@ -34,13 +48,3 @@ nix-env --file '<nixpkgs>' --install --attr llama-cpp
 For non-flake enabled installs.
 
 This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-## Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-
-```sh
-flox install llama-cpp
-```
-
-Flox follows the nixpkgs build of llama.cpp.

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/embedding/embedding.cpp

@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -236,9 +236,24 @@ int main(int argc, char ** argv) {
                 LOG("\n");
             }
         } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            const uint32_t n_cls_out = llama_model_n_cls_out(model);
+            std::vector<std::string> cls_out_labels;
+
+            for (uint32_t i = 0; i < n_cls_out; i++) {
+                const char * label = llama_model_cls_label(model, i);
+                const std::string label_i(label == nullptr ? "" : label);
+                cls_out_labels.emplace_back(label_i.empty() ? std::to_string(i) : label_i);
+            }
+
             for (int j = 0; j < n_embd_count; j++) {
-                // NOTE: if you change this log - update the tests in ci/run.sh
-                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                for (uint32_t i = 0; i < n_cls_out; i++) {
+                    // NOTE: if you change this log - update the tests in ci/run.sh
+                    if (n_cls_out == 1) {
+                        LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                    } else {
+                        LOG("rerank score %d: %8.3f [%s]\n", j, emb[j * n_embd + i], cls_out_labels[i].c_str());
+                    }
+                }
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding

examples/gritlm/gritlm.cpp

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
+    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
 }

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_self_clear(context)
+        llama_memory_clear(llama_get_memory(context), true)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/lookahead/lookahead.cpp

@@ -60,6 +60,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // Tokenize the prompt
@@ -94,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+        llama_memory_seq_cp(mem, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -427,17 +429,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
+        llama_memory_seq_rm(mem, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_self_seq_keep(ctx, seq_id_best);
-            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_memory_seq_keep(mem, seq_id_best);
+            llama_memory_seq_cp  (mem, seq_id_best, 0, -1, -1);
+            llama_memory_seq_rm  (mem, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+                llama_memory_seq_cp(mem, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -181,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
+        llama_memory_seq_rm(llama_get_memory(ctx), 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/parallel/parallel.cpp

@@ -194,6 +194,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
@@ -259,7 +261,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+            llama_memory_seq_cp(mem, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -286,9 +288,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_self_seq_rm(ctx, i, -1, -1);
+                llama_memory_seq_rm(mem, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                llama_memory_seq_cp(mem, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -447,8 +449,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
+                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,6 +126,8 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
+    auto * mem = llama_get_memory(ctx);
+
     // fill the KV cache
     for (int i = 0; i < n_ctx; i += n_batch) {
         if (i > 0 && n_grp > 1) {
@@ -133,10 +135,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_self_seq_add(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_self_seq_div(ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_memory_seq_add(mem, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_memory_seq_div(mem, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -166,10 +168,10 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+        llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+        llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -195,10 +197,10 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+            llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+            llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
     }
 

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_process(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), false);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/save-load-state/save-load-state.cpp

@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_self_clear(ctx3);
+        llama_memory_clear(llama_get_memory(ctx3), true);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");

examples/speculative-simple/speculative-simple.cpp

@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -142,6 +142,8 @@ int main(int argc, char ** argv) {
         }
     }
 
+    auto * mem_tgt = llama_get_memory(ctx_tgt);
+    auto * mem_dft = llama_get_memory(ctx_dft);
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -420,14 +422,14 @@ int main(int argc, char ** argv) {
             {
                 LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
 
-                llama_kv_self_seq_keep(ctx_dft, s_keep);
-                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_dft, 0);
+                llama_memory_seq_keep(mem_dft, s_keep);
+                llama_memory_seq_cp  (mem_dft, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_dft, 0);
 
-                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_self_seq_keep(ctx_tgt, s_keep);
-                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_tgt, 0);
+                llama_memory_seq_rm  (mem_tgt, s_keep, n_past_tgt, -1);
+                llama_memory_seq_keep(mem_tgt, s_keep);
+                llama_memory_seq_cp  (mem_tgt, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +446,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_memory_seq_rm(mem_dft, 0, n_past_dft, -1);
             // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode(ctx_dft, batch_dft);
 
@@ -503,8 +505,8 @@ int main(int argc, char ** argv) {
                     if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
-                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
+                        llama_memory_seq_cp(mem_dft, s, n_seq_cur, -1, -1);
 
                         // all previous tokens from this branch are now also part of the new branch
                         for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -585,9 +587,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_self_seq_keep(ctx_tgt, 0);
+            llama_memory_seq_keep(mem_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_memory_seq_cp(mem_tgt, 0, s, -1, -1);
             }
 
             // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());

ggml/CMakeLists.txt

@@ -137,7 +137,7 @@ set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
 set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
-if (WIN32)
+if (MINGW)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 

ggml/src/CMakeLists.txt

@@ -125,7 +125,6 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -318,7 +318,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
             endif()
 
-            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)

ggml/src/ggml-cpu/ops.cpp

@@ -8132,8 +8132,8 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
         #define WKV_VECTOR_SIZE 4
     #endif
 
-    int wkv_vector_size;
     #ifdef WKV_VECTOR_SIZE
+        int wkv_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             wkv_vector_size = svcntw();
         #else
@@ -8348,8 +8348,8 @@ static void ggml_compute_forward_gla_f32(
         #define GLA_VECTOR_SIZE 4
     #endif
 
-    int gla_vector_size;
     #ifdef GLA_VECTOR_SIZE
+        int gla_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             gla_vector_size = svcntw();
         #else

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
             KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
         }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1144,7 +1144,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
     const char * src_ptr = (const char *) src->data;
     char       * dst_ptr = (char       *) dst;
 
@@ -1427,8 +1426,6 @@ static void ggml_cuda_op_mul_mat(
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
     ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
     ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
 
@@ -1750,7 +1747,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.

ggml/src/ggml-metal/ggml-metal.m

@@ -4766,6 +4766,8 @@ static bool ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    const int is_q = ggml_is_quantized(src1->type) ? 1 : 0;
+
                     // 2*(2*ncpsg + nqptg)*(nsg)
                     // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
                     //
@@ -4773,7 +4775,7 @@ static bool ggml_metal_encode_node(
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(2*ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -4810,9 +4812,9 @@ static bool ggml_metal_encode_node(
                     // and store the soft_max values and the mask
                     //
                     // ne00*(nsg)
-                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
+                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {

ggml/src/ggml-metal/ggml-metal.metal

@@ -3328,14 +3328,14 @@ kernel void kernel_flash_attn_ext(
     constexpr short NW  = N_SIMDWIDTH;
     constexpr short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
 
-    const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
-    const short T  = DK + 2*TS; // shared memory size per query in (half)
+    const short TS = nsg*SH;      // shared memory size per query in (s_t == float)
+    const short T  = 2*DK + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*DK); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*DK); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*DK); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*DK); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +                0*DK); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +                0*DK); // same as above but in q4_t
+    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +                0*DK); // reuse query data for accumulation
+    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +                0*DK); // same as above but in o4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + 2*Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
     threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
@@ -3354,7 +3354,7 @@ kernel void kernel_flash_attn_ext(
             if (iq1 + j < args.ne01) {
                 sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*DK4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = 0;
             }
         }
     }
@@ -3634,9 +3634,6 @@ kernel void kernel_flash_attn_ext(
 
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        float S = { 0.0f };
-        float M = { -__FLT_MAX__/2 };
-
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // each simdgroup stores its output to shared memory, reusing sq
@@ -3657,12 +3654,12 @@ kernel void kernel_flash_attn_ext(
                 const float M0 = ss[j*TS +         1];
                 const float M1 = ss[j*TS + sg*SH + 1];
 
-                M = max(M0, M1);
+                const float M = max(M0, M1);
 
                 const float ms0 = exp(M0 - M);
                 const float ms1 = exp(M1 - M);
 
-                S = S0*ms0 + S1*ms1;
+                const float S = S0*ms0 + S1*ms1;
 
                 if (tiisg == 0) {
                     ss[j*TS + 0] = S;
@@ -3701,16 +3698,18 @@ kernel void kernel_flash_attn_ext(
         }
     }
 
-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
-    if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
-            const float S = ss[j*TS + 0];
+    for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
+        const float S = 1.0f/sf[j*TS + 0];
 
-            for (short i = tiisg; i < DV4; i += NW) {
-                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S;
-            }
+        device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
+
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[i] = (float4) so4[j*DV4 + i]*S;
         }
     }
 }
@@ -3719,12 +3718,22 @@ kernel void kernel_flash_attn_ext(
 //       template to be able to explore different combinations
 //
 #define FA_TYPES \
-    half,  half4,   simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    float,          simdgroup_float8x8, \
-    float,          simdgroup_float8x8, \
-    half,  half4,   simdgroup_half8x8
+    float,  float4,    simdgroup_float8x8, \
+    half,   half4x4,   simdgroup_half8x8,  \
+    half,   half4x4,   simdgroup_half8x8,  \
+    float,             simdgroup_float8x8, \
+    float,             simdgroup_float8x8, \
+    float,  float4,    simdgroup_float8x8
+    //half,   half4,     simdgroup_half8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    float,  float4,    simdgroup_float8x8
+    //half,   half4,     simdgroup_half8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
@@ -3739,15 +3748,15 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3801,6 +3810,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
+#undef FA_TYPES_BF
 
 template<
     typename q4_t,  // query types in shared memory
@@ -3847,12 +3857,12 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                  0*DK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                  0*DK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 + sgitg*SH       + Q*DK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 + sgitg*SH       + Q*DK); // same as above but in s4_t
-    threadgroup float * sm  = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
-    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + sgitg*DV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*DK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*DK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*DK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*DK); // same as above but in s4_t
+    threadgroup float * sm  = (threadgroup float *) (shmem_f16 +   sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
+    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*DV       + Q*T);  // scratch buffer for the results
 
     // store the result for all queries in local memory (the O matrix from the paper)
     o4_t lo[DV4/NL];
@@ -4157,7 +4167,7 @@ kernel void kernel_flash_attn_ext_vec(
            half4,  \
     float,         \
     float, float4, \
-           half4
+           float4
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -95,6 +95,12 @@ set(GGML_OPENCL_KERNELS
     sub
     sum_rows
     transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    pad
+    repeat
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -315,6 +315,12 @@ struct ggml_backend_opencl_context {
     cl_program program_softmax_4_f16;
     cl_program program_argsort_f32_i32;
     cl_program program_sum_rows_f32;
+    cl_program program_repeat;
+    cl_program program_pad;
+    cl_program program_tanh;
+    cl_program program_upscale;
+    cl_program program_concat;
+    cl_program program_tsembd;
 
     cl_kernel kernel_add, kernel_add_row;
     cl_kernel kernel_mul, kernel_mul_row;
@@ -351,6 +357,15 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
+    cl_kernel kernel_repeat;
+    cl_kernel kernel_pad;
+    cl_kernel kernel_tanh_f32_nd;
+    cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_upscale;
+    cl_kernel kernel_upscale_bilinear;
+    cl_kernel kernel_concat_f32_contiguous;
+    cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_timestep_embedding;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // Transpose kernels
@@ -1097,6 +1112,150 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+        // repeat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "repeat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("repeat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_repeat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_repeat = clCreateKernel(backend_ctx->program_repeat, "kernel_repeat", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: repeat kernel source not found or empty. Repeat operations will not be available.\n");
+            backend_ctx->program_repeat = nullptr;
+            backend_ctx->kernel_repeat = nullptr;
+        }
+    }
+
+    // pad
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "pad.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("pad.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_pad =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_pad = clCreateKernel(backend_ctx->program_pad, "kernel_pad", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: pad kernel source not found or empty. Pad operations will not be available.\n");
+            backend_ctx->program_pad = nullptr;
+            backend_ctx->kernel_pad = nullptr;
+        }
+    }
+
+    // tanh
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tanh.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tanh.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tanh =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_tanh_f32_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_tanh_f16_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: tanh kernel source not found or empty. Tanh operation will not be available.\n");
+            backend_ctx->program_tanh = nullptr;
+            backend_ctx->kernel_tanh_f32_nd = nullptr;
+            backend_ctx->kernel_tanh_f16_nd = nullptr;
+        }
+    }
+
+    // upscale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "upscale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("upscale.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_upscale =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_upscale = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale", &err), err));
+            if (backend_ctx->program_upscale) {
+                 cl_int err_bilinear;
+                 backend_ctx->kernel_upscale_bilinear = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale_bilinear", &err_bilinear);
+                 if (err_bilinear != CL_SUCCESS) {
+                    GGML_LOG_WARN("ggml_opencl: kernel_upscale_bilinear not found in upscale.cl. Bilinear upscale will not be available. Error: %d\n", err_bilinear);
+                    backend_ctx->kernel_upscale_bilinear = nullptr;
+                 }
+            } else {
+                backend_ctx->kernel_upscale_bilinear = nullptr;
+            }
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: upscale kernel source not found or empty. Upscale operations will not be available.\n");
+            backend_ctx->program_upscale = nullptr;
+            backend_ctx->kernel_upscale = nullptr;
+            backend_ctx->kernel_upscale_bilinear = nullptr;
+        }
+    }
+
+    // concat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "concat.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("concat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_concat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+            CL_CHECK((backend_ctx->kernel_concat_f32_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_contiguous", &err), err));
+            CL_CHECK((backend_ctx->kernel_concat_f32_non_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_non_contiguous", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: concat kernel source not found or empty. Concat operations will not be available.\n");
+            backend_ctx->program_concat = nullptr;
+            backend_ctx->kernel_concat_f32_contiguous = nullptr;
+            backend_ctx->kernel_concat_f32_non_contiguous = nullptr;
+        }
+    }
+
+    // timestep_embedding
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tsembd.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("tsembd.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tsembd =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_timestep_embedding = clCreateKernel(backend_ctx->program_tsembd, "kernel_timestep_embedding", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: timestep_embedding kernel source not found or empty. This op will not be available.\n");
+            backend_ctx->program_tsembd = nullptr;
+            backend_ctx->kernel_timestep_embedding = nullptr;
+        }
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1863,7 +2022,12 @@ static bool ggml_backend_opencl_cpy_tensor_async(ggml_backend_t backend, const g
 }
 
 static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+
+    cl_event evt;
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, 0, nullptr, &evt));
+    CL_CHECK(clWaitForEvents(1, &evt));
+    CL_CHECK(clReleaseEvent(evt));
 }
 
 // Syncronizes the 'backend_ctx's device with others so that commands
@@ -1976,9 +2140,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU_QUICK:
-                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 case GGML_UNARY_OP_SIGMOID:
                     return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_TANH:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
                 default:
                     return false;
             }
@@ -1988,6 +2155,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
             return true;
+        case GGML_OP_REPEAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
+        case GGML_OP_PAD:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   op->src[0]->ne[3] == 1 && op->ne[3] == 1;
+        case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONCAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_MUL_MAT:
@@ -2052,7 +2230,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
     /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
     /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
-    /* .synchronize             = */ NULL,  /* ggml_backend_opencl_synchronize */
+    /* .synchronize             = */ ggml_backend_opencl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_update       = */ NULL,
@@ -4108,6 +4286,536 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0,
 #endif
 }
 
+static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_tanh_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_tanh_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_tanh");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0]; const int ne01 = src0->ne[1]; const int ne02 = src0->ne[2]; const int ne03 = src0->ne[3];
+    const cl_ulong nb00 = src0->nb[0]; const cl_ulong nb01 = src0->nb[1]; const cl_ulong nb02 = src0->nb[2]; const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0]; const int ne11 = dst->ne[1]; const int ne12 = dst->ne[2]; const int ne13 = dst->ne[3];
+    const cl_ulong nb10 = dst->nb[0]; const cl_ulong nb11 = dst->nb[1]; const cl_ulong nb12 = dst->nb[2]; const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == src0->type);
+
+    UNUSED(src1_shape_def);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_repeat == nullptr) {
+        GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int src0_ne0 = src0->ne[0]; const int src0_ne1 = src0->ne[1]; const int src0_ne2 = src0->ne[2]; const int src0_ne3 = src0->ne[3];
+    const cl_ulong src0_nb0 = src0->nb[0]; const cl_ulong src0_nb1 = src0->nb[1]; const cl_ulong src0_nb2 = src0->nb[2]; const cl_ulong src0_nb3 = src0->nb[3];
+
+    const int dst_ne0 = dst->ne[0]; const int dst_ne1 = dst->ne[1]; const int dst_ne2 = dst->ne[2]; const int dst_ne3 = dst->ne[3];
+    const cl_ulong dst_nb0 = dst->nb[0]; const cl_ulong dst_nb1 = dst->nb[1]; const cl_ulong dst_nb2 = dst->nb[2]; const cl_ulong dst_nb3 = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_repeat;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &src0_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &src0_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &src0_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &src0_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong),  &src0_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong),  &src0_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &src0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &src0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &dst_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &dst_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &dst_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dst_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &dst_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &dst_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &dst_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &dst_nb3));
+
+    size_t gws0 = dst_ne1 > 0 ? (size_t)dst_ne1 : 1;
+    size_t gws1 = dst_ne2 > 0 ? (size_t)dst_ne2 : 1;
+    size_t gws2 = dst_ne3 > 0 ? (size_t)dst_ne3 : 1;
+
+    size_t global_work_size[] = { gws0, gws1, gws2 };
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_pad == nullptr) {
+        GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int s_ne0 = src0->ne[0];
+    const int s_ne1 = src0->ne[1];
+    const int s_ne2 = src0->ne[2];
+
+    const int d_ne0 = dst->ne[0];
+    const int d_ne1 = dst->ne[1];
+    const int d_ne2 = dst->ne[2];
+
+    cl_kernel kernel = backend_ctx->kernel_pad;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &s_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &s_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &s_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne2));
+
+    size_t lws0 = 64;
+    size_t gws0 = (( (size_t)d_ne0 + lws0 - 1 ) / lws0) * lws0;
+
+    size_t global_work_size[] = { gws0, (size_t)d_ne1, (size_t)d_ne2 };
+    size_t local_work_size[]  = { lws0, 1, 1 };
+
+    size_t * local_work_size_ptr = local_work_size;
+     if (d_ne0 % lws0 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    cl_kernel kernel = nullptr;
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        kernel = backend_ctx->kernel_upscale;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: nearest upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        kernel = backend_ctx->kernel_upscale_bilinear;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: bilinear upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else {
+        GGML_LOG_WARN("%s: unsupported upscale mode %d, skipping OpenCL execution.\n", __func__, mode);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne00_src = src0->ne[0];
+    const int ne01_src = src0->ne[1];
+
+    const int ne10_dst = dst->ne[0];
+    const int ne11_dst = dst->ne[1];
+    const int ne12_dst = dst->ne[2];
+    const int ne13_dst = dst->ne[3];
+
+    const float sf0 = (float)dst->ne[0] / src0->ne[0];
+    const float sf1 = (float)dst->ne[1] / src0->ne[1];
+    const float sf2 = (float)dst->ne[2] / src0->ne[2];
+    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00_src));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01_src));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+    }
+
+
+    size_t dst_total_elements = (size_t)ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    if (dst_total_elements == 0) {
+        return;
+    }
+    size_t global_work_size[] = { dst_total_elements, 1, 1 };
+    size_t local_work_size_pref = 256;
+    size_t local_work_size[] = { MIN(local_work_size_pref, dst_total_elements), 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (dst_total_elements % local_work_size[0] != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], 1, 1};
+    size_t profiling_lws[3] = {local_work_size_ptr ? local_work_size[0] : 0, 1, 1};
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_concat_f32_contiguous == nullptr || backend_ctx->kernel_concat_f32_non_contiguous == nullptr) {
+        GGML_LOG_WARN("%s: concat kernels not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra0_cl = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1_cl = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad_cl = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra0_cl->offset + src0->view_offs;
+    cl_ulong off_src1 = extra1_cl->offset + src1->view_offs;
+    cl_ulong off_dst  = extrad_cl->offset + dst->view_offs;
+
+    const int32_t dim = ((const int32_t *) dst->op_params)[0];
+    GGML_ASSERT(dim >= 0 && dim <= 3);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+        if (dim == 3) {
+
+            size_t nbytes_src0 = ggml_nbytes(src0);
+            size_t nbytes_src1 = ggml_nbytes(src1);
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra0_cl->data_device, extrad_cl->data_device,
+                                         off_src0, off_dst, nbytes_src0, 0, NULL, NULL));
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra1_cl->data_device, extrad_cl->data_device,
+                                         off_src1, off_dst + nbytes_src0, nbytes_src1, 0, NULL, NULL));
+        } else {
+
+            cl_kernel kernel = backend_ctx->kernel_concat_f32_contiguous;
+            size_t global_work_size[3];
+
+            for (int i3 = 0; i3 < dst->ne[3]; ++i3) {
+                cl_ulong current_off_src0 = off_src0 + (i3 * src0->nb[3]);
+                cl_ulong current_off_src1 = off_src1 + (i3 * src1->nb[3]);
+                cl_ulong current_off_dst  = off_dst  + (i3 * dst->nb[3]);
+
+                int d_ne00 = src0->ne[0]; int d_ne01 = src0->ne[1]; int d_ne02 = src0->ne[2];
+                int d_ne10 = src1->ne[0]; int d_ne11 = src1->ne[1]; int d_ne12 = src1->ne[2];
+                int d_ne0  = dst->ne[0];  int d_ne1  = dst->ne[1];  int d_ne2  = dst->ne[2];
+
+                CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &current_off_src0));
+                CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &current_off_src1));
+                CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &current_off_dst));
+                CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &d_ne00));
+                CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne01));
+                CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne02));
+                CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne10));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &d_ne11));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &d_ne12));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &d_ne0));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &d_ne1));
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &d_ne2));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dim));
+
+                global_work_size[0] = d_ne0;
+                global_work_size[1] = d_ne1;
+                global_work_size[2] = d_ne2;
+
+                CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+            }
+        }
+    } else {
+        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
+
+        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+
+        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+
+        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
+
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_src1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
+
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),    &nb03));
+
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),    &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),    &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
+
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
+        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong),    &d_nb3));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &dim));
+
+        size_t global_work_size_nc[] = { d_ne1 > 0 ? (size_t)d_ne1 : 1,
+                                         d_ne2 > 0 ? (size_t)d_ne2 : 1,
+                                         d_ne3 > 0 ? (size_t)d_ne3 : 1 };
+
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size_nc, NULL, 0, NULL, NULL));
+    }
+}
+
+static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_timestep_embedding == nullptr) {
+        GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int logical_dim = dst->op_params[0];
+    const int max_period  = dst->op_params[1];
+    const int dst_nb1_bytes = dst->nb[1];
+
+    cl_kernel kernel = backend_ctx->kernel_timestep_embedding;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &dst_nb1_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &logical_dim));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &max_period));
+
+    size_t gws0 = (size_t)(((logical_dim + 1) / 2) + 1);
+
+    size_t gws1 = (size_t)src0->ne[0];
+
+    size_t global_work_size[] = {gws0, gws1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, &evt)); // Pass 2 for 2D problem
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], global_work_size[1], 1};
+    size_t profiling_lws[3] = {0,0,0}; // Reflects NULL LWS
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, NULL)); // Pass 2 for 2D problem
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -5667,6 +6375,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_sigmoid;
                     break;
+                case GGML_UNARY_OP_TANH:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_tanh;
+                    break;
                 default:
                     return false;
             } break;
@@ -5694,6 +6408,36 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_group_norm;
             break;
+                case GGML_OP_REPEAT:
+             if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_repeat;
+            break;
+        case GGML_OP_PAD:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_pad(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_UPSCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_upscale(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_CONCAT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_concat;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_timestep_embedding(backend, tensor->src[0], tensor);
+            return true;
         case GGML_OP_MUL_MAT:
             if (!any_on_device && !ggml_cl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
                 return false;

ggml/src/ggml-opencl/kernels/concat.cl

@@ -0,0 +1,109 @@
+kernel void kernel_concat_f32_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
+    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
+    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
+    int dim
+) {
+    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
+    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
+    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+
+    int i0 = get_global_id(0); // Index along dst's 0th dimension
+    int i1 = get_global_id(1); // Index along dst's 1st dimension
+    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+
+    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
+        return;
+    }
+
+    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
+    ulong src_idx;
+
+    if (dim == 0) {
+        if (i0 < d_ne00) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 1) {
+        if (i1 < d_ne01) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 2) {
+        if (i2 < d_ne02) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+
+            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    }
+}
+
+kernel void kernel_concat_f32_non_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+
+    long ne00, long ne01, long ne02, long ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
+
+    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
+    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
+    int dim
+) {
+    global const char * src0_base = p_src0 + off_src0;
+    global const char * src1_base = p_src1 + off_src1;
+    global char * dst_base        = p_dst + off_dst;
+
+    long current_i1 = get_global_id(0); // Index for dst_dim_1
+    long current_i2 = get_global_id(1); // Index for dst_dim_2
+    long current_i3 = get_global_id(2); // Index for dst_dim_3
+
+    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
+        return;
+    }
+
+    global const float * x_val_ptr;
+    global float * y_val_ptr;
+
+    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
+        bool use_src0;
+        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
+
+        if (dim == 0) {
+            use_src0 = (current_i0 < ne00);
+            if (!use_src0) { s_i0 = current_i0 - ne00; }
+        } else if (dim == 1) {
+            use_src0 = (current_i1 < ne01);
+            if (!use_src0) { s_i1 = current_i1 - ne01; }
+        } else if (dim == 2) {
+            use_src0 = (current_i2 < ne02);
+            if (!use_src0) { s_i2 = current_i2 - ne02; }
+        } else { // dim == 3
+            use_src0 = (current_i3 < ne03);
+            if (!use_src0) { s_i3 = current_i3 - ne03; }
+        }
+
+        if (use_src0) {
+            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+        } else {
+            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+        }
+
+        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
+        *y_val_ptr = *x_val_ptr;
+    }
+}

ggml/src/ggml-opencl/kernels/pad.cl

@@ -0,0 +1,30 @@
+kernel void kernel_pad(
+        global const void * src0_ptr,
+        ulong src0_offset,
+        global void * dst_ptr,
+        ulong dst_offset,
+        int s_ne0, int s_ne1, int s_ne2,
+        int d_ne0, int d_ne1, int d_ne2
+) {
+    global const float * src0 = (global const float *)((global const char *)src0_ptr + src0_offset);
+    global float * dst = (global float *)((global char *)dst_ptr + dst_offset);
+
+    int nidx   = get_global_id(0);
+    int idx_d1 = get_group_id(1);
+    int idx_d2 = get_group_id(2);
+
+    if (nidx >= d_ne0) {
+        return;
+    }
+
+    int dst_el_offset = nidx + idx_d1 * d_ne0 + idx_d2 * d_ne0 * d_ne1;
+
+    bool in_src_bounds = (nidx < s_ne0) && (idx_d1 < s_ne1) && (idx_d2 < s_ne2);
+
+    if (in_src_bounds) {
+        int src_el_offset = nidx + idx_d1 * s_ne0 + idx_d2 * s_ne0 * s_ne1;
+        dst[dst_el_offset] = src0[src_el_offset];
+    } else {
+        dst[dst_el_offset] = 0.0f;
+    }
+}

ggml/src/ggml-opencl/kernels/repeat.cl

@@ -0,0 +1,39 @@
+kernel void kernel_repeat(
+    global const char * src0_data_in,
+    global       char * dst_data_in,
+    ulong src0_offset,
+    ulong dst_offset,
+    int src0_ne0, int src0_ne1, int src0_ne2, int src0_ne3,
+    ulong src0_nb0, ulong src0_nb1, ulong src0_nb2, ulong src0_nb3,
+    int dst_ne0, int dst_ne1, int dst_ne2, int dst_ne3,
+    ulong dst_nb0, ulong dst_nb1, ulong dst_nb2, ulong dst_nb3
+) {
+    global const char * src0_data = src0_data_in + src0_offset;
+    global       char * dst_data  = dst_data_in + dst_offset;
+
+    const int d3 = get_global_id(2);
+    const int d2 = get_global_id(1);
+    const int d1 = get_global_id(0);
+
+    if (d3 >= dst_ne3 || d2 >= dst_ne2 || d1 >= dst_ne1) {
+        return;
+    }
+
+    const int s3 = d3 % src0_ne3;
+    const int s2 = d2 % src0_ne2;
+    const int s1 = d1 % src0_ne1;
+
+    const global char * p_src0_slice = src0_data + (ulong)s3*src0_nb3 + (ulong)s2*src0_nb2 + (ulong)s1*src0_nb1;
+    global char * p_dst_slice  = dst_data  + (ulong)d3*dst_nb3 + (ulong)d2*dst_nb2 + (ulong)d1*dst_nb1;
+
+    for (int d0 = 0; d0 < dst_ne0; ++d0) {
+        // Determine source index for dimension 0 based on tiling/broadcasting.
+        const int s0 = d0 % src0_ne0;
+
+        const global char * restrict current_src_el_ptr = p_src0_slice + (ulong)s0*src0_nb0;
+        global char * restrict current_dst_el_ptr  = p_dst_slice  + (ulong)d0*dst_nb0;
+        for (int k = 0; k < src0_nb0; ++k) {
+            current_dst_el_ptr[k] = current_src_el_ptr[k];
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/tanh.cl

@@ -0,0 +1,63 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+kernel void kernel_tanh_f32_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+    int ne10, int ne11, int ne12, int ne13,
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_tanh_f16_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+    int ne10, int ne11, int ne12, int ne13,
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/tsembd.cl

@@ -0,0 +1,48 @@
+kernel void kernel_timestep_embedding(
+    global const void * p_timesteps,
+    ulong off_timesteps,
+    global void * p_dst,
+    ulong off_dst,
+    int dst_nb1_bytes,
+    int logical_dim,
+    int max_period
+) {
+    int local_i;
+    int local_j;
+    int local_half_dim;
+    float local_timestep_val;
+    float local_freq;
+    float local_arg;
+    global float * local_embed_data_ptr;
+    global const float * local_timesteps_input_ptr;
+    global float * local_dst_output_base_ptr;
+
+    local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps);
+    local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst);
+
+    local_i = get_global_id(1);
+    local_j = get_global_id(0);
+
+    local_half_dim = logical_dim / 2;
+    local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
+
+    if (logical_dim % 2 != 0 && local_j == ((logical_dim + 1) / 2)) {
+        local_embed_data_ptr[logical_dim] = 0.0f;
+    }
+
+    if (local_j >= local_half_dim) {
+        return;
+    }
+
+    local_timestep_val = local_timesteps_input_ptr[local_i];
+
+    if (local_half_dim == 0) {
+        local_freq = 1.0f;
+    } else {
+        local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim);
+    }
+
+    local_arg = local_timestep_val * local_freq;
+    local_embed_data_ptr[local_j] = cos(local_arg);
+    local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg);
+}

ggml/src/ggml-opencl/kernels/upscale.cl

@@ -0,0 +1,121 @@
+kernel void kernel_upscale(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne10,
+    int ne11,
+    int ne12,
+    int ne13,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10 * ne11 * ne12 * ne13;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = index / (ne10 * ne11 * ne12);
+
+    int i00 = (int)(i10 / sf0);
+    int i01 = (int)(i11 / sf1);
+    int i02 = (int)(i12 / sf2);
+    int i03 = (int)(i13 / sf3);
+
+    ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00;
+    global const float * src_element_ptr = (global const float *)(src_base + offset_src_element);
+
+    dst_base[index] = *src_element_ptr;
+}
+
+kernel void kernel_upscale_bilinear(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne00_src,
+    int ne01_src,
+    int ne10_dst,
+    int ne11_dst,
+    int ne12_dst,
+    int ne13_dst,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10_dst = index % ne10_dst;
+    int i11_dst = (index / ne10_dst) % ne11_dst;
+    int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    int i02_src = (int)(i12_dst / sf2);
+    int i03_src = (int)(i13_dst / sf3);
+
+    const float pixel_offset = 0.5f;
+
+    float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    long y0_src = (long)floor(y_src_f);
+    long y1_src = y0_src + 1;
+
+    y0_src = max(0L, min(y0_src, (long)ne01_src - 1));
+    y1_src = max(0L, min(y1_src, (long)ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    long x0_src = (long)floor(x_src_f);
+    long x1_src = x0_src + 1;
+
+    x0_src = max(0L, min(x0_src, (long)ne00_src - 1));
+    x1_src = max(0L, min(x1_src, (long)ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst_base[index] = result;
+}

ggml/src/ggml-sycl/cpy.cpp

@@ -1,8 +1,12 @@
 #include "cpy.hpp"
 
 #include <float.h>
+#include <string>
 
 #include "dequantize.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
+#include "ggml.h"
 
 static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) {
@@ -116,6 +120,15 @@ static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
     }
 }
 
+/* quantized type same copy */
+template<typename T>
+static void cpy_blck_q_q(const char * cxi, char * cdsti) {
+    const T * xi = (const T *) cxi;
+    T * dsti = (T *) cdsti;
+    *dsti = *xi;
+}
+
+
 static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *) (cdsti);
 
@@ -311,6 +324,34 @@ template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const
     }
 }
 
+
+template <typename T, int qk>
+static void cpy_q_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02,
+                      const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11,
+                      const int ne12, const int nb10, const int nb11, const int nb12, const int nb13,
+                      const sycl::nd_item<3> & item_ct1) {
+    const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03      = i / (ne00 * ne01 * ne02);
+    const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
+    const int i00      = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00;
+    const int x_offset = (i00 / qk) * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03;
+
+
+    const int i13        = i / (ne10 * ne11 * ne12);
+    const int i12        = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11);
+    const int i11        = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10;
+    const int i10        = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10;
+    const int dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13;
+
+    cpy_blck_q_q<T>(cx + x_offset, cdst + dst_offset);
+}
+
 template <cpy_kernel_t cpy_blck, int qk>
 static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02,
                       const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11,
@@ -322,6 +363,7 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00
         return;
     }
 
+
     const int i03      = i / (ne00 * ne01 * ne02);
     const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
     const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
@@ -615,6 +657,70 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
     }
 }
 
+static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+
+   const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+   stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
     scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
@@ -632,8 +738,10 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
 
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
-
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+    if ((src0->type == src1->type) && (ggml_is_contiguous(src0) && ggml_is_contiguous(src1))) {
+        GGML_SYCL_DEBUG("%s: memcpy path\n", __func__);
+        main_stream->memcpy(src1_ddc, src0_ddc, ggml_nbytes(src0));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
@@ -684,6 +792,16 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
         ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                  nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_q8_0_q8_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_q5_0_q5_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_Q5_1) {
+        ggml_cpy_q5_1_q5_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_q4_0_q4_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_q4_1_q4_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else {
         GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type),
                        ggml_type_name(src1->type));

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -1434,6 +1434,59 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
     reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
 }
 
+template <int ElementsPerWI>
+static __dpct_inline__ void quantize_and_reorder_q8_1(const float * __restrict__ x, void * reordered_q8_tensor,
+                                                      const int kx, const int kx_padded, const sycl::nd_item<1> & it) {
+    /*
+        Quantizes and reorders the resultant q8 tensor in a per row fashion
+        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
+    */
+
+    auto subgroup_id = it.get_group(0);
+    auto wi_id       = it.get_local_id(0);
+
+    const int num_blocks_per_row = kx / QK8_1;
+    auto      row                = subgroup_id / num_blocks_per_row;
+    auto      col                = subgroup_id % num_blocks_per_row;
+
+    auto row_offset = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
+    auto col_offset = QK8_1 * col + wi_id * ElementsPerWI;
+
+    auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
+    auto ds_ptr    = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
+
+    sycl::vec<float, ElementsPerWI>  wi_f32_vals;
+    sycl::vec<int8_t, ElementsPerWI> quantized_values;
+
+    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
+    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
+
+    float sum  = 0.0f;
+    float amax = 0.0f;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        sum += wi_f32_vals[i];
+        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
+        quantized_values[i] = 0;
+    }
+    sum     = sycl::reduce_over_group(it.get_group(), sum, sycl::plus<float>());
+    amax    = sycl::reduce_over_group(it.get_group(), amax, sycl::maximum<float>());
+    float d = amax == 0 ? 1 : amax / 127;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
+    }
+
+    d = amax == 0 ? 0 : d;
+
+    *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
+    if (wi_id == 0) {
+        *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
+    }
+}
+
 static void mul_mat_p021_f16_f32(
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y,
@@ -1718,23 +1771,30 @@ static  void pool2d_nchw_kernel(
         o_ptr[cur_oh * ow + cur_ow] = res;
 }
 
-static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
-                                   const int ky, const int kx_padded,
-                                   queue_ptr stream) {
-    const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
-    const sycl::range<3> num_blocks(1, ky, block_num_x);
-    int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
-    static_assert(QK8_1 % WARP_SIZE == 0);
-    const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+static void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
+                                   bool reorder_q8_tensor, queue_ptr stream) {
+    if (reorder_q8_tensor) {
+        auto local_range      = std::size_t(WARP_SIZE);
+        auto num_quant_blocks = ky * (kx / QK8_1);
+        auto global_range     = num_quant_blocks * local_range;
+        stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
+                             [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 quantize_and_reorder_q8_1<QK8_1 / WARP_SIZE>(x, vy, kx, kx_padded, it);
+                             });
+    } else {
+        const int            block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
+        const sycl::range<3> num_blocks(1, ky, block_num_x);
+        int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
+        static_assert(QK8_1 % WARP_SIZE == 0);
+        const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
+        {
+            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
-            sycl::nd_range<3>(num_blocks * block_size, block_size),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
-            });
+            stream->parallel_for(sycl::nd_range<3>(num_blocks * block_size, block_size),
+                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                     quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
+                                 });
+        }
     }
 }
 
@@ -2446,9 +2506,10 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
             dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
 
             if (src1_on_device && src1_is_contiguous) {
+                bool reorder_q8_tensor = src0->extra && ((ggml_tensor_extra_gpu *)src0->extra)->optimized_feature.reorder;
                 scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                      /*num_src=*/2, " : converting src1 to Q8_1");
-                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, reorder_q8_tensor, stream);
                 /*
                 DPCT1010:90: SYCL uses exceptions to report errors and does not
                 use the error codes. The call was replaced with 0. You need to
@@ -2554,7 +2615,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 if (convert_src1_to_q8_1 && !src1_is_contiguous) {
                     scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                          /*num_src=*/2, " : converting src1 to Q8_1");
-                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, false, stream);
                     /*
                     DPCT1010:92: SYCL uses exceptions to report errors and does
                     not use the error codes. The call was replaced with 0. You
@@ -4165,6 +4226,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
+                if (src0_type == src1_type && (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) && src0_type != GGML_TYPE_BF16) {
+                    return true;
+                }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
@@ -4210,6 +4274,21 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
+                if(src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
                 return false;
             }
         case GGML_OP_CONCAT:

ggml/src/ggml-sycl/mmvq.cpp

@@ -29,8 +29,6 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
     static_assert(blocks_per_subgroup > 0);
     static_assert(block_elements_per_subgroup > 0);
 
-    const block_q8_1 * y = (const block_q8_1 *) vy;
-
     float partial_sum = 0.0f;
     for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
         const int ibx       = row * blocks_per_row + i;  // x block index
@@ -40,13 +38,15 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
 
         // Y block index that aligns with ibx
         const int iby = i * block_type::block_to_q8_1_ratio();
+        const int8_t* q8_1_quant_ptr = (const int8_t*)vy + iby * QK8_1;
+        const sycl::half2* q8_1_ds_ptr = (const sycl::half2*)((const char*)vy + ncols + iby * sizeof(sycl::half2));
 
 #pragma unroll
         for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) {
             // x block quant index when casting the quants to int
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs, nblocks);
         }
     }
 

ggml/src/ggml-sycl/vecdotq.hpp

@@ -285,21 +285,21 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     }
 
     __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int /* nblocks */) {
+                     const int8_t* q8_1_quant_ptr, const sycl::half2* q8_1_ds, const int & iqs, int /* nblocks */) {
         const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
         const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
         int             v[q4_0_traits::vdr_mmvq];
         int             u[2 * q4_0_traits::vdr_mmvq];
 
-#pragma unroll
 
+#pragma unroll
         for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) {
             v[i]         = get_int_from_uint8(bq4_0, iqs + i);
-            u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-            u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + q4_0_traits::qi);
+            u[2 * i + 0] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i);
+            u[2 * i + 1] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i + q4_0_traits::qi);
         }
 
-        return vec_dot_q4_0_q8_1_impl(v, u, d, bq8_1->ds);
+        return vec_dot_q4_0_q8_1_impl(v, u, d, *q8_1_ds);
     };
 };
 
@@ -347,7 +347,7 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     using q4_k_traits = typename q4_k_block::traits;
 
     float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int nblocks) {
+                     const int8_t* q8_1_quant_ptr, const sycl::half2* q8_1_ds, const int & iqs, int nblocks) {
         const int ib = ibx_offset / (QK_K / 2);
 
         const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
@@ -360,7 +360,38 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
         const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
         const uint16_t * scales     = (const uint16_t *) scs;
 
-        return vec_dot_q4_K_q8_1_common(q4, scales, *dms, bq8_1, iqs);
+        int   v[2];
+        int   u[2 * QR4_K];
+        float d8[QR4_K];
+
+        v[0] = q4[0];
+        v[1] = q4[4];
+
+        uint16_t  aux[2];
+        const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+        if (j < 2) {
+            aux[0] = scales[j + 0] & 0x3f3f;
+            aux[1] = scales[j + 2] & 0x3f3f;
+        } else {
+            aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+            aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+        }
+
+        const uint8_t * sc = (const uint8_t *) aux;
+        const uint8_t * m  = sc + 2;
+
+        for (int i = 0; i < QR4_K; ++i) {
+            const int8_t* quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            sycl::half2 ds_values = *(q8_1_ds + bq8_offset + i);
+
+            d8[i]                   = ds_values[0];
+
+            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
+            u[2 * i + 0]   = q8[0];
+            u[2 * i + 1]   = q8[4];
+        }
+
+        return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, *dms, d8);
     }
 };
 

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -196,6 +196,7 @@ enum vk_device_architecture {
     AMD_RDNA1,
     AMD_RDNA2,
     AMD_RDNA3,
+    INTEL_XE2,
 };
 
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
@@ -246,6 +247,34 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             }
             return vk_device_architecture::AMD_RDNA2;
         }
+    } else if (props.vendorID == VK_VENDOR_ID_INTEL) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool subgroup_size_control = false;
+
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                subgroup_size_control = true;
+            }
+        }
+
+        if (!subgroup_size_control) {
+            return vk_device_architecture::OTHER;
+        }
+
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
+        props2.pNext = &subgroup_size_control_props;
+        device.getProperties2(&props2);
+
+        if (subgroup_size_control_props.minSubgroupSize == 16) {
+            // Xe2 architecture uses SIMD16 while previous Xe and Gen architecture uses SIMD8.
+            // Minimum subgroup size matches the SIMD width so we distinguish architecture by checking this value.
+            // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
+            // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
+            return vk_device_architecture::INTEL_XE2;
+        }
     }
     return vk_device_architecture::OTHER;
 }
@@ -396,6 +425,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_count_equal_i32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
+    vk_pipeline pipeline_conv_transpose_1d_f32;
     vk_pipeline pipeline_pool2d_f32;
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
@@ -444,7 +474,7 @@ struct vk_device_struct {
     // for GGML_VK_PERF_LOGGER
     std::unique_ptr<vk_perf_logger> perf_logger;
     vk::QueryPool query_pool;
-    uint32_t num_queries;
+    int32_t num_queries;
 
     ~vk_device_struct() {
         VK_LOG_DEBUG("destroy device " << name);
@@ -706,6 +736,21 @@ struct vk_op_timestep_embedding_push_constants {
     uint32_t max_period;
 };
 
+struct vk_op_conv_transpose_1d_push_constants {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+};
+
 struct vk_op_pool2d_push_constants {
     uint32_t IW; uint32_t IH;
     uint32_t OW; uint32_t OH;
@@ -2726,6 +2771,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_conv_transpose_1d_f32, "conv_transpose_1d_f32", conv_transpose_1d_f32_len, conv_transpose_1d_f32_data, "main", 3, sizeof(vk_op_conv_transpose_1d_push_constants), {1, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
@@ -4061,7 +4108,33 @@ static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bo
     return s;
 }
 
-static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
+template <typename T> size_t push_constant_size(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    GGML_UNUSED(t);
+    return sizeof(T);
+}
+template <typename T> size_t push_constant_size(const std::vector<T> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * t.size();
+}
+template <typename T, uint32_t N> size_t push_constant_size(const std::array<T, N> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * N;
+}
+
+template <typename T> const T *push_constant_data(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    return &t;
+}
+template <typename T> const T *push_constant_data(const std::vector<T> &t) {
+    return t.data();
+}
+template <typename T, uint32_t N> const T *push_constant_data(const std::array<T, N> &t) {
+    return t.data();
+}
+
+template <typename T>
+static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, const T &push_constants, std::array<uint32_t, 3> elements) {
     const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
     const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
     const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
@@ -4077,7 +4150,7 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
     vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
     ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
 
-    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
+    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size(push_constants), push_constant_data(push_constants));
     subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
     subctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
                                 pipeline->layout,
@@ -4540,7 +4613,7 @@ static void ggml_vk_matmul(
     ggml_vk_sync_buffers(subctx);
     if (split_k == 1) {
         const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
         return;
     }
 
@@ -4548,10 +4621,10 @@ static void ggml_vk_matmul(
 
     const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);
     const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
 }
 
 static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) {
@@ -4599,7 +4672,7 @@ static void ggml_vk_matmul_id(
     ggml_vk_sync_buffers(subctx);
     const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
                                               nei0, nei1, nbi1, ne11, padded_n };
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
 }
 
 static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
@@ -4720,7 +4793,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     };
     init_pushconst_fastdiv(pc);
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
 }
 
 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
@@ -4739,7 +4812,7 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
     vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
 
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(uint32_t), &ne, { ne, 1, 1 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -4939,7 +5012,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     if (y_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -5155,7 +5228,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                               { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
-                              sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+                              pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 }
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5243,7 +5316,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     }
 
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, workgroups_z });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, workgroups_z });
 }
 
 static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5326,7 +5399,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
-        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5542,7 +5615,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
-            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     if (y_non_contig) {
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -5762,7 +5835,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
         { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
         vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
-        sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z });
+        pc, { groups_x, (uint32_t)nei0, groups_z });
 }
 
 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
@@ -6112,7 +6185,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // there's no more than one tile of rows (i.e. workgroups_x would have been
                                     // one). We reuse workgroups_x to mean the number of splits, so we need to
                                     // cancel out the divide by wg_denoms[0].
-                                    sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
+                                    pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
         const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
@@ -6121,7 +6194,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
-                                    pc2.size() * uint32_t{sizeof(uint32_t)}, pc2.data(), { (uint32_t)ne1, 1, 1 });
+                                    pc2, { (uint32_t)ne1, 1, 1 });
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -6131,7 +6204,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE},
                                         vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                     },
-                                    sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x, workgroups_y, workgroups_z });
+                                    pc, { workgroups_x, workgroups_y, workgroups_z });
     }
 }
 
@@ -6392,6 +6465,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_timestep_embedding_f32;
         }
         return nullptr;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_conv_transpose_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_POOL_2D:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_pool2d_f32;
@@ -6726,6 +6804,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             uint32_t half_ceil = (dim + 1) / 2;
             elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
         } break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        {
+            elements = {uint32_t(src0->ne[1]), 1, 1}; // parallelize in {Cout, 1, 1}
+        } break;
     case GGML_OP_POOL_2D:
         {
             const uint32_t N = dst->ne[3];
@@ -6800,7 +6882,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
 
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
         // Empty src2 is possible in rope, but the shader needs a buffer
         vk_subbuffer subbuf_z;
@@ -6811,26 +6893,26 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
 
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_IM2COL) {
         // im2col uses only src1 and dst buffers
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_COUNT_EQUAL) {
         ggml_vk_sync_buffers(subctx);
         // count_equal assumes that destination buffer is initialized with zeroes
         ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src2) {
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (use_src1) {
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else {
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     }
 }
 
@@ -6999,7 +7081,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
             vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
             vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
             vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
+        }, pc, elements);
     } else if (version == 7) {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
             vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
@@ -7010,7 +7092,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
             vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
             vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] },
             vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, sizeof(vk_op_rwkv_wkv7_push_constants), &pc, elements);
+        }, pc, elements);
     } else {
         // shouldn't happen
         GGML_ASSERT(false);
@@ -7147,7 +7229,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
         vk_subbuffer{ d_GM, gm_offset, gm_size },
         vk_subbuffer{ d_GV, gv_offset, gv_size },
         vk_subbuffer{ d_P, p_offset, p_size },
-    }, sizeof(vk_op_push_constants), &pc, elements);
+    }, pc, elements);
 }
 
 static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
@@ -7529,6 +7611,37 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
     }, dryrun);
 }
 
+static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    // src0: (K, Cout, Cin, 1) -- kernel
+    // src1: (L, Cin, 1, 1) -- input
+    // dst: (*, Cout, 1, 1)
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    const int32_t s0 = dst->op_params[0];
+
+    vk_op_conv_transpose_1d_push_constants p{};
+    p.Cout = static_cast<uint32_t>(ne01);
+    p.Cin = static_cast<uint32_t>(ne02);
+    p.K = static_cast<uint32_t>(ne00);
+    p.L = static_cast<uint32_t>(ne10);
+    p.KL = static_cast<uint32_t>(ne0);
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.s0 = static_cast<uint32_t>(s0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
+}
+
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
     const int32_t k1 = dst->op_params[1];
@@ -8005,7 +8118,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
-    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc, { (uint32_t)ne, 1, 1});
     ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
@@ -8600,6 +8713,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
@@ -8664,6 +8778,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_POOL_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
@@ -8835,6 +8950,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_TIMESTEP_EMBEDDING:
         ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
@@ -8963,6 +9082,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
@@ -9513,8 +9633,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             if (ctx->device->query_pool) {
                 ctx->device->device.destroyQueryPool(ctx->device->query_pool);
             }
-            VkQueryPoolCreateInfo query_create_info = { VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO };
-            query_create_info.queryType = VK_QUERY_TYPE_TIMESTAMP;
+            vk::QueryPoolCreateInfo query_create_info;
+            query_create_info.queryType = vk::QueryType::eTimestamp;
             query_create_info.queryCount = cgraph->n_nodes + 100;
             ctx->device->query_pool = ctx->device->device.createQueryPool(query_create_info);
             ctx->device->num_queries = query_create_info.queryCount;
@@ -9600,7 +9720,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
         // Get the results and pass them to the logger
         std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
-        ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait);
+        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
         for (int i = 0; i < cgraph->n_nodes; i++) {
             if (!ggml_vk_is_empty(cgraph->nodes[i])) {
                 ctx->device->perf_logger->log_timing(cgraph->nodes[i], uint64_t((timestamps[i+1] - timestamps[i]) * ctx->device->properties.limits.timestampPeriod));
@@ -10024,6 +10144,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_OPT_STEP_ADAMW:
             return true;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         default:
             return false;
     }
@@ -10170,8 +10292,9 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
     switch (props.vendorID) {
     case VK_VENDOR_ID_INTEL:
-        // Intel drivers don't support coopmat properly yet
-        return false;
+        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
+        // while some older hardware (ex. Arc A770) has performance regressions
+        return arch == vk_device_architecture::INTEL_XE2;
     case VK_VENDOR_ID_AMD:
         if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
             // Workaround for AMD proprietary driver reporting support on all GPUs
@@ -10515,6 +10638,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t dim = tensor->op_params[0];
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
+    } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t p0 = tensor->op_params[1];
+        const int32_t d0 = tensor->op_params[2];
+        tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
     } else if (tensor->op == GGML_OP_POOL_2D) {
         enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
         const int32_t k0 = tensor->op_params[1];

ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp

@@ -0,0 +1,98 @@
+#version 450
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};     // dst - result      [KL, Cout]
+
+layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+} p;
+
+
+uint32_t Cout_idx = gl_WorkGroupID.x;
+const uint32_t bs = gl_WorkGroupSize.x;
+uint32_t tid = gl_LocalInvocationID.x;
+// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
+uint32_t tmp_len = bs*p.s0+p.K;
+shared D_TYPE tmp[4096];
+
+uint splitWork(uint workSize){
+    return (bs + workSize -1) / bs;
+}
+
+void main(){
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        if(idx < tmp_len){
+            tmp[idx] = 0.0;
+        }
+    }
+
+    uint32_t L_blocks = splitWork(p.L);
+    for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
+        if(L_block_id > 0){
+            barrier();
+            // Shift values in tmp to the current processing window
+            for(int i = 0; i < splitWork(tmp_len); i++){
+                uint32_t idx = i*bs+tid;
+                if(idx >= bs*p.s0 && idx < tmp_len){
+                    tmp[idx-bs*p.s0] = tmp[idx];
+                    tmp[idx] = 0.0;
+                }else if(idx >= p.K && idx < bs*p.s0){
+                    tmp[idx] = 0.0;
+                }
+            }
+        }
+        barrier();
+
+        // Save contributions of the block to tmp
+        uint32_t L_idx = L_block_id*bs + tid;
+        for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
+            D_TYPE dp = 0.0;
+            for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
+                A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
+                if(L_idx < p.L){
+                    B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
+                    dp = fma(elemKrn, elemInp, dp);
+                }
+            }
+            tmp[tid*p.s0 + K_idx] += dp;
+            barrier();
+        }
+
+        // Save the computed values except the last block that can have different size
+        uint32_t KLb_idx = L_block_id*bs*p.s0;
+        if(L_block_id < L_blocks-1){
+            for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
+                uint32_t sh_idx = p.s0*tid+s0_idx;
+                uint32_t KL_idx = KLb_idx+sh_idx;
+                if(KL_idx < p.KL){
+                    data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
+                }
+            }
+        }
+    }
+
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
+        if(KL_idx < p.KL){
+            data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -622,6 +622,8 @@ void process_shaders() {
 
     string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"},  {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+
     string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

ggml/src/gguf.cpp

@@ -347,6 +347,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     int64_t n_tensors = 0;
 
     if (ok && gr.read(ctx->version)) {
+        if (ok && ctx->version == 0) {
+            GGML_LOG_ERROR("%s: bad GGUF version: %" PRIu32 "\n", __func__, ctx->version);
+            ok = false;
+        }
+
         /*
          * bit layout is different when reading non-native endian models.
          * assuming that the GGUF version is 3, the non-native endian model
@@ -354,18 +359,16 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
          * the last 4 hexadecimal digits to check if the model is the same
          * endianness as the host system.
         */
-        if ((ctx->version & 0x0000FFFF) == 0x00000000) {
+        if (ok && (ctx->version & 0x0000FFFF) == 0x00000000) {
             GGML_LOG_ERROR("%s: failed to load model: this GGUF file version %" PRIu32 " is extremely large, is there a mismatch between the host and model endianness?\n", __func__, ctx->version);
-            gguf_free(ctx);
-            return nullptr;
+            ok = false;
         }
 
-        GGML_ASSERT(ctx->version > 0 && ctx->version <= 65535);
-        if (ctx->version == 1) {
+        if (ok && ctx->version == 1) {
             GGML_LOG_ERROR("%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
             ok = false;
         }
-        if (ctx->version > GGUF_VERSION) {
+        if (ok && ctx->version > GGUF_VERSION) {
             GGML_LOG_ERROR("%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
                 __func__, ctx->version, GGUF_VERSION);
             ok = false;

gguf-py/gguf/gguf_writer.py

@@ -935,6 +935,9 @@ class GGUFWriter:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
+        self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
+
     # for vision models
 
     def add_clip_has_vision_encoder(self, value: bool) -> None:

include/llama.h

@@ -61,7 +61,10 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
-    struct llama_kv_cache;
+
+    typedef struct llama_memory_i * llama_memory_t;
+
+    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -493,9 +496,11 @@ extern "C" {
     DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
 
     LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
-    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
     LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
 
+    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
+
     LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
     LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
 
@@ -509,6 +514,13 @@ extern "C" {
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
 
+    // Returns the number of classifier outputs (only valid for classifier models)
+    // Undefined behavior for non-classifier models
+    LLAMA_API uint32_t llama_model_n_cls_out(const struct llama_model * model);
+
+    // Returns label of classifier output by index (<n_cls_out). Returns nullptr if no label provided
+    LLAMA_API const char * llama_model_cls_label(const struct llama_model * model, uint32_t i);
+
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_vocab * vocab);
 
     LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab);
@@ -609,7 +621,81 @@ extern "C" {
                          int32_t   il_end);
 
     //
-    // KV cache
+    // Memory
+    //
+
+    // Clear the memory contents
+    // If data == true, the data buffers will also be cleared together with the metadata
+    LLAMA_API void llama_memory_clear(
+            llama_memory_t mem,
+                      bool data);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API bool llama_memory_seq_rm(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_cp(
+            llama_memory_t mem,
+              llama_seq_id seq_id_src,
+              llama_seq_id seq_id_dst,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_memory_seq_keep(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_add(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                 llama_pos delta);
+
+    // Integer division of the positions by factor of `d > 1`
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_div(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                       int d);
+
+    // Returns the smallest position present in the memory for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_min(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Returns the largest position present in the memory for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_max(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Check if the memory supports shifting
+    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
+
+    //
+    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
     //
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
@@ -622,86 +708,95 @@ extern "C" {
                "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_self_clear(
-            struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_clear(
+                struct llama_context * ctx),
+            "Use llama_memory_clear() instead");
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_self_seq_rm(
+    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_rm() instead");
 
     // Copy all tokens that belong to the specified sequence to another sequence
     // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_cp(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
             struct llama_context * ctx,
                     llama_seq_id   seq_id_src,
                     llama_seq_id   seq_id_dst,
                        llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_cp() instead");
 
     // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_self_seq_keep(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_keep() instead");
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_add(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
-                       llama_pos   delta);
+                       llama_pos   delta),
+            "Use llama_memory_seq_add() instead");
 
     // Integer division of the positions by factor of `d > 1`
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_div(
+    DEPRECATED(void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
-                             int   d);
+                             int   d),
+            "Use llama_memory_seq_div() instead");
 
     // Returns the smallest position present in the KV cache for the specified sequence
     // This is typically non-zero only for SWA caches
     // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_min() instead");
 
     // Returns the largest position present in the KV cache for the specified sequence
     // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_max() instead");
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    LLAMA_API DEPRECATED(void llama_kv_self_defrag(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
             "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
 
     // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
+            "use llama_memory_can_shift() instead");
 
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API DEPRECATED(void llama_kv_self_update(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
             "simply remove this call, updates are applied lazily on the next llama_decode()");
 
     //
@@ -709,7 +804,7 @@ extern "C" {
     //
 
     // Returns the *actual* size in bytes of the state
-    // (logits, embedding and kv_cache)
+    // (logits, embedding and memory)
     // Only use when saving the state, not when restoring it, otherwise the size may be too small.
     LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
     LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -765,12 +860,12 @@ extern "C" {
                           size_t   n_token_count),
         "use llama_state_save_file instead");
 
-    // Get the exact size needed to copy the KV cache of a single sequence
+    // Get the exact size needed to copy the state of a single sequence
     LLAMA_API size_t llama_state_seq_get_size(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
-    // Copy the KV cache of a single sequence into the specified buffer
+    // Copy the state of a single sequence into the specified buffer
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
@@ -836,16 +931,16 @@ extern "C" {
     // For encode-decoder contexts, processes the batch using the encoder.
     // Can store the encoder output internally for later use by the decoder's cross-attention layers.
     //   0 - success
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // < 0 - error. the memory state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
     // Process a batch of tokens.
-    // Requires KV cache.
+    // Requires the context to have a memory.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the memory state is restored to the state before this call
     //    0 - success
     //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
     //    2 - aborted
@@ -916,7 +1011,7 @@ extern "C" {
 
     // Get the embeddings for a sequence id
     // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[n_cls_out] with the rank(s) of the sequence
     // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 

src/CMakeLists.txt

@@ -20,7 +20,6 @@ add_library(llama
             llama-hparams.cpp
             llama-impl.cpp
             llama-io.cpp
-            llama-kv-cache.cpp
             llama-kv-cache-unified.cpp
             llama-kv-cache-unified-iswa.cpp
             llama-kv-cache-recurrent.cpp

src/llama-arch.cpp

@@ -200,7 +200,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
     { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
-    { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,      "tokenizer.chat_template.%s"              },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
@@ -1707,8 +1706,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
 std::string LLM_KV::operator()(llm_kv kv) const {
-    return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
-        : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+
+    if (suffix != nullptr) {
+        name += ".";
+        name += suffix;
+    }
+
+    return name;
 }
 
 std::string LLM_TN_IMPL::str() const {

src/llama-arch.h

@@ -196,7 +196,6 @@ enum llm_kv {
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
-    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,

src/llama-context.cpp

@@ -2,9 +2,9 @@
 
 #include "llama-impl.h"
 #include "llama-io.h"
+#include "llama-memory.h"
 #include "llama-mmap.h"
 #include "llama-model.h"
-#include "llama-kv-cache.h"
 
 #include <cinttypes>
 #include <cstring>
@@ -123,7 +123,7 @@ llama_context::llama_context(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    if (!params.swa_full && cparams.n_seq_max > 1) {
+    if (!params.swa_full && cparams.n_seq_max > 1 && hparams.is_swa_any()) {
         LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
                 __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
     }
@@ -277,10 +277,9 @@ llama_context::llama_context(
         int n_nodes_tg  = -1;
 
         // simulate full KV cache
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-        const auto kv_state = kv_self->init_full();
-        if (!kv_state) {
+        const auto mstate = memory->init_full();
+        if (!mstate) {
             throw std::runtime_error("failed to initialize KV cache");
         }
 
@@ -288,7 +287,7 @@ llama_context::llama_context(
 
         // reserve pp graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -299,7 +298,7 @@ llama_context::llama_context(
 
         // reserve with tg graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(1, 1, 1, kv_state.get());
+            auto * gf = graph_reserve(1, 1, 1, mstate.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -310,7 +309,7 @@ llama_context::llama_context(
 
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -419,40 +418,68 @@ uint32_t llama_context::n_threads_batch() const {
     return cparams.n_threads_batch;
 }
 
-llama_kv_cache * llama_context::get_kv_self() {
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-    return kv_self;
+llama_memory_t llama_context::get_memory() const {
+    return memory.get();
 }
 
-const llama_kv_cache * llama_context::get_kv_self() const {
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-    return kv_self;
+// deprecated
+void llama_context::kv_self_defrag_sched() {
+    if (!memory) {
+        return;
+    }
+
+    memory_force_optimize = true;
 }
 
-bool llama_context::kv_self_update() {
+// deprecated
+bool llama_context::kv_self_update(bool optimize) {
     if (!memory) {
         return false;
     }
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    {
+        // TODO: remove in the future
+        optimize |= memory_force_optimize;
+        memory_force_optimize = false;
 
-    if (!kv_self->update(*this)) {
-        // no updates have been performed
-        return false;
+        const auto mstate = memory->init_update(this, optimize);
+        switch (mstate->get_status()) {
+            case LLAMA_MEMORY_STATUS_SUCCESS:
+                {
+                    // noop
+                } break;
+            case LLAMA_MEMORY_STATUS_NO_UPDATE:
+                {
+                    // no updates need to be performed
+                    return false;
+                }
+            case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+            case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+                {
+                    LLAMA_LOG_ERROR("%s: failed to prepare memory update\n", __func__);
+                    return false;
+                }
+        }
+
+        if (!mstate->apply()) {
+            LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__);
+        }
     }
 
-    // if the KV cache did any computation, we have to reserve a new worst-case graph
-    const auto kv_state = kv_self->init_full();
-    if (!kv_state) {
-        throw std::runtime_error("failed to initialize KV cache");
-    }
+    // if the memory module did any computation, we have to reserve a new worst-case graph
+    {
+        const auto mstate = memory->init_full();
+        if (!mstate) {
+            throw std::runtime_error("failed to initialize memory state");
+        }
 
-    const uint32_t n_seqs   = cparams.n_seq_max;
-    const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+        const uint32_t n_seqs   = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
-    auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
-    if (!gf) {
-        LLAMA_LOG_ERROR("%s: failed to reserve graph after the KV cache update\n", __func__);
+        auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mstate.get());
+        if (!gf) {
+            LLAMA_LOG_ERROR("%s: failed to reserve graph after the memory update\n", __func__);
+        }
     }
 
     return true;
@@ -814,16 +841,17 @@ int llama_context::encode(llama_batch & inp_batch) {
                 } break;
             case LLAMA_POOLING_TYPE_RANK:
                 {
-                    // extract the rerank score - a single float per sequence
+                    // extract the rerank score - n_cls_out floats per sequence
                     auto & embd_seq_out = embd_seq;
+                    const uint32_t n_cls_out = hparams.n_cls_out;
 
                     for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
                         const llama_seq_id seq_id = ubatch.seq_id[s][0];
                         if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                             continue;
                         }
-                        embd_seq_out[seq_id].resize(1);
-                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
+                        embd_seq_out[seq_id].resize(n_cls_out);
+                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_id)*sizeof(float), n_cls_out*sizeof(float));
                     }
                 } break;
             case LLAMA_POOLING_TYPE_UNSPECIFIED:
@@ -880,10 +908,8 @@ int llama_context::decode(llama_batch & inp_batch) {
         }
     }
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
     // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->seq_pos_max(0) + 1);
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : memory->seq_pos_max(0) + 1);
 
     const llama_batch & batch = batch_allocr.batch;
 
@@ -940,42 +966,49 @@ int llama_context::decode(llama_batch & inp_batch) {
         n_outputs_all = 1;
     }
 
+    bool did_optimize = false;
+
     // handle any pending defrags/shifts
-    kv_self_update();
+    kv_self_update(false);
 
-    llama_memory_state_ptr kv_state;
-
-    bool did_defrag = false;
+    llama_memory_state_ptr mstate;
 
     while (true) {
-        kv_state = kv_self->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ n_outputs_all == n_tokens_all);
-        if (!kv_state) {
+        mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ n_outputs_all == n_tokens_all);
+        if (!mstate) {
             return -2;
         }
 
-        switch (kv_state->get_status()) {
+        switch (mstate->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
                 {
                 } break;
+            case LLAMA_MEMORY_STATUS_NO_UPDATE:
+                {
+                    LLAMA_LOG_ERROR("%s: unexpected memory state status: %d\n", __func__, mstate->get_status());
+
+                    return -2;
+                }
             case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
                 {
-                    if (!did_defrag) {
-                        did_defrag = true;
+                    if (!did_optimize) {
+                        did_optimize = true;
 
-                        kv_self->defrag_sched(-1.0f);
-                        if (kv_self_update()) {
-                            LLAMA_LOG_DEBUG("%s: failed to init batch of size %d, retrying after defrag\n", __func__, batch.n_tokens);
+                        if (kv_self_update(true)) {
+                            LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, batch.n_tokens);
 
                             continue;
                         }
                     }
 
-                    LLAMA_LOG_WARN("%s: failed to find KV cache slot for batch of size %d\n", __func__, batch.n_tokens);
+                    LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, batch.n_tokens);
 
                     return 1;
                 }
             case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
                 {
+                    LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, batch.n_tokens);
+
                     return -2;
                 }
         }
@@ -992,7 +1025,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     int64_t n_outputs_prev = 0;
 
     do {
-        const auto & ubatch = kv_state->get_ubatch();
+        const auto & ubatch = mstate->get_ubatch();
 
         // count the outputs in this u_batch
         {
@@ -1015,11 +1048,14 @@ int llama_context::decode(llama_batch & inp_batch) {
         ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
 
         ggml_status status;
-        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, kv_state.get(), status);
+        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mstate.get(), status);
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
-            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES] = { std::numeric_limits<llama_pos>::max() };
+            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
+            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+                pos_min[s] = std::numeric_limits<llama_pos>::max();
+            }
 
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
@@ -1034,7 +1070,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 
                 LLAMA_LOG_WARN("%s: removing KV cache entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);
 
-                llama_kv_self_seq_rm(this, s, pos_min[s], -1);
+                memory->seq_rm(s, pos_min[s], -1);
             }
 
             switch (status) {
@@ -1128,7 +1164,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         }
 
         n_outputs_prev += n_outputs;
-    } while (kv_state->next());
+    } while (mstate->next());
 
     // set to total number of outputs in the batch, for use in llama_get_logits_ith
     n_outputs = n_outputs_all;
@@ -1137,7 +1173,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     {
         bool sorted_output = true;
 
-        auto & out_ids = kv_state->out_ids();
+        auto & out_ids = mstate->out_ids();
 
         GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
 
@@ -1189,11 +1225,6 @@ int llama_context::decode(llama_batch & inp_batch) {
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
-    // decide if we need to defrag the kv cache
-    if (cparams.defrag_thold > 0.0f) {
-        kv_self->defrag_sched(cparams.defrag_thold);
-    }
-
     // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
     // overlap with device computation.
     ggml_backend_sched_reset(sched.get());
@@ -1810,11 +1841,9 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-    if (kv_self != nullptr) {
+    if (memory != nullptr) {
         LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
-        kv_self->state_write(io);
+        memory->state_write(io);
     }
 
     return io.n_bytes();
@@ -1901,9 +1930,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     if (memory) {
         LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
 
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_read(io);
+        memory->state_read(io);
     }
 
     return io.n_bytes();
@@ -1913,9 +1940,7 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_write(io, seq_id);
+        memory->state_write(io, seq_id);
     }
 
     return io.n_bytes();
@@ -1925,9 +1950,7 @@ size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_read(io, seq_id);
+        memory->state_read(io, seq_id);
     }
 
     return io.n_bytes();
@@ -2032,9 +2055,7 @@ void llama_context::opt_epoch_iter(
     const uint32_t n_batch  = std::min(this->n_batch(),  n_ctx);
     const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-    kv_self->clear();
+    memory->clear(true);
 
     for (uint32_t pos_ctx = 0; pos_ctx < n_ctx; pos_ctx += n_batch) {
         batch.n_tokens = n_batch;
@@ -2057,8 +2078,8 @@ void llama_context::opt_epoch_iter(
 
         int64_t n_outputs_all = n_tokens_all;
 
-        auto kv_state = kv_self->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ true);
-        if (!kv_state || kv_state->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
+        auto mstate = memory->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ true);
+        if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
         }
@@ -2071,17 +2092,17 @@ void llama_context::opt_epoch_iter(
 
         uint32_t pos_batch = 0;
         do {
-            const auto & ubatch = kv_state->get_ubatch();
+            const auto & ubatch = mstate->get_ubatch();
 
             n_outputs = ubatch.n_tokens;
 
-            if (!kv_state->apply()) {
+            if (!mstate->apply()) {
                 LLAMA_LOG_ERROR("%s: failed to update the memory state\n", __func__);
                 break;
             }
 
             auto * gf = graph_init();
-            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, kv_state.get());
+            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate.get());
 
             struct ggml_context * ctx_compute_opt;
             {
@@ -2116,7 +2137,7 @@ void llama_context::opt_epoch_iter(
             ggml_free(ctx_compute_opt);
 
             pos_batch += ubatch.n_tokens;
-        } while (kv_state->next());
+        } while (mstate->next());
     }
 }
 
@@ -2277,13 +2298,14 @@ const llama_model * llama_get_model(const llama_context * ctx) {
     return &ctx->get_model();
 }
 
+// deprecated
 llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
-    return ctx->get_kv_self();
+    return dynamic_cast<llama_kv_cache *>(ctx->get_memory());
 }
 
 // deprecated
 void llama_kv_self_update(llama_context * ctx) {
-    ctx->kv_self_update();
+    ctx->kv_self_update(false);
 }
 
 enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
@@ -2398,13 +2420,118 @@ int32_t llama_apply_adapter_cvec(
     return res ? 0 : -1;
 }
 
+//
+// memory
+//
+
+llama_memory_t llama_get_memory(const struct llama_context * ctx) {
+    return ctx->get_memory();
+}
+
+void llama_memory_clear(llama_memory_t mem, bool data) {
+    if (!mem) {
+        return;
+    }
+
+    mem->clear(data);
+}
+
+bool llama_memory_seq_rm(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1) {
+    if (!mem) {
+        return true;
+    }
+
+    return mem->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_seq_cp(
+        llama_memory_t mem,
+          llama_seq_id seq_id_src,
+          llama_seq_id seq_id_dst,
+             llama_pos p0,
+             llama_pos p1) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_seq_keep(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_keep(seq_id);
+}
+
+void llama_memory_seq_add(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1,
+             llama_pos delta) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_add(seq_id, p0, p1, delta);
+}
+
+void llama_memory_seq_div(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1,
+                   int d) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_seq_pos_min(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return -1;
+    }
+
+    return mem->seq_pos_min(seq_id);
+}
+
+llama_pos llama_memory_seq_pos_max(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return -1;
+    }
+
+    return mem->seq_pos_max(seq_id);
+}
+
+bool llama_memory_can_shift(llama_memory_t mem) {
+    if (!mem) {
+        return false;
+    }
+
+    return mem->get_can_shift();
+}
+
 //
 // kv cache
 //
 
 // deprecated
 int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    const auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return 0;
     }
@@ -2426,7 +2553,7 @@ int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
 // deprecated
 // note: this is the same as above - will be removed anyway, so it's ok
 int32_t llama_kv_self_used_cells(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    const auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return 0;
     }
@@ -2445,115 +2572,119 @@ int32_t llama_kv_self_used_cells(const llama_context * ctx) {
     return res;
 }
 
+// deprecated
 void llama_kv_self_clear(llama_context * ctx) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->clear();
+    llama_memory_clear(kv, true);
 }
 
+// deprecated
 bool llama_kv_self_seq_rm(
         llama_context * ctx,
          llama_seq_id   seq_id,
             llama_pos   p0,
             llama_pos   p1) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return true;
     }
 
-    return kv->seq_rm(seq_id, p0, p1);
+    return llama_memory_seq_rm(kv, seq_id, p0, p1);
 }
 
+// deprecated
 void llama_kv_self_seq_cp(
         llama_context * ctx,
          llama_seq_id   seq_id_src,
          llama_seq_id   seq_id_dst,
             llama_pos   p0,
             llama_pos   p1) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    llama_memory_seq_cp(kv, seq_id_src, seq_id_dst, p0, p1);
 }
 
+// deprecated
 void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_keep(seq_id);
+    llama_memory_seq_keep(kv, seq_id);
 }
 
+// deprecated
 void llama_kv_self_seq_add(
         llama_context * ctx,
          llama_seq_id   seq_id,
             llama_pos   p0,
             llama_pos   p1,
             llama_pos   delta) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_add(seq_id, p0, p1, delta);
+    llama_memory_seq_add(kv, seq_id, p0, p1, delta);
 }
 
+// deprecated
 void llama_kv_self_seq_div(
         llama_context * ctx,
          llama_seq_id   seq_id,
             llama_pos   p0,
             llama_pos   p1,
                   int   d) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_div(seq_id, p0, p1, d);
+    llama_memory_seq_div(kv, seq_id, p0, p1, d);
 }
 
+// deprecated
 llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return -1;
     }
 
-    return kv->seq_pos_min(seq_id);
+    return llama_memory_seq_pos_min(kv, seq_id);
 }
 
+// deprecated
 llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return -1;
     }
 
-    return kv->seq_pos_max(seq_id);
+    return llama_memory_seq_pos_max(kv, seq_id);
 }
 
 // deprecated
 void llama_kv_self_defrag(llama_context * ctx) {
-    auto * kv = ctx->get_kv_self();
-    if (!kv) {
-        return;
-    }
-
     // force defrag
-    kv->defrag_sched(-1.0f);
+    ctx->kv_self_defrag_sched();
 }
 
+// deprecated
 bool llama_kv_self_can_shift(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return false;
     }
 
-    return kv->get_can_shift();
+    return llama_memory_can_shift(kv);
 }
 
 // llama state API

src/llama-context.h

@@ -13,13 +13,12 @@
 #include <vector>
 
 struct llama_model;
-struct llama_kv_cache;
 
 class llama_io_read_i;
 class llama_io_write_i;
 
-class llama_memory_i;
-class llama_memory_state_i;
+struct llama_memory_i;
+struct llama_memory_state_i;
 
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
@@ -47,12 +46,12 @@ struct llama_context {
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;
 
-          llama_kv_cache * get_kv_self();
-    const llama_kv_cache * get_kv_self() const;
+    llama_memory_t get_memory() const;
 
     // return true of the KV cache was updated
     // TODO: remove
-    bool kv_self_update();
+    bool kv_self_update(bool optimize);
+    void kv_self_defrag_sched();
 
     enum llama_pooling_type pooling_type() const;
 
@@ -231,6 +230,9 @@ private:
 
     std::unique_ptr<llama_memory_i> memory;
 
+    // TODO: temporary, until the llama_kv_self_defrag() API is removed
+    bool memory_force_optimize = false;
+
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;

src/llama-graph.cpp

@@ -659,6 +659,28 @@ ggml_tensor * llm_graph_context::build_ffn(
                 cur = ggml_mul(ctx0, x0, x1);
                 cb(cur, "ffn_mul", il);
             } break;
+        case LLM_FFN_GEGLU:
+            {
+                // Split into two equal parts
+                int64_t split_point = cur->ne[0] / 2;
+                ggml_tensor * output_ffn_up = ggml_cont(ctx0, ggml_view_2d(
+                                                ctx0, cur, split_point,
+                                                cur->ne[1], cur->nb[1], 0
+                                            ));
+                ggml_tensor * output_ffn_gate = ggml_cont(ctx0, ggml_view_2d(
+                                                ctx0, cur, split_point,
+                                                cur->ne[1], cur->nb[1],
+                                                split_point * ggml_element_size(cur)
+                                            ));
+
+                // Apply GELU activation function to the first part
+                output_ffn_up = ggml_gelu(ctx0, output_ffn_up);
+                cb(output_ffn_up, "ffn_gelu", il);
+
+                // Element-wise multiplication between the activated part and the gate part
+                cur = ggml_mul(ctx0, output_ffn_up, output_ffn_gate);
+                cb(cur, "ffn_geglu", il);
+            } break;
     }
 
     if (gate && type_gate == LLM_FFN_PAR) {
@@ -769,9 +791,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
 
     if (weight_before_ffn) {
-        // TODO: this is a workaround as we don't yet have a repeat op that takes custom dim (ggml_repeat_4d)
-        ggml_tensor * repeated = ggml_new_tensor_3d(ctx0, cur->type, n_embd, n_expert_used, n_tokens);
-        repeated = ggml_repeat(ctx0, cur, repeated); // [n_embd, n_expert_used, n_tokens]
+        // repeat cur to [n_embd, n_expert_used, n_tokens]
+        ggml_tensor * repeated = ggml_repeat_4d(ctx0, cur, n_embd, n_expert_used, n_tokens, 1);
         cur = ggml_mul(ctx0, repeated, weights);
         cb(cur, "ffn_moe_weighted", il);
     }

src/llama-graph.h

@@ -17,7 +17,7 @@ struct ggml_tensor;
 struct llama_ubatch;
 struct llama_cparams;
 
-class llama_memory_state_i;
+struct llama_memory_state_i;
 
 class llama_kv_cache_unified_state;
 class llama_kv_cache_unified_iswa_state;
@@ -36,6 +36,7 @@ enum llm_ffn_op_type {
     LLM_FFN_RELU,
     LLM_FFN_RELU_SQR,
     LLM_FFN_SWIGLU,
+    LLM_FFN_GEGLU,
 };
 
 enum llm_ffn_gate_type {

src/llama-kv-cache-recurrent.cpp

@@ -1,6 +1,7 @@
 #include "llama-kv-cache-recurrent.h"
 
 #include "llama-impl.h"
+#include "llama-io.h"
 #include "llama-batch.h"
 #include "llama-model.h"
 
@@ -116,18 +117,21 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
     }
 }
 
-void llama_kv_cache_recurrent::clear() {
+void llama_kv_cache_recurrent::clear(bool data) {
     for (int32_t i = 0; i < (int32_t) size; ++i) {
         cells[i].pos = -1;
         cells[i].seq_id.clear();
         cells[i].src = -1;
         cells[i].tail = -1;
     }
+
     head = 0;
     used = 0;
 
-    for (auto & buf : bufs) {
-        ggml_backend_buffer_clear(buf.get(), 0);
+    if (data) {
+        for (auto & buf : bufs) {
+            ggml_backend_buffer_clear(buf.get(), 0);
+        }
     }
 }
 
@@ -386,6 +390,13 @@ llama_memory_state_ptr llama_kv_cache_recurrent::init_full() {
     return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
 }
 
+llama_memory_state_ptr llama_kv_cache_recurrent::init_update(llama_context * lctx, bool optimize) {
+    GGML_UNUSED(lctx);
+    GGML_UNUSED(optimize);
+
+    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_NO_UPDATE);
+}
+
 bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
     // simply remember the full state because it is very small for this type of cache
     // TODO: optimize
@@ -419,17 +430,6 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
     return success;
 }
 
-bool llama_kv_cache_recurrent::update(llama_context & lctx) {
-    GGML_UNUSED(lctx);
-    // noop
-    return false;
-}
-
-void llama_kv_cache_recurrent::defrag_sched(float thold) {
-    GGML_UNUSED(thold);
-    // noop
-}
-
 bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
     const uint32_t n_tokens = ubatch.n_tokens;
     const uint32_t n_seqs   = ubatch.n_seqs;
@@ -726,7 +726,7 @@ void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq
 
     if (!res) {
         if (seq_id == -1) {
-            clear();
+            clear(true);
         } else {
             seq_rm(seq_id, -1, -1);
         }
@@ -883,7 +883,7 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
             return false;
         }
 
-        clear();
+        clear(true);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             kv_cell & cell = cells[i];

src/llama-kv-cache-recurrent.h

@@ -2,7 +2,7 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache.h"
+#include "llama-memory.h"
 
 #include <set>
 #include <vector>
@@ -13,7 +13,7 @@
 
 // TODO: extract the KV cache state used for graph computation into llama_kv_cache_recurrent_state_i
 //       see the implementation of llama_kv_cache_unified_state_i for an example how to do it
-class llama_kv_cache_recurrent : public llama_kv_cache {
+class llama_kv_cache_recurrent : public llama_memory_i {
 public:
     llama_kv_cache_recurrent(
             const llama_model & model,
@@ -29,7 +29,17 @@ public:
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_pooled,
+            bool logits_all) override;
+
+    llama_memory_state_ptr init_full() override;
+
+    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    void clear(bool data) override;
 
     bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
     void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
@@ -40,22 +50,6 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
     bool prepare(const std::vector<llama_ubatch> & ubatches);
 
     // find a contiguous slot of kv cells and emplace the ubatch there

src/llama-kv-cache-unified-iswa.cpp

@@ -52,9 +52,9 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
             hparams.n_swa, hparams.swa_type);
 }
 
-void llama_kv_cache_unified_iswa::clear() {
-    kv_base->clear();
-    kv_swa ->clear();
+void llama_kv_cache_unified_iswa::clear(bool data) {
+    kv_base->clear(data);
+    kv_swa ->clear(data);
 }
 
 bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
@@ -123,26 +123,16 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch
 
     assert(heads_base.size() == heads_swa.size());
 
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_SUCCESS,
+    return std::make_unique<llama_kv_cache_unified_iswa_state>(
             this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
 }
 
 llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() {
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+    return std::make_unique<llama_kv_cache_unified_iswa_state>(this);
 }
 
-bool llama_kv_cache_unified_iswa::update(llama_context & lctx) {
-    bool res = false;
-
-    res = res | kv_base->update(lctx);
-    res = res | kv_swa ->update(lctx);
-
-    return res;
-}
-
-void llama_kv_cache_unified_iswa::defrag_sched(float thold) {
-    kv_base->defrag_sched(thold);
-    kv_swa ->defrag_sched(thold);
+llama_memory_state_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_unified_iswa_state>(this, lctx, optimize);
 }
 
 bool llama_kv_cache_unified_iswa::get_can_shift() const {
@@ -174,26 +164,38 @@ llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(llama_memory_status status) : status(status) {}
 
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
-        llama_memory_status status,
-        llama_kv_cache_unified_iswa * kv) : status(status) {
-    state_base.reset(new llama_kv_cache_unified_state(status, kv->get_base()));
-    state_swa .reset(new llama_kv_cache_unified_state(status, kv->get_swa ()));
+        llama_kv_cache_unified_iswa * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS) {
+    state_base = kv->get_base()->init_full();
+    state_swa  = kv->get_swa ()->init_full();
+
+    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
+}
+
+llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
+        llama_kv_cache_unified_iswa * kv,
+        llama_context * lctx,
+        bool optimize) : status(LLAMA_MEMORY_STATUS_SUCCESS) {
+    state_base = kv->get_base()->init_update(lctx, optimize);
+    state_swa  = kv->get_swa ()->init_update(lctx, optimize);
+
+    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
 }
 
 llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
-        llama_memory_status status,
         llama_kv_cache_unified_iswa * kv,
         llama_sbatch sbatch,
         std::vector<uint32_t> heads_base,
         std::vector<uint32_t> heads_swa,
         std::vector<llama_ubatch> ubatches)
-    : status(status),
-    sbatch(std::move(sbatch)),
-    ubatches(std::move(ubatches)) {
-        // note: here we copy the ubatches. not sure if this is ideal
-        state_base.reset(new llama_kv_cache_unified_state(status, kv->get_base(), {}, std::move(heads_base), this->ubatches));
-        state_swa .reset(new llama_kv_cache_unified_state(status, kv->get_swa (), {}, std::move(heads_swa),  this->ubatches));
-    }
+        : status(LLAMA_MEMORY_STATUS_SUCCESS),
+        sbatch(std::move(sbatch)),
+        ubatches(std::move(ubatches)) {
+    // note: here we copy the ubatches. not sure if this is ideal
+    state_base.reset(new llama_kv_cache_unified_state(kv->get_base(), {}, std::move(heads_base), this->ubatches));
+    state_swa .reset(new llama_kv_cache_unified_state(kv->get_swa (), {}, std::move(heads_swa),  this->ubatches));
+
+    status = llama_memory_status_combine(state_base->get_status(), state_swa->get_status());
+}
 
 llama_kv_cache_unified_iswa_state:: ~llama_kv_cache_unified_iswa_state() = default;
 
@@ -233,17 +235,18 @@ llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const {
 
 const llama_ubatch & llama_kv_cache_unified_iswa_state::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
     return ubatches[i_next];
 }
 
 const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_base() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return state_base.get();
+    return static_cast<const llama_kv_cache_unified_state *>(state_base.get());
 }
 
 const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_swa()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return state_swa.get();
+    return static_cast<const llama_kv_cache_unified_state *>(state_swa.get());
 }

src/llama-kv-cache-unified-iswa.h

@@ -11,7 +11,7 @@
 // utilizes two instances of llama_kv_cache_unified
 //   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
 
-class llama_kv_cache_unified_iswa : public llama_kv_cache {
+class llama_kv_cache_unified_iswa : public llama_memory_i {
 public:
     llama_kv_cache_unified_iswa(
             const llama_model & model,
@@ -31,7 +31,19 @@ public:
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_pooled,
+            bool logits_all) override;
+
+    llama_memory_state_ptr init_full() override;
+
+    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
 
     bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
     void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
@@ -42,24 +54,6 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
-    bool get_can_shift() const override;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -86,12 +80,16 @@ public:
 
     // used to create a full-cache state
     llama_kv_cache_unified_iswa_state(
-            llama_memory_status status,
             llama_kv_cache_unified_iswa * kv);
 
+    // used to create an update state
+    llama_kv_cache_unified_iswa_state(
+            llama_kv_cache_unified_iswa * kv,
+            llama_context * lctx,
+            bool optimize);
+
     // used to create a state from a batch
     llama_kv_cache_unified_iswa_state(
-            llama_memory_status status,
             llama_kv_cache_unified_iswa * kv,
             llama_sbatch sbatch,
             std::vector<uint32_t> heads_base,
@@ -120,7 +118,7 @@ public:
     const llama_kv_cache_unified_state * get_swa()  const;
 
 private:
-    const llama_memory_status status;
+    llama_memory_status status;
 
     //llama_kv_cache_unified_iswa * kv;
 
@@ -131,6 +129,6 @@ private:
 
     std::vector<llama_ubatch> ubatches;
 
-    std::unique_ptr<llama_kv_cache_unified_state> state_base;
-    std::unique_ptr<llama_kv_cache_unified_state> state_swa;
+    llama_memory_state_ptr state_base;
+    llama_memory_state_ptr state_swa;
 };

src/llama-kv-cache-unified.cpp

@@ -1,6 +1,7 @@
 #include "llama-kv-cache-unified.h"
 
 #include "llama-impl.h"
+#include "llama-io.h"
 #include "llama-model.h"
 #include "llama-context.h"
 
@@ -128,13 +129,15 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     }
 }
 
-void llama_kv_cache_unified::clear() {
+void llama_kv_cache_unified::clear(bool data) {
     cells.reset();
 
     head = 0;
 
-    for (auto & buf : bufs) {
-        ggml_backend_buffer_clear(buf.get(), 0);
+    if (data) {
+        for (auto & buf : bufs) {
+            ggml_backend_buffer_clear(buf.get(), 0);
+        }
     }
 }
 
@@ -149,12 +152,27 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    for (uint32_t i = 0; i < cells.size(); ++i) {
-        if (!cells.pos_in(i, p0, p1)) {
-            continue;
-        }
+    if (seq_id >= 0) {
+        for (uint32_t i = 0; i < cells.size(); ++i) {
+            if (!cells.pos_in(i, p0, p1)) {
+                continue;
+            }
+
+            if (cells.seq_has(i, seq_id) && cells.seq_rm(i, seq_id)) {
+                if (new_head == cells.size()) {
+                    new_head = i;
+                }
+            }
+        }
+    } else {
+        // match any sequence
+        for (uint32_t i = 0; i < cells.size(); ++i) {
+            if (!cells.pos_in(i, p0, p1)) {
+                continue;
+            }
+
+            cells.rm(i);
 
-        if (cells.seq_has(i, seq_id) && cells.seq_rm(i, seq_id)) {
             if (new_head == cells.size()) {
                 new_head = i;
             }
@@ -305,16 +323,49 @@ llama_memory_state_ptr llama_kv_cache_unified::init_batch(
         return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
     }
 
-    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_SUCCESS,
+    return std::make_unique<llama_kv_cache_unified_state>(
             this, std::move(sbatch), std::move(heads), std::move(ubatches));
 }
 
 llama_memory_state_ptr llama_kv_cache_unified::init_full() {
-    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+    return std::make_unique<llama_kv_cache_unified_state>(this);
 }
 
-std::vector<uint32_t> llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
-    std::vector<uint32_t> res;
+llama_memory_state_ptr llama_kv_cache_unified::init_update(llama_context * lctx, bool optimize) {
+    bool do_shift = get_has_shift();
+
+    defrag_info dinfo;
+
+    // see if we need to defrag
+    {
+        bool do_defrag = optimize;
+
+        const auto thold = lctx->get_cparams().defrag_thold;
+
+        if (!do_defrag && thold > 0.0f) {
+            const auto n_kv = cells.used_max_p1();
+
+            // - do not defrag small contexts (i.e. < 2048 tokens)
+            // - count the padding towards the number of used tokens
+            const float fragmentation = n_kv >= 2048 ? std::max(0.0f, 1.0f - (float(cells.get_used() + n_pad)/n_kv)) : 0.0f;
+
+            if (fragmentation > thold) {
+                LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
+
+                do_defrag = true;
+            }
+        }
+
+        if (do_defrag) {
+            dinfo = defrag_prepare(lctx->graph_max_nodes());
+        }
+    }
+
+    return std::make_unique<llama_kv_cache_unified_state>(this, lctx, do_shift, std::move(dinfo));
+}
+
+llama_kv_cache_unified::ubatch_heads llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
+    llama_kv_cache_unified::ubatch_heads res;
 
     struct state {
         uint32_t head_old; // old position of the head, before placing the ubatch
@@ -359,12 +410,12 @@ std::vector<uint32_t> llama_kv_cache_unified::prepare(const std::vector<llama_ub
     return res;
 }
 
-bool llama_kv_cache_unified::update(llama_context & lctx) {
+bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const defrag_info & dinfo) {
     bool updated = false;
 
-    auto * sched = lctx.get_sched();
+    auto * sched = lctx->get_sched();
 
-    if (cells.get_has_shift()) {
+    if (do_shift) {
         if (!get_can_shift()) {
             GGML_ABORT("The current KV cache / model configuration does not support K-shift");
         }
@@ -375,9 +426,9 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
         if (hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
             ggml_backend_sched_reset(sched);
 
-            auto * gf = lctx.graph_init();
+            auto * gf = lctx->graph_init();
 
-            auto res = build_graph_shift(lctx.get_cparams(), lctx.get_ctx_compute(), gf);
+            auto res = build_graph_shift(lctx->get_cparams(), lctx->get_ctx_compute(), gf);
             if (!res) {
                 LLAMA_LOG_ERROR("%s: failed to build graph for K-shift\n", __func__);
                 return updated;
@@ -390,7 +441,7 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
 
             res->set_inputs(nullptr);
 
-            if (lctx.graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
+            if (lctx->graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
                 LLAMA_LOG_ERROR("%s: failed to compute K-shift\n", __func__);
                 return updated;
             }
@@ -401,56 +452,55 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
         cells.reset_shift();
     }
 
-    if (do_defrag) {
+    if (!dinfo.empty()) {
         LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
 
-        if (defrag_prepare(lctx.graph_max_nodes())) {
-            ggml_backend_sched_reset(sched);
+        // apply moves:
+        {
+            const auto n_kv = dinfo.ids.size();
 
-            auto * gf = lctx.graph_init();
+            for (uint32_t i = 0; i < n_kv; ++i) {
+                assert(dinfo.ids[i] <= n_kv);
 
-            auto res = build_graph_defrag(lctx.get_cparams(), lctx.get_ctx_compute(), gf);
-            if (!res) {
-                LLAMA_LOG_ERROR("%s: failed to build graph for defrag\n", __func__);
-                return updated;
+                if (dinfo.ids[i] == n_kv) {
+                    continue;
+                }
+
+                cells.mv(i, dinfo.ids[i]);
             }
 
-            if (!ggml_backend_sched_alloc_graph(sched, gf)) {
-                LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
-                return updated;
-            }
-
-            res->set_inputs(nullptr);
-
-            if (lctx.graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
-                LLAMA_LOG_ERROR("%s: failed to compute defrag\n", __func__);
-                return updated;
-            }
-
-            updated = true;
+            // reset the head so we can find the first free slot during the next ubatch
+            head = 0;
         }
 
-        do_defrag = false;
+        ggml_backend_sched_reset(sched);
+
+        auto * gf = lctx->graph_init();
+
+        auto res = build_graph_defrag(lctx->get_cparams(), lctx->get_ctx_compute(), gf, dinfo);
+        if (!res) {
+            LLAMA_LOG_ERROR("%s: failed to build graph for defrag\n", __func__);
+            return updated;
+        }
+
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
+            return updated;
+        }
+
+        res->set_inputs(nullptr);
+
+        if (lctx->graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
+            LLAMA_LOG_ERROR("%s: failed to compute defrag\n", __func__);
+            return updated;
+        }
+
+        updated = true;
     }
 
     return updated;
 }
 
-void llama_kv_cache_unified::defrag_sched(float thold) {
-    const auto n_kv = cells.used_max_p1();
-
-    // - do not defrag small contexts (i.e. < 2048 tokens)
-    // - count the padding towards the number of used tokens
-    const float fragmentation = n_kv >= 2048 ? std::max(0.0f, 1.0f - (float(cells.get_used() + n_pad)/n_kv)) : 0.0f;
-
-    // queue defragmentation for next llama_kv_cache_update
-    if (fragmentation > thold) {
-        LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
-
-        do_defrag = true;
-    }
-}
-
 int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
@@ -597,6 +647,10 @@ uint32_t llama_kv_cache_unified::get_size() const {
     return cells.size();
 }
 
+bool llama_kv_cache_unified::get_has_shift() const {
+    return cells.get_has_shift();
+}
+
 uint32_t llama_kv_cache_unified::get_n_kv() const {
     return std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad)));
 }
@@ -926,12 +980,13 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
 }
 
 llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-        const llama_cparams & cparams,
-               ggml_context * ctx,
-                ggml_cgraph * gf) const {
+                const llama_cparams & cparams,
+                       ggml_context * ctx,
+                        ggml_cgraph * gf,
+                  const defrag_info & dinfo) const {
     auto res = std::make_unique<llm_graph_result>();
 
-    const auto & ids = defrag_info.ids;
+    const auto & ids = dinfo.ids;
 
 #if 0
     // CPU defrag
@@ -1072,7 +1127,7 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
     return res;
 }
 
-bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+llama_kv_cache_unified::defrag_info llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) const {
     const uint32_t n_layer = layers.size();
 
     const uint32_t n_kv   = cells.used_max_p1();
@@ -1093,14 +1148,9 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
     const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer);
 
     // determine which KV cells to move where
-    //
-    //  cell i moves to ids[i]
-    //
-    //  if ids[i] == i || ids[i] == n_kv, then cell i is not moved
-    //
-    auto & ids = defrag_info.ids;
+    defrag_info res;
+    auto & ids = res.ids;
 
-    ids.clear();
     ids.resize(n_kv, n_kv);
 
     for (uint32_t i0 = 0; i0 < n_used; ++i0) {
@@ -1164,11 +1214,6 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
             // this cell goes to (i0 + nf)
             ids[i1] = i0 + nf;
 
-            // move the cell meta data
-            cells.mv(i1, i0 + nf);
-
-            head = n_used;
-
             if (!cont) {
                 n_moves++;
                 cont = true;
@@ -1191,14 +1236,14 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
     }
 
     if (n_moves == 0) {
-        return false;
+        return {};
     }
 
     LLAMA_LOG_DEBUG("%s: (tmp log) KV defrag cell moves: %u\n", __func__, n_moves);
 
     LLAMA_LOG_DEBUG("%s: expected gf nodes: %u\n", __func__, 6*n_moves*n_layer);
 
-    return true;
+    return res;
 }
 
 bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
@@ -1276,7 +1321,7 @@ void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_i
 
     if (!res) {
         if (seq_id == -1) {
-            clear();
+            clear(true);
         } else {
             seq_rm(seq_id, -1, -1);
         }
@@ -1457,7 +1502,7 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
             return false;
         }
 
-        clear();
+        clear(true);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -1621,24 +1666,27 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
 llama_kv_cache_unified_state::llama_kv_cache_unified_state(llama_memory_status status) : status(status) {}
 
 llama_kv_cache_unified_state::llama_kv_cache_unified_state(
-            llama_memory_status status,
-            llama_kv_cache_unified * kv) : status(status), kv(kv) {
-        n_kv = kv->get_size();
-        head = 0;
-    }
+        llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
+    n_kv = kv->get_size();
+    head = 0;
+}
 
 llama_kv_cache_unified_state::llama_kv_cache_unified_state(
-            llama_memory_status status,
-            llama_kv_cache_unified * kv,
-            llama_sbatch sbatch,
-            std::vector<uint32_t> heads,
-            std::vector<llama_ubatch> ubatches)
-            : status(status),
-              kv(kv),
-              sbatch(std::move(sbatch)),
-              heads(std::move(heads)),
-              ubatches(std::move(ubatches)) {
+        llama_kv_cache_unified * kv,
+        llama_context * lctx,
+        bool do_shift,
+        defrag_info dinfo) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)) {
+    if (!do_shift && dinfo.empty()) {
+        status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
+}
+
+llama_kv_cache_unified_state::llama_kv_cache_unified_state(
+        llama_kv_cache_unified * kv,
+        llama_sbatch sbatch,
+        llama_kv_cache_unified::ubatch_heads heads,
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sbatch(std::move(sbatch)), heads(std::move(heads)), ubatches(std::move(ubatches)) {
+}
 
 llama_kv_cache_unified_state::~llama_kv_cache_unified_state() = default;
 
@@ -1655,6 +1703,13 @@ bool llama_kv_cache_unified_state::next() {
 bool llama_kv_cache_unified_state::apply() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
+    // no ubatches -> this is a KV cache update
+    if (ubatches.empty()) {
+        kv->update(lctx, do_shift, dinfo);
+
+        return true;
+    }
+
     kv->apply_ubatch(heads[i_next], ubatches[i_next]);
 
     n_kv = kv->get_n_kv();

src/llama-kv-cache-unified.h

@@ -2,8 +2,8 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache.h"
 #include "llama-kv-cells.h"
+#include "llama-memory.h"
 
 #include <unordered_map>
 #include <vector>
@@ -17,13 +17,26 @@ struct llama_context;
 // llama_kv_cache_unified
 //
 
-class llama_kv_cache_unified : public llama_kv_cache {
+class llama_kv_cache_unified : public llama_memory_i {
 public:
     static uint32_t get_padding(const llama_cparams & cparams);
 
     // this callback is used to filter out layers that should not be included in the cache
     using layer_filter_cb = std::function<bool(int32_t il)>;
 
+    using ubatch_heads = std::vector<uint32_t>;
+
+    struct defrag_info {
+        bool empty() const {
+            return ids.empty();
+        }
+
+        // contains information about which cell moves where:
+        //  - cell i moves to ids[i]
+        //  - if ids[i] == i || ids[i] == ids.size(), then cell i is not moved
+        std::vector<uint32_t> ids;
+    };
+
     llama_kv_cache_unified(
             const llama_model &  model,
               layer_filter_cb && filter,
@@ -43,7 +56,19 @@ public:
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_pooled,
+            bool logits_all) override;
+
+    llama_memory_state_ptr init_full() override;
+
+    llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
 
     bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
     void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
@@ -54,24 +79,6 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
-    bool get_can_shift() const override;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -83,6 +90,8 @@ public:
 
     uint32_t get_size() const;
 
+    bool get_has_shift() const;
+
     //
     // graph_build API
     //
@@ -103,7 +112,9 @@ public:
 
     // find places for the provided ubatches in the cache, returns the head locations
     // return empty vector on failure
-    std::vector<uint32_t> prepare(const std::vector<llama_ubatch> & ubatches);
+    ubatch_heads prepare(const std::vector<llama_ubatch> & ubatches);
+
+    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
 
     // return the cell position where we can insert the ubatch
     // return -1 on failure to find a contiguous slot of kv cells
@@ -133,8 +144,7 @@ private:
         ggml_tensor * v;
     };
 
-    bool do_defrag = false;
-    bool v_trans   = true;  // the value tensor is transposed
+    bool v_trans = true;  // the value tensor is transposed
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
@@ -160,13 +170,8 @@ private:
     // model layer id -> KV cache layer id
     std::unordered_map<int32_t, int32_t> map_layer_ids;
 
-    // defrag
-    struct {
-        std::vector<uint32_t> ids;
-    } defrag_info;
-
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
+    // return non-empty vector if cells have been moved
+    defrag_info defrag_prepare(int32_t n_max_nodes) const;
 
     size_t total_size() const;
 
@@ -192,7 +197,8 @@ private:
     llm_graph_result_ptr build_graph_defrag(
             const llama_cparams & cparams,
                    ggml_context * ctx,
-                    ggml_cgraph * gf) const;
+                    ggml_cgraph * gf,
+              const defrag_info & dinfo) const;
 
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
@@ -203,20 +209,29 @@ private:
 
 class llama_kv_cache_unified_state : public llama_memory_state_i {
 public:
+    // some shorthands
+    using ubatch_heads = llama_kv_cache_unified::ubatch_heads;
+    using defrag_info  = llama_kv_cache_unified::defrag_info;
+
     // used for errors
     llama_kv_cache_unified_state(llama_memory_status status);
 
     // used to create a full-cache state
     llama_kv_cache_unified_state(
-            llama_memory_status status,
             llama_kv_cache_unified * kv);
 
-    // used to create a state from a batch
+    // used to create an update state
+    llama_kv_cache_unified_state(
+            llama_kv_cache_unified * kv,
+            llama_context * lctx,
+            bool do_shift,
+            defrag_info dinfo);
+
+    // used to create a decode state from a batch
     llama_kv_cache_unified_state(
-            llama_memory_status status,
             llama_kv_cache_unified * kv,
             llama_sbatch sbatch,
-            std::vector<uint32_t> heads,
+            ubatch_heads heads,
             std::vector<llama_ubatch> ubatches);
 
     virtual ~llama_kv_cache_unified_state();
@@ -253,16 +268,30 @@ public:
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
 private:
-    const llama_memory_status status;
+    llama_memory_status status;
 
     llama_kv_cache_unified * kv;
+    llama_context * lctx;
+
+    //
+    // update state
+    //
+
+    bool do_shift = false;
+
+    defrag_info dinfo;
+
+    //
+    // batch processing state
+    //
 
     llama_sbatch sbatch;
 
     // the index of the next ubatch to process
     size_t i_next = 0;
 
-    std::vector<uint32_t> heads;
+    ubatch_heads heads;
+
     std::vector<llama_ubatch> ubatches;
 
     //

src/llama-kv-cache.cpp

@@ -1 +0,0 @@
-#include "llama-kv-cache.h"

src/llama-kv-cache.h

@@ -1,44 +0,0 @@
-#pragma once
-
-#include "llama.h"
-#include "llama-io.h"
-#include "llama-memory.h"
-
-struct llama_kv_cache : public llama_memory_i {
-    virtual ~llama_kv_cache() = default;
-
-    // split the input batch into a set of ubatches and verify that they can fit into the cache
-    // return a state object containing the ubatches and KV cache state required to process them
-    // check the llama_memory_state_i::get_status() for the result
-    virtual llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) = 0;
-
-    // simulate full cache, used for allocating worst-case compute buffers
-    virtual llama_memory_state_ptr init_full() = 0;
-
-    // process any pending defrag/shift/etc. operations
-    // optionally call once before processing a new batch
-    // return true if any operations were performed
-    virtual bool update(llama_context & lctx) = 0;
-
-    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
-    // TODO: change to
-    //   llama_memory_state_ptr init_defrag(float thold) = 0;
-    //
-    virtual void defrag_sched(float thold) = 0;
-
-    // getters
-    virtual bool get_can_shift() const = 0;
-
-    bool get_can_edit() const override { return get_can_shift(); }
-
-    //
-    // state write/read
-    //
-
-    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
-    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
-};

src/llama-memory.cpp

@@ -1 +1,42 @@
 #include "llama-memory.h"
+
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1) {
+    bool has_update = false;
+
+    switch (s0) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s0;
+            }
+    }
+
+    switch (s1) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s1;
+            }
+    }
+
+    // if either status has an update, then the combined status has an update
+    return has_update ? LLAMA_MEMORY_STATUS_SUCCESS : LLAMA_MEMORY_STATUS_NO_UPDATE;
+}

src/llama-memory.h

@@ -7,6 +7,9 @@
 
 struct llama_ubatch;
 
+class llama_io_write_i;
+class llama_io_read_i;
+
 struct llama_memory_params {
     // kv cache
     ggml_type type_k;
@@ -16,32 +19,17 @@ struct llama_memory_params {
     bool swa_full;
 };
 
-// general concept of LLM memory
-// the KV cache is a type of LLM memory, but there can be other types
-class llama_memory_i {
-public:
-    virtual ~llama_memory_i() = default;
-
-    virtual void clear() = 0;
-
-    virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
-    virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;
-    virtual void seq_keep(llama_seq_id seq_id) = 0;
-    virtual void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) = 0;
-    virtual void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) = 0;
-
-    virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
-    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
-
-    virtual bool get_can_edit() const = 0;
-};
-
 enum llama_memory_status {
     LLAMA_MEMORY_STATUS_SUCCESS = 0,
+    LLAMA_MEMORY_STATUS_NO_UPDATE,
     LLAMA_MEMORY_STATUS_FAILED_PREPARE,
     LLAMA_MEMORY_STATUS_FAILED_COMPUTE,
 };
 
+// helper function for combining the status of two memory states
+// useful for implementing hybrid memory types (e.g. iSWA)
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1);
+
 // the interface for managing the memory state during batch processing
 // this interface is implemented per memory type. see:
 //   - llama_kv_cache_unified_state
@@ -51,8 +39,7 @@ enum llama_memory_status {
 // the only method that can mutate the memory and the memory state is llama_memory_i::apply()
 //
 // TODO: rename to llama_memory_context_i ?
-class llama_memory_state_i {
-public:
+struct llama_memory_state_i {
     virtual ~llama_memory_state_i() = default;
 
     // consume the current ubatch from the state and proceed to the next one
@@ -69,8 +56,63 @@ public:
     // get the current ubatch
     virtual const llama_ubatch & get_ubatch() const = 0;
 
-    // get the status of the memory state
+    // get the status of the memory state - used for error handling and checking if any updates would be applied
     virtual llama_memory_status get_status() const = 0;
 };
 
 using llama_memory_state_ptr = std::unique_ptr<llama_memory_state_i>;
+
+// general concept of LLM memory
+// the KV cache is a type of LLM memory, but there can be other types
+struct llama_memory_i {
+    virtual ~llama_memory_i() = default;
+
+    // split the input batch into a set of ubatches and verify that they can fit into the cache
+    // return a state object containing the ubatches and KV cache state required to process them
+    // check the llama_memory_state_i::get_status() for the result
+    virtual llama_memory_state_ptr init_batch(
+            const llama_batch & batch,
+            uint32_t n_ubatch,
+            bool embd_pooled,
+            bool logits_all) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual llama_memory_state_ptr init_full() = 0;
+
+    // prepare for any pending memory updates, such as shifts, defrags, etc.
+    // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
+    virtual llama_memory_state_ptr init_update(llama_context * lctx, bool optimize) = 0;
+
+    // getters
+    virtual bool get_can_shift() const = 0;
+
+    //
+    // ops
+    //
+
+    // if data == true, the data buffers will also be cleared together with the metadata
+    virtual void clear(bool data) = 0;
+
+    virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;
+    virtual void seq_keep(llama_seq_id seq_id) = 0;
+    virtual void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) = 0;
+    virtual void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) = 0;
+
+    virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
+    virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
+};
+
+using llama_memory_ptr = std::unique_ptr<llama_memory_i>;
+
+// TODO: temporary until the llama_kv_cache is removed from the public API
+struct llama_kv_cache : public llama_memory_i {
+    virtual ~llama_kv_cache() = default;
+};

src/llama-mmap.cpp

@@ -401,7 +401,7 @@ struct llama_mmap::impl {
                 }
             }
 #else
-            throw std::runtime_error("PrefetchVirtualMemory unavailable");
+            LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
 #endif
         }
     }

src/llama-model-loader.cpp

@@ -288,9 +288,10 @@ namespace GGUFMeta {
 
     template<typename T>
     bool llama_model_loader::get_arr(const std::string & key, std::vector<T> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -298,28 +299,40 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
-        result.resize(arr_info.length);
-        result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+            result.clear();
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result.emplace_back(value);
+            }
+        } else {
+            result.resize(arr_info.length);
+            result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        }
 
         return true;
     }
 
     template<typename T, size_t N_MAX>
     bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -327,22 +340,32 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {
             throw std::runtime_error(format("array length %u for key %s exceeds max %u", (uint32_t) arr_info.length, key.c_str(), (uint32_t) N_MAX));
         }
 
-        std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result[i] = value;
+            }
+        } else {
+            std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        }
 
         return true;
     }
@@ -352,6 +375,8 @@ namespace GGUFMeta {
         return get_arr(llm_kv(kid), result, required);
     }
 
+    template bool llama_model_loader::get_arr<std::vector<std::string>>(enum llm_kv kid, std::vector<std::string> & result, bool required);
+
     template<typename T>
     bool llama_model_loader::get_key(const std::string & key, T & result, bool required) {
         auto it = kv_overrides.find(key);

src/llama-model.cpp

@@ -543,6 +543,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     uint32_t n_vocab = 0;
     ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);
 
+    // for classifier models
+    ml.get_arr(LLM_KV_CLASSIFIER_OUTPUT_LABELS, classifier_labels, false);
+    if (!classifier_labels.empty()) {
+        hparams.n_cls_out = classifier_labels.size();
+    }
+
     // arch-specific KVs
     switch (arch) {
         case LLM_ARCH_LLAMA:
@@ -686,7 +692,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type, false);
-                ml.get_arr_n(LLM_KV_CLASSIFIER_OUTPUT_LABELS, hparams.n_cls_out, false);
 
                 switch (hparams.n_layer) {
                     case 3:
@@ -956,6 +961,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 46: type = LLM_TYPE_27B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                }
+
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
+                hparams.f_attention_scale = type == LLM_TYPE_27B
+                    ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
+                    : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
             } break;
         case LLM_ARCH_GEMMA3:
             {
@@ -976,6 +986,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
 
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
                 hparams.f_attention_scale = type == LLM_TYPE_27B
                     ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
                     : 1.0f / std::sqrt(float(hparams.n_embd_head_k));
@@ -4356,6 +4367,15 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
         LLAMA_LOG_INFO("%s: ssm_dt_rank      = %u\n",     __func__, hparams.ssm_dt_rank);
         LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms   = %d\n",     __func__, hparams.ssm_dt_b_c_rms);
+
+        if (!classifier_labels.empty()) {
+            LLAMA_LOG_INFO("%s: n_cls_out        = %u\n", __func__, hparams.n_cls_out);
+
+            size_t i = 0;
+            for (auto label : classifier_labels) {
+                LLAMA_LOG_INFO("%s: cls_label[%2zu]    = %s\n", __func__, i++, label.c_str());
+            }
+        }
     }
 
     LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, type_name().c_str());
@@ -8484,14 +8504,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
-                // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
-                switch (model.type) {
-                    case LLM_TYPE_2B:
-                    case LLM_TYPE_9B:
-                    case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); break;
-                    default: GGML_ABORT("fatal error");
-                };
-                cb(Qcur, "Qcur_scaled", il);
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
@@ -8632,9 +8645,12 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
             cur = build_norm(cur,
@@ -13600,6 +13616,18 @@ int32_t llama_model_n_swa(const llama_model * model) {
     return model->hparams.n_swa;
 }
 
+uint32_t llama_model_n_cls_out(const struct llama_model * model) {
+    return model->hparams.n_cls_out;
+}
+
+const char * llama_model_cls_label(const struct llama_model * model, uint32_t i) {
+    if (i < model->classifier_labels.size()) {
+        return model->classifier_labels[i].c_str();
+    }
+
+    return nullptr;
+}
+
 // deprecated
 int32_t llama_n_ctx_train(const llama_model * model) {
     return llama_model_n_ctx_train(model);
@@ -13760,7 +13788,7 @@ uint64_t llama_model_size(const llama_model * model) {
 }
 
 const char * llama_model_chat_template(const llama_model * model, const char * name) {
-    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N)
+    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE)
         : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
     const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {

src/llama-model.h

@@ -329,6 +329,9 @@ struct llama_model {
     llama_hparams hparams = {};
     llama_vocab   vocab;
 
+    // for classifier models
+    std::vector<std::string> classifier_labels;
+
     struct ggml_tensor * tok_embd   = nullptr;
     struct ggml_tensor * type_embd  = nullptr;
     struct ggml_tensor * pos_embd   = nullptr;

src/llama-vocab.cpp

@@ -2080,9 +2080,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
         std::string model_name;
         std::string tokenizer_pre;
+        std::string general_arch;
 
         ml.get_key(LLM_KV_GENERAL_NAME,  model_name,    false);
         ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false);
+        ml.get_key(LLM_KV_GENERAL_ARCHITECTURE, general_arch, false);
 
         // model name to lowercase
         std::transform(model_name.begin(), model_name.end(), model_name.begin(),
@@ -2091,9 +2093,16 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         );
 
-        // set attributes by model/tokenizer name
-        if (_contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})) {
-            _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+        // set attributes by model/tokenizer/architecture name
+        if (false
+                || _contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})
+                || _contains_any(general_arch, {"nomic-bert-moe"})
+           ) {
+            if (token_to_id.count("<mask>") == 0) {
+                LLAMA_LOG_WARN("%s: Mask token is missing in vocab, please reconvert model!\n", __func__);
+            } else {
+                _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+            }
         } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
             for (auto id : cache_special_tokens) {
                 _set_tokenid_attr(id, LLAMA_TOKEN_ATTR_RSTRIP, true);

tests/test-backend-ops.cpp

@@ -2706,8 +2706,8 @@ struct test_conv_transpose_1d : public test_case {
         return VARS_TO_STR5(ne_input, ne_kernel, s0, p0, d0);
     }
 
-    test_conv_transpose_1d(std::array<int64_t, 4> ne_input = {197, 32, 1, 1}, // [input_width, input_height, input_channels, 1]
-                           std::array<int64_t, 4> ne_kernel = {16, 32, 32, 1}, // [kernel_width, kernel_height, input_channels, 1]
+    test_conv_transpose_1d(std::array<int64_t, 4> ne_input = {197, 32, 1, 1}, // [input_width, input_channels, 1 /* assert in cpu kernel*/, 1 (should be batch)]
+                           std::array<int64_t, 4> ne_kernel = {16, 32, 32, 1}, // [kernel_width, output_channels, input_channels, 1 (should be batch)]
                            int s0 = 1, int p0 = 0, int d0 = 1)
         : ne_input(ne_input), ne_kernel(ne_kernel), s0(s0), p0(p0), d0(d0) {}
 
@@ -4029,6 +4029,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
 
+    for(uint32_t Cout : {1, 9}){
+        for(uint32_t Cin : {1, 7}){
+            for(uint32_t K : {1, 3, 1337}){
+                for(uint32_t L : {1, 2, 13}){
+                    for(uint32_t s0: {1, 2, 3}){
+                        test_cases.emplace_back(new test_conv_transpose_1d({L,Cin,1,1}, {K,Cout,Cin,1}, s0, 0, 1));
+                    }
+                }
+            }
+        }
+    }
+
     test_cases.emplace_back(new test_conv_transpose_1d());
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));

tests/test-chat.cpp

@@ -19,8 +19,8 @@
 using json = nlohmann::ordered_json;
 
 static std::ostream & operator<<(std::ostream & os, const common_chat_msg_diff & diff) {
-    // os << "reasoning_content_delta: " << diff.reasoning_content_delta << '\n';
     os << "{ content_delta: " << diff.content_delta << "; ";
+    os << "reasoning_content_delta: " << diff.reasoning_content_delta << "; ";
     if (diff.tool_call_index != std::string::npos) {
         os << "tool_call_index: " << diff.tool_call_index << "; ";
         os << "tool_call_delta.name: " << diff.tool_call_delta.name << "; ";

tests/test-gguf.cpp

@@ -16,6 +16,7 @@ constexpr int offset_has_data    = 3000;
 
 enum handcrafted_file_type {
     HANDCRAFTED_HEADER_BAD_MAGIC           =  10,
+    HANDCRAFTED_HEADER_BAD_VERSION_0       =  15,
     HANDCRAFTED_HEADER_BAD_VERSION_1       =  20,
     HANDCRAFTED_HEADER_BAD_VERSION_FUTURE  =  30,
     HANDCRAFTED_HEADER_BAD_N_TENSORS       =  40,
@@ -51,6 +52,7 @@ enum handcrafted_file_type {
 static std::string handcrafted_file_type_name(const enum handcrafted_file_type hft) {
     switch (hft) {
         case HANDCRAFTED_HEADER_BAD_MAGIC:           return "HEADER_BAD_MAGIC";
+        case HANDCRAFTED_HEADER_BAD_VERSION_0:       return "HEADER_BAD_VERSION_0";
         case HANDCRAFTED_HEADER_BAD_VERSION_1:       return "HEADER_BAD_VERSION_1";
         case HANDCRAFTED_HEADER_BAD_VERSION_FUTURE:  return "HEADER_BAD_VERSION_FUTURE";
         case HANDCRAFTED_HEADER_BAD_N_KV:            return "HEADER_BAD_N_KV";
@@ -171,7 +173,10 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         helper_write(file, GGUF_MAGIC, 4);
     }
 
-    if (hft == HANDCRAFTED_HEADER_BAD_VERSION_1) {
+    if (hft == HANDCRAFTED_HEADER_BAD_VERSION_0) {
+        const uint32_t version = 0;
+        helper_write(file, version);
+    } else if (hft == HANDCRAFTED_HEADER_BAD_VERSION_1) {
         const uint32_t version = 1;
         helper_write(file, version);
     } else if (hft == HANDCRAFTED_HEADER_BAD_VERSION_FUTURE) {
@@ -660,6 +665,7 @@ static std::pair<int, int> test_handcrafted_file(const unsigned int seed) {
 
     const std::vector<handcrafted_file_type> hfts = {
         HANDCRAFTED_HEADER_BAD_MAGIC,
+        HANDCRAFTED_HEADER_BAD_VERSION_0,
         HANDCRAFTED_HEADER_BAD_VERSION_1,
         HANDCRAFTED_HEADER_BAD_VERSION_FUTURE,
         HANDCRAFTED_HEADER_BAD_N_KV,

tools/batched-bench/batched-bench.cpp

@@ -57,6 +57,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto * mem = llama_get_memory(ctx);
+
     const int32_t n_kv_max = llama_n_ctx(ctx);
 
     llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
@@ -132,7 +134,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_self_clear(ctx);
+                llama_memory_clear(mem, false);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +143,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                        llama_memory_seq_cp(mem, 0, i, -1, -1);
                     }
                 }
 

tools/cvector-generator/cvector-generator.cpp

@@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;

tools/imatrix/imatrix.cpp

@@ -498,7 +498,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

tools/llama-bench/llama-bench.cpp

@@ -1900,7 +1900,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), false);
 
         // cool off before the test
         if (params.delay) {
@@ -1948,7 +1948,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_self_clear(ctx);
+            llama_memory_clear(llama_get_memory(ctx), false);
 
             if (t.n_depth > 0) {
                 if (params.progress) {

tools/main/main.cpp

@@ -147,6 +147,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
     auto chat_templates = common_chat_templates_init(model, params.chat_template);
 
@@ -351,7 +353,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1);
     }
 
     LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
@@ -599,8 +601,8 @@ int main(int argc, char ** argv) {
                     LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_self_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_memory_seq_rm (mem, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_memory_seq_add(mem, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -623,9 +625,9 @@ int main(int argc, char ** argv) {
                     LOG_DBG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_self_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_memory_seq_add(mem, 0, ga_i,                n_past,              ib*bd);
+                    llama_memory_seq_div(mem, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_memory_seq_add(mem, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 

tools/mtmd/mtmd-cli.cpp

@@ -70,6 +70,7 @@ struct mtmd_cli_context {
     llama_model       * model;
     llama_context     * lctx;
     const llama_vocab * vocab;
+    common_sampler    * smpl;
     llama_batch         batch;
     int                 n_batch;
 
@@ -89,8 +90,9 @@ struct mtmd_cli_context {
         model = llama_init.model.get();
         lctx = llama_init.context.get();
         vocab = llama_model_get_vocab(model);
+        smpl = common_sampler_init(model, params.sampling);
         n_threads = params.cpuparams.n_threads;
-        batch = llama_batch_init(params.n_batch, 0, 1);
+        batch = llama_batch_init(1, 0, 1); // batch for next token generation
         n_batch = params.n_batch;
 
         if (!model || !lctx) {
@@ -118,6 +120,11 @@ struct mtmd_cli_context {
         }
     }
 
+    ~mtmd_cli_context() {
+        llama_batch_free(batch);
+        common_sampler_free(smpl);
+    }
+
     void init_vision_context(common_params & params) {
         const char * clip_path = params.mmproj.path.c_str();
         mtmd_context_params mparams = mtmd_context_params_default();
@@ -153,7 +160,7 @@ struct mtmd_cli_context {
     }
 };
 
-static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
+static int generate_response(mtmd_cli_context & ctx, int n_predict) {
     llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
         if (i > n_predict || !g_is_generating || g_is_interrupted) {
@@ -161,9 +168,9 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
             break;
         }
 
-        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        llama_token token_id = common_sampler_sample(ctx.smpl, ctx.lctx, -1);
         generated_tokens.push_back(token_id);
-        common_sampler_accept(smpl, token_id, true);
+        common_sampler_accept(ctx.smpl, token_id, true);
 
         if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
             LOG("\n");
@@ -261,7 +268,6 @@ int main(int argc, char ** argv) {
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
 
-    struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling);
     int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
 
     // Ctrl+C handling
@@ -300,7 +306,7 @@ int main(int argc, char ** argv) {
         if (eval_message(ctx, msg, true)) {
             return 1;
         }
-        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
+        if (!g_is_interrupted && generate_response(ctx, n_predict)) {
             return 1;
         }
 
@@ -336,7 +342,7 @@ int main(int argc, char ** argv) {
             }
             if (line == "/clear") {
                 ctx.n_past = 0;
-                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                llama_memory_seq_rm(llama_get_memory(ctx.lctx), 0, 1, -1); // keep BOS
                 LOG("Chat history cleared\n\n");
                 continue;
             }
@@ -366,7 +372,7 @@ int main(int argc, char ** argv) {
                 return 1;
             }
             if (g_is_interrupted) break;
-            if (generate_response(ctx, smpl, n_predict)) {
+            if (generate_response(ctx, n_predict)) {
                 return 1;
             }
             content.clear();

tools/mtmd/mtmd-helper.cpp

@@ -311,6 +311,7 @@ int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
         GGML_ABORT("chunk type not supported");
     }
 
+    llama_batch_free(text_batch);
     return 0;
 }
 

tools/perplexity/perplexity.cpp

@@ -361,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -547,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params &
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1217,7 +1217,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1592,7 +1592,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1782,7 +1782,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 

tools/run/run.cpp

@@ -939,7 +939,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_kv_self_seq_pos_max(llama_data.context.get(), 0) == 0;
+    const bool is_first = llama_memory_seq_pos_max(llama_get_memory(llama_data.context.get()), 0) == 0;
 
     const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
     prompt_tokens.resize(n_prompt_tokens);
@@ -955,7 +955,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
     const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_kv_self_seq_pos_max(ctx.get(), 0);
+    const int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx.get()), 0);
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");

tools/server/server.cpp

@@ -360,7 +360,7 @@ struct server_task {
                 params.oaicompat_chat_syntax.format = defaults.oaicompat_chat_syntax.format;
             }
             params.oaicompat_chat_syntax.reasoning_format = params_base.reasoning_format;
-            params.oaicompat_chat_syntax.reasoning_in_content = params.stream;
+            params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (params_base.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
             params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
             params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
         }
@@ -2006,7 +2006,7 @@ struct server_context {
             }
         }
 
-        if (!llama_kv_self_can_shift(ctx)) {
+        if (!llama_memory_can_shift(llama_get_memory(ctx))) {
             if (params_base.ctx_shift) {
                 params_base.ctx_shift = false;
                 SRV_WRN("%s\n", "ctx_shift is not supported by this context, it will be disabled");
@@ -2016,6 +2016,11 @@ struct server_context {
                 params_base.n_cache_reuse = 0;
                 SRV_WRN("%s\n", "cache_reuse is not supported by this context, it will be disabled");
             }
+
+            if (!params_base.speculative.model.path.empty()) {
+                SRV_ERR("%s\n", "err: speculative decode is not supported by this context");
+                return false;
+            }
         }
 
         return true;
@@ -2219,7 +2224,7 @@ struct server_context {
         SRV_DBG("%s", "clearing KV cache\n");
 
         // clear the entire KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
         clean_kv_cache = false;
     }
 
@@ -2905,7 +2910,7 @@ struct server_context {
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
+                    llama_memory_seq_rm(llama_get_memory(ctx), slot->id, -1, -1);
                     slot->cache_tokens.clear();
 
                     auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2980,8 +2985,8 @@ struct server_context {
 
                 SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard);
 
-                llama_kv_self_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
-                llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
+                llama_memory_seq_rm (llama_get_memory(ctx), slot.id, n_keep            , n_keep + n_discard);
+                llama_memory_seq_add(llama_get_memory(ctx), slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 // add generated tokens to cache
                 {
@@ -3184,8 +3189,8 @@ struct server_context {
 
                                             const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
 
-                                            llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c);
-                                            llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift);
+                                            llama_memory_seq_rm (llama_get_memory(ctx), slot.id, head_p, head_c);
+                                            llama_memory_seq_add(llama_get_memory(ctx), slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens.set_token(head_p + i, slot.cache_tokens[head_c + i]);
@@ -3203,13 +3208,11 @@ struct server_context {
                                 }
                             } else {
                                 // if we don't cache the prompt, we have to remove the entire KV cache
-                                llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                 slot.n_past = 0;
-                                slot.cache_tokens.clear(); // TODO: not needed, will be cleared later via "keep_first()"
                             }
 
                             if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
-                                const auto pos_min = llama_kv_self_seq_pos_min(ctx, slot.id);
+                                const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
                                     SLT_ERR(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min);
                                     GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
@@ -3220,7 +3223,6 @@ struct server_context {
                                     SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min, n_swa);
                                     SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
                                             "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
-                                    llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                     slot.n_past = 0;
                                 }
                             }
@@ -3245,9 +3247,9 @@ struct server_context {
                     }
 
                     // keep only the common part
-                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1)) {
                         // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
+                        llama_memory_seq_rm(llama_get_memory(ctx), slot.id, -1, -1);
 
                         // there is no common part left
                         slot.n_past = 0;
@@ -3587,7 +3589,7 @@ struct server_context {
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert({ids.begin(), ids.end() - 1});
 
-                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1);
 
                 for (size_t i = 0; i < ids.size(); ++i) {
                     completion_token_output result;

tools/server/tests/unit/test_tool_call.py

@@ -499,13 +499,12 @@ def do_test_calc_result(server: ServerProcess, result_override: str | None, n_pr
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize("n_predict,reasoning_format,stream,expect_reasoning_content,expect_content,hf_repo,template_override", [
-    (128, 'deepseek',   CompletionMode.NORMAL,   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
-    (128,  None,        CompletionMode.NORMAL,   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
-    (1024, 'deepseek',  CompletionMode.NORMAL,   "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    (1024, 'deepseek',  CompletionMode.STREAMED, None, "^<think>I need to calculate [\\s\\S]*?</think>To find the sum of [\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    (1024, 'deepseek',  CompletionMode.NORMAL,   "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*",                                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
-    (1024, 'deepseek',  CompletionMode.STREAMED, None, "^<think>First, I [\\s\\S]*?</think>To find the sum of[\\s\\S]*",              "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+@pytest.mark.parametrize("stream", [CompletionMode.NORMAL, CompletionMode.STREAMED])
+@pytest.mark.parametrize("n_predict,reasoning_format,expect_reasoning_content,expect_content,hf_repo,template_override", [
+    (128, 'deepseek',   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (128,  None,        None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (1024, 'deepseek',  "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (1024, 'deepseek',  "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*",                                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
     # (1024, 'none',      CompletionMode.NORMAL,   None, "^(<think>\\s*)?I need[\\s\\S]*?</think>\\s*To find[\\s\\S]*",                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
     # (128,  'deepseek',  None, "^Okay, let me figure out the sum of 102 and 7[\\s\\S]*",                      "bartowski/Qwen_QwQ-32B-GGUF:Q4_K_M",                None),
 ])

tools/server/tests/utils.py

@@ -308,10 +308,12 @@ class ServerProcess:
         stream = data.get('stream', False)
         if stream:
             content: list[str] = []
+            reasoning_content: list[str] = []
             tool_calls: list[dict] = []
             finish_reason: Optional[str] = None
 
             content_parts = 0
+            reasoning_content_parts = 0
             tool_call_parts = 0
             arguments_parts = 0
 
@@ -322,6 +324,10 @@ class ServerProcess:
                     assert len(choice['delta']['content']) > 0, f'Expected non empty content delta!'
                     content.append(choice['delta']['content'])
                     content_parts += 1
+                if choice['delta'].get('reasoning_content') is not None:
+                    assert len(choice['delta']['reasoning_content']) > 0, f'Expected non empty reasoning_content delta!'
+                    reasoning_content.append(choice['delta']['reasoning_content'])
+                    reasoning_content_parts += 1
                 if choice['delta'].get('finish_reason') is not None:
                     finish_reason = choice['delta']['finish_reason']
                 for tc in choice['delta'].get('tool_calls', []):
@@ -349,8 +355,10 @@ class ServerProcess:
                         tool_call['function']['name'] = tool_call['function'].get('name', '') + fct['name']
                     if fct.get('arguments') is not None:
                         tool_call['function']['arguments'] += fct['arguments']
+                        arguments_parts += 1
+                    tool_call_parts += 1
 
-            print(f'Streamed response had {content_parts} content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
+            print(f'Streamed response had {content_parts} content parts, {reasoning_content_parts} reasoning_content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
             result = dict(
                 choices=[
                     dict(
@@ -359,6 +367,7 @@ class ServerProcess:
                         message=dict(
                             role='assistant',
                             content=''.join(content) if content else None,
+                            reasoning_content=''.join(reasoning_content) if reasoning_content else None,
                             tool_calls=tool_calls if tool_calls else None,
                         ),
                     )