common : Fix corrupted memory error on json grammar initialization (#16038)

Initalizing RESERVED_NAME in is_reserved_name() is not thread
safe and leads to corrupted memory when used from multiple threads
as can be seen in the asan trace below. This fixes the initialization
to make it thread-safe.

    #0 0x000100abd018 in std::__1::pair<std::__1::__hash_iterator<std::__1::__hash_node<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, void*>*>, bool> std::__1::__hash_table<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::hash<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::equal_to<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::allocator<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>>::__emplace_unique_key_args<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&>(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&) __hash_table:1565
    #1 0x000100ab0320 in SchemaConverter::visit(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&) json-schema-to-grammar.cpp:802
    #2 0x000100aafc48 in std::__1::__function::__func<build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&)::$_2, std::__1::allocator<build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&)::$_2>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> (std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>::operator()(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&) function.h:319
    #3 0x000100a2c938 in std::__1::__function::__func<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0::operator()(common_grammar_builder const&) const::'lambda'(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&), std::__1::allocator<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0::operator()(common_grammar_builder const&) const::'lambda'(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>, void (nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)>::operator()(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&) function.h:319
    #4 0x000100a139f8 in foreach_function(nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&, std::__1::function<void (nlohmann::json_abi_v3_12_0::basic_json<nlohmann::json_abi_v3_12_0::ordered_map, std::__1::vector, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, bool, long long, unsigned long long, double, std::__1::allocator, nlohmann::json_abi_v3_12_0::adl_serializer, std::__1::vector<unsigned char, std::__1::allocator<unsigned char>>, void> const&)> const&) chat.cpp:762
    #5 0x000100a2a7f4 in std::__1::__function::__func<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0, std::__1::allocator<common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool)::$_0>, void (common_grammar_builder const&)>::operator()(common_grammar_builder const&) function.h:319
    #6 0x000100aa98f4 in build_grammar(std::__1::function<void (common_grammar_builder const&)> const&, common_grammar_options const&) json-schema-to-grammar.cpp:982
    #7 0x0001009c9314 in common_chat_params_init_llama_3_x(minja::chat_template const&, templates_params const&, bool) chat.cpp:1110
    #8 0x0001009b8afc in common_chat_templates_apply_jinja(common_chat_templates const*, common_chat_templates_inputs const&) chat.cpp:1992
    #9 0x0001009b533c in common_chat_templates_apply(common_chat_templates const*, common_chat_templates_inputs const&) chat.cpp:2074
    #10 0x000100810120 in llamacpp_apply_chat_template+0x724 (predict_oai-98384e17fb94e863:arm64+0x100090120)
    ...

==45482==Register values:
 x[0] = 0x00006020004147f8   x[1] = 0x00006080000013c8   x[2] = 0x0000000000000000   x[3] = 0x0000604006289738
 x[4] = 0x0000000000000002   x[5] = 0x0000000000000001   x[6] = 0x04034000004b4000   x[7] = 0x0000000000000001
 x[8] = 0xbebebebebebebebe   x[9] = 0x17d7d7d7d7d7d7d7  x[10] = 0x00000c04000828ff  x[11] = 0x0000000000000001
x[12] = 0x000000002018d383  x[13] = 0x0000000000000000  x[14] = 0xfa0000000000fafa  x[15] = 0x000010700001ffff
x[16] = 0x000000019dc012c0  x[17] = 0x00000001021284f8  x[18] = 0x0000000000000000  x[19] = 0x00000001700acdc0
x[20] = 0x0000000000000002  x[21] = 0x000000002018d384  x[22] = 0x16dd16fd2e731151  x[23] = 0x0000007000020000
x[24] = 0x0000000100c69c08  x[25] = 0x0000000100c69c20  x[26] = 0x00006080000013c7  x[27] = 0x0000000100c69c00
x[28] = 0x00000001700acd60     fp = 0x00000001700aceb0     lr = 0x0000000100abce30     sp = 0x00000001700acd60
AddressSanitizer can not provide additional info.
SUMMARY: AddressSanitizer: SEGV __hash_table:1565 in std::__1::pair<std::__1::__hash_iterator<std::__1::__hash_node<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, void*>*>, bool> std::__1::__hash_table<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::hash<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::equal_to<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>, std::__1::allocator<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>>::__emplace_unique_key_args<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&>(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&)
Thread T5 created by T0 here:
    #0 0x0001020b99d4 in pthread_create+0x5c (libclang_rt.asan_osx_dynamic.dylib:arm64e+0x359d4)
    #1 0x000100873910 in std::sys::pal::unix::thread::Thread::new::h77254fdd87a28e05+0x118 (predict_oai-98384e17fb94e863:arm64+0x1000f3910)
    #2 0x0001007c7a1c in test::run_test::haeb3c2bcd5ed6cf6+0x76c (predict_oai-98384e17fb94e863:arm64+0x100047a1c)
    #3 0x0001007aedb0 in test::console::run_tests_console::he9d142d704f3a986+0x149c (predict_oai-98384e17fb94e863:arm64+0x10002edb0)
    #4 0x0001007c5758 in test::test_main::hf86a5e20735245b9+0x118 (predict_oai-98384e17fb94e863:arm64+0x100045758)
    #5 0x0001007c5da0 in test::test_main_static::h61ee9c8fd30abca0+0x54 (predict_oai-98384e17fb94e863:arm64+0x100045da0)
    ...

==45482==ABORTING

.clang-format

@@ -22,7 +22,14 @@ AllowShortIfStatementsOnASingleLine: Never
 AllowShortLambdasOnASingleLine: Inline
 AllowShortLoopsOnASingleLine: false
 AlwaysBreakBeforeMultilineStrings: true
-BinPackArguments: false
+# Treat CUDA keywords/attributes as "attribute macros" and avoid breaking lines inside them
+AttributeMacros:
+  - __host__
+  - __device__
+  - __global__
+  - __forceinline__
+  - __launch_bounds__
+BinPackArguments: true
 BinPackParameters: false # OnePerLine
 BitFieldColonSpacing: Both
 BreakBeforeBraces: Custom # Attach

.devops/rocm.Dockerfile

@@ -4,7 +4,7 @@ ARG UBUNTU_VERSION=24.04
 ARG ROCM_VERSION=6.4
 ARG AMDGPU_VERSION=6.4
 
-# Target the CUDA build image
+# Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
 ### Build image
@@ -15,16 +15,13 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # This is mostly tied to rocBLAS supported archs.
 # gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
 # gfx906 is deprecated
-#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html
+#check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
 
-ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102'
-#ARG ROCM_DOCKER_ARCH=gfx1100
+ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
+#ARG ROCM_DOCKER_ARCH='gfx1151'
 
-# Set nvcc architectured
+# Set ROCm architectures
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
-# Enable ROCm
-# ENV CC=/opt/rocm/llvm/bin/clang
-# ENV CXX=/opt/rocm/llvm/bin/clang++
 
 RUN apt-get update \
     && apt-get install -y \
@@ -39,8 +36,16 @@ WORKDIR /app
 
 COPY . .
 
+RUN git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
+    cmake -S . -B build \
+        -DGGML_HIP=ON \
+        -DGGML_HIP_ROCWMMA_FATTN=ON \
+        -DCMAKE_HIP_FLAGS="-I$(pwd)/rocwmma/library/include/" \
+        -DAMDGPU_TARGETS="$ROCM_DOCKER_ARCH" \
+        -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON \
+        -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -2,14 +2,30 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Install build tools
-RUN apt update && apt install -y git build-essential cmake wget
+# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
 
-# Install Vulkan SDK and cURL
-RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
-    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-noble.list https://packages.lunarg.com/vulkan/lunarg-vulkan-noble.list && \
-    apt update -y && \
-    apt-get install -y vulkan-sdk libcurl4-openssl-dev curl
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget xz-utils
+
+# Install Vulkan SDK
+ARG VULKAN_VERSION=1.4.321.1
+RUN ARCH=$(uname -m) && \
+    wget -qO /tmp/vulkan-sdk.tar.xz https://sdk.lunarg.com/sdk/download/${VULKAN_VERSION}/linux/vulkan-sdk-linux-${ARCH}-${VULKAN_VERSION}.tar.xz && \
+    mkdir -p /opt/vulkan && \
+    tar -xf /tmp/vulkan-sdk.tar.xz -C /tmp --strip-components=1 && \
+    mv /tmp/${ARCH}/* /opt/vulkan/ && \
+    rm -rf /tmp/*
+
+# Install cURL and Vulkan SDK dependencies
+RUN apt install -y libcurl4-openssl-dev curl \
+    libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev
+
+# Set environment variables
+ENV VULKAN_SDK=/opt/vulkan
+ENV PATH=$VULKAN_SDK/bin:$PATH
+ENV LD_LIBRARY_PATH=$VULKAN_SDK/lib:$LD_LIBRARY_PATH
+ENV CMAKE_PREFIX_PATH=$VULKAN_SDK:$CMAKE_PREFIX_PATH
+ENV PKG_CONFIG_PATH=$VULKAN_SDK/lib/pkgconfig:$PKG_CONFIG_PATH
 
 # Build it
 WORKDIR /app

.github/copilot-instructions.md

@@ -0,0 +1,262 @@
+# Copilot Instructions for llama.cpp
+
+## Repository Overview
+
+llama.cpp is a large-scale C/C++ project for efficient LLM (Large Language Model) inference with minimal setup and dependencies. The project enables running language models on diverse hardware with state-of-the-art performance.
+
+**Key Facts:**
+- **Primary language**: C/C++ with Python utility scripts
+- **Size**: ~200k+ lines of code across 1000+ files
+- **Architecture**: Modular design with main library (`libllama`) and 40+ executable tools/examples
+- **Core dependency**: ggml tensor library (vendored in `ggml/` directory)
+- **Backends supported**: CPU (AVX/NEON optimized), CUDA, Metal, Vulkan, SYCL, ROCm, MUSA
+- **License**: MIT
+
+## Build Instructions
+
+### Prerequisites
+- CMake 3.14+ (primary build system)
+- C++17 compatible compiler (GCC 13.3+, Clang, MSVC)
+- Optional: ccache for faster compilation
+
+### Basic Build (CPU-only)
+**ALWAYS run these commands in sequence:**
+```bash
+cmake -B build
+cmake --build build --config Release -j $(nproc)
+```
+
+**Build time**: ~10 minutes on 4-core system with ccache enabled, ~25 minutes without ccache.
+
+**Important Notes:**
+- The Makefile is deprecated - always use CMake
+- ccache is automatically detected and used if available
+- Built binaries are placed in `build/bin/`
+- Parallel builds (`-j`) significantly reduce build time
+
+### Backend-Specific Builds
+For CUDA support:
+```bash
+cmake -B build -DGGML_CUDA=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+For Metal (macOS):
+```bash
+cmake -B build -DGGML_METAL=ON
+cmake --build build --config Release -j $(nproc)
+```
+
+**Important Note**: While all backends can be built as long as the correct requirements for that backend are installed, you will not be able to run them without the correct hardware. The only backend that can be run for testing and validation is the CPU backend.
+
+### Debug Builds
+Single-config generators:
+```bash
+cmake -B build -DCMAKE_BUILD_TYPE=Debug
+cmake --build build
+```
+
+Multi-config generators:
+```bash
+cmake -B build -G "Xcode"
+cmake --build build --config Debug
+```
+
+### Common Build Issues
+- **Issue**: Network tests fail in isolated environments
+  **Solution**: Expected behavior - core functionality tests will still pass
+
+## Testing
+
+### Running Tests
+```bash
+ctest --test-dir build --output-on-failure -j $(nproc)
+```
+
+**Test suite**: 38 tests covering tokenizers, grammar parsing, sampling, backends, and integration
+**Expected failures**: 2-3 tests may fail if network access is unavailable (they download models)
+**Test time**: ~30 seconds for passing tests
+
+### Server Unit Tests
+Run server-specific unit tests after building the server:
+```bash
+# Build the server first
+cmake --build build --target llama-server
+
+# Navigate to server tests and run
+cd tools/server/tests
+source ../../../.venv/bin/activate
+./tests.sh
+```
+**Server test dependencies**: The `.venv` environment includes the required dependencies for server unit tests (pytest, aiohttp, etc.). Tests can be run individually or with various options as documented in `tools/server/tests/README.md`.
+
+### Test Categories
+- Tokenizer tests: Various model tokenizers (BERT, GPT-2, LLaMA, etc.)
+- Grammar tests: GBNF parsing and validation
+- Backend tests: Core ggml operations across different backends
+- Integration tests: End-to-end workflows
+
+### Manual Testing Commands
+```bash
+# Test basic inference
+./build/bin/llama-cli --version
+
+# Test model loading (requires model file)
+./build/bin/llama-cli -m path/to/model.gguf -p "Hello" -n 10
+```
+
+## Code Quality and Linting
+
+### C++ Code Formatting
+**ALWAYS format C++ code before committing:**
+```bash
+git clang-format
+```
+
+Configuration is in `.clang-format` with these key rules:
+- 4-space indentation
+- 120 column limit
+- Braces on same line for functions
+- Pointer alignment: `void * ptr` (middle)
+- Reference alignment: `int & ref` (middle)
+
+### Python Code
+**ALWAYS activate the Python environment in `.venv` and use tools from that environment:**
+```bash
+# Activate virtual environment
+source .venv/bin/activate
+```
+
+Configuration files:
+- `.flake8`: flake8 settings (max-line-length=125, excludes examples/tools)
+- `pyrightconfig.json`: pyright type checking configuration
+
+### Pre-commit Hooks
+Run before committing:
+```bash
+pre-commit run --all-files
+```
+
+## Continuous Integration
+
+### GitHub Actions Workflows
+Key workflows that run on every PR:
+- `.github/workflows/build.yml`: Multi-platform builds
+- `.github/workflows/server.yml`: Server functionality tests
+- `.github/workflows/python-lint.yml`: Python code quality
+- `.github/workflows/python-type-check.yml`: Python type checking
+
+### Local CI Validation
+**Run full CI locally before submitting PRs:**
+```bash
+mkdir tmp
+
+# CPU-only build
+bash ./ci/run.sh ./tmp/results ./tmp/mnt
+```
+
+**CI Runtime**: 30-60 minutes depending on backend configuration
+
+### Triggering CI
+Add `ggml-ci` to commit message to trigger heavy CI workloads on the custom CI infrastructure.
+
+## Project Layout and Architecture
+
+### Core Directories
+- **`src/`**: Main llama library implementation (`llama.cpp`, `llama-*.cpp`)
+- **`include/`**: Public API headers, primarily `include/llama.h`
+- **`ggml/`**: Core tensor library (submodule with custom GGML framework)
+- **`examples/`**: 30+ example applications and tools
+- **`tools/`**: Additional development and utility tools (server benchmarks, tests)
+- **`tests/`**: Comprehensive test suite with CTest integration
+- **`docs/`**: Detailed documentation (build guides, API docs, etc.)
+- **`scripts/`**: Utility scripts for CI, data processing, and automation
+- **`common/`**: Shared utility code used across examples
+
+### Key Files
+- **`CMakeLists.txt`**: Primary build configuration
+- **`include/llama.h`**: Main C API header (~2000 lines)
+- **`src/llama.cpp`**: Core library implementation (~8000 lines)
+- **`CONTRIBUTING.md`**: Coding guidelines and PR requirements
+- **`.clang-format`**: C++ formatting rules
+- **`.pre-commit-config.yaml`**: Git hook configuration
+
+### Built Executables (in `build/bin/`)
+Primary tools:
+- **`llama-cli`**: Main inference tool
+- **`llama-server`**: OpenAI-compatible HTTP server
+- **`llama-quantize`**: Model quantization utility
+- **`llama-perplexity`**: Model evaluation tool
+- **`llama-bench`**: Performance benchmarking
+- **`llama-convert-llama2c-to-ggml`**: Model conversion utilities
+
+### Configuration Files
+- **CMake**: `CMakeLists.txt`, `cmake/` directory
+- **Linting**: `.clang-format`, `.clang-tidy`, `.flake8`
+- **CI**: `.github/workflows/`, `ci/run.sh`
+- **Git**: `.gitignore` (includes build artifacts, models, cache)
+
+### Dependencies
+- **System**: OpenMP, libcurl (for model downloading)
+- **Optional**: CUDA SDK, Metal framework, Vulkan SDK, Intel oneAPI
+- **Bundled**: httplib, json (header-only libraries in vendored form)
+
+## Common Validation Steps
+
+### After Making Changes
+1. **Format code**: `git clang-format`
+2. **Build**: `cmake --build build --config Release`
+3. **Test**: `ctest --test-dir build --output-on-failure`
+4. **Server tests** (if modifying server): `cd tools/server/tests && source ../../../.venv/bin/activate && ./tests.sh`
+5. **Manual validation**: Test relevant tools in `build/bin/`
+
+### Performance Validation
+```bash
+# Benchmark inference performance
+./build/bin/llama-bench -m model.gguf
+
+# Evaluate model perplexity
+./build/bin/llama-perplexity -m model.gguf -f dataset.txt
+```
+
+### Backend Validation
+```bash
+# Test backend operations
+./build/bin/test-backend-ops
+```
+
+## Environment Setup
+
+### Required Tools
+- CMake 3.14+ (install via system package manager)
+- Modern C++ compiler with C++17 support
+- Git (for submodule management)
+- Python 3.9+ with virtual environment (`.venv` is provided)
+
+### Optional but Recommended
+- ccache: `apt install ccache` or `brew install ccache`
+- clang-format 15+: Usually included with LLVM/Clang installation
+- pre-commit: `pip install pre-commit`
+
+### Backend-Specific Requirements
+- **CUDA**: NVIDIA CUDA Toolkit 11.2+
+- **Metal**: Xcode command line tools (macOS only)
+- **Vulkan**: Vulkan SDK
+- **SYCL**: Intel oneAPI toolkit
+
+## Important Guidelines
+
+### Code Changes
+- **Minimal dependencies**: Avoid adding new external dependencies
+- **Cross-platform compatibility**: Test on Linux, macOS, Windows when possible
+- **Performance focus**: This is a performance-critical inference library
+- **API stability**: Changes to `include/llama.h` require careful consideration
+
+### Git Workflow
+- Always create feature branches from `master`
+- **Never** commit build artifacts (`build/`, `.ccache/`, `*.o`, `*.gguf`)
+- Use descriptive commit messages following project conventions
+
+### Trust These Instructions
+Only search for additional information if these instructions are incomplete or found to be incorrect. This document contains validated build and test procedures that work reliably across different environments.
+

.github/workflows/build-riscv-native.yml

@@ -1,10 +1,11 @@
 name: Build on RISCV Linux Machine by Cloud-V
 on:
+  pull_request:
   workflow_dispatch:
   workflow_call:
 
 jobs:
-  bianbu-riscv64-native: # Bianbu 2.2
+  debian-13-riscv64-native: # Bianbu 2.2
     runs-on: self-hosted
 
     steps:
@@ -20,24 +21,40 @@ jobs:
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
                   g++-14-riscv64-linux-gnu \
+                  ccache \
                   cmake
 
+      - name: Setup ccache
+        run: |
+          mkdir -p $HOME/.ccache
+          ccache -M 5G -d $HOME/.ccache
+          export CCACHE_LOGFILE=/home/runneruser/ccache_debug/ccache.log
+          export CCACHE_DEBUGDIR="/home/runneruser/ccache_debug"
+          echo "$GITHUB_WORKSPACE"
+          echo "CCACHE_LOGFILE=$CCACHE_LOGFILE" >> $GITHUB_ENV
+          echo "CCACHE_DEBUGDIR=$CCACHE_DEBUGDIR" >> $GITHUB_ENV
+          echo "CCACHE_BASEDIR=$GITHUB_WORKSPACE" >> $GITHUB_ENV
+          echo "CCACHE_DIR=$HOME/.ccache" >> $GITHUB_ENV
+
       - 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=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+          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=riscv64 \
+            -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+            -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+            -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+            -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-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

@@ -56,7 +56,7 @@ env:
 
 jobs:
   macOS-latest-cmake-arm64:
-    runs-on: macos-14
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -88,6 +88,7 @@ jobs:
             -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
+          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
 
       - name: Test
         id: cmake_test
@@ -126,7 +127,8 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -136,7 +138,7 @@ jobs:
           ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-arm64-webgpu:
-    runs-on: macos-14
+    runs-on: macos-latest
 
     steps:
       - name: Clone
@@ -709,6 +711,7 @@ jobs:
 
   macOS-latest-swift:
     runs-on: macos-latest
+    needs: ios-xcode-build
 
     strategy:
       matrix:
@@ -725,6 +728,12 @@ jobs:
           key: macOS-latest-swift
           evict-old-files: 1d
 
+      - name: Download xcframework artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+
       - name: Dependencies
         id: depends
         continue-on-error: true
@@ -746,11 +755,6 @@ jobs:
             -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          ./build-xcframework.sh
-
   windows-msys2:
     runs-on: windows-2025
 
@@ -1050,9 +1054,13 @@ jobs:
         run:  examples/sycl/win-build-sycl.bat
 
   windows-latest-cmake-hip:
-    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-2022
 
+    env:
+      # The ROCm version must correspond to the version used in the HIP SDK.
+      ROCM_VERSION: "6.4.2"
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
+
     steps:
       - name: Clone
         id: checkout
@@ -1061,23 +1069,46 @@ jobs:
       - name: Clone rocWMMA repository
         id: clone_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+          git clone https://github.com/rocm/rocwmma --branch rocm-${{ env.ROCM_VERSION }} --depth 1
 
-      - name: Install
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
         id: verify
         run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1141,8 +1172,17 @@ jobs:
         run: |
           ./build-xcframework.sh
 
+      - name: Upload xcframework artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+          retention-days: 1
+
       - name: Build Xcode project
-        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+        run: |
+          xcodebuild -downloadPlatform iOS
+          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
 
   android-build:
     runs-on: ubuntu-latest

.github/workflows/close-issue.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - uses: actions/stale@v5
         with:
-          exempt-issue-labels: "refactoring,help wanted,good first issue,research,bug,roadmap"
+          exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
           days-before-issue-stale: 30
           days-before-issue-close: 14
           stale-issue-label: "stale"

.github/workflows/copilot-setup-steps.yml

@@ -39,6 +39,10 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install build-essential libcurl4-openssl-dev
+          # Install git-clang-format script for formatting only changed code
+          wget -O /tmp/git-clang-format https://raw.githubusercontent.com/llvm/llvm-project/release/18.x/clang/tools/clang-format/git-clang-format
+          sudo cp /tmp/git-clang-format /usr/local/bin/git-clang-format
+          sudo chmod +x /usr/local/bin/git-clang-format
 
       - name: Set up Python
         uses: actions/setup-python@v5
@@ -50,4 +54,4 @@ jobs:
           python3 -m venv .venv
           .venv/bin/activate
           pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
-          pip install flake8 pyright
+          pip install flake8 pyright pre-commit

.github/workflows/release.yml

@@ -108,7 +108,8 @@ jobs:
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Determine tag name
@@ -528,11 +529,14 @@ jobs:
   windows-hip:
     runs-on: windows-2022
 
+    env:
+      HIPSDK_INSTALLER_VERSION: "25.Q3"
+
     strategy:
       matrix:
         include:
           - name: "radeon"
-            gpu_targets: "gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
@@ -542,29 +546,52 @@ jobs:
       - name: Clone rocWMMA repository
         id: clone_rocwmma
         run: |
-          git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
+          git clone https://github.com/rocm/rocwmma --branch develop --depth 1
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\Program Files\AMD\ROCm
+          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: windows-latest-cmake-hip-${{ matrix.name }}-x64
+          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ matrix.name }}-x64
           evict-old-files: 1d
 
-      - name: Install
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          $completed = $proc.WaitForExit(600000)
+          if (-not $completed) {
+              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
+              $proc.Kill()
+              exit 1
+          }
+          if ($proc.ExitCode -ne 0) {
+              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
+              exit 1
+          }
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
         id: verify
         run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+          # Find and test ROCm installation
+          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
+          if (-not $clangPath) {
+            Write-Error "ROCm installation not found"
+            exit 1
+          }
+          & $clangPath.FullName --version
 
       - name: Build
         id: cmake_build
@@ -585,9 +612,12 @@ jobs:
             -DLLAMA_CURL=OFF
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
+          md "build\bin\hipblaslt\library"
           cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
 
       - name: Pack artifacts
         id: pack_artifacts

.gitignore

@@ -147,3 +147,4 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
+.ccache/

CMakeLists.txt

@@ -58,6 +58,12 @@ if (MSVC)
     add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
 endif()
 
+if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
+    set(LLAMA_TOOLS_INSTALL_DEFAULT OFF)
+else()
+    set(LLAMA_TOOLS_INSTALL_DEFAULT ${LLAMA_STANDALONE})
+endif()
+
 #
 # option list
 #
@@ -82,6 +88,7 @@ option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
+option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
 
 # 3rd party libs
 option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)

CONTRIBUTING.md

@@ -16,6 +16,9 @@
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
 - Consider adding yourself to [CODEOWNERS](CODEOWNERS)
+- Let authors, who are also collaborators, merge their own PRs
+- When merging a PR by a contributor, make sure you have a good understanding of the changes
+- Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 
 # Coding guidelines
 

README.md

@@ -137,6 +137,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
 - [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
+- [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
 
 #### Multimodal
 
@@ -151,6 +152,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 - [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
+- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
 
 </details>
 

ci/run.sh

@@ -106,7 +106,7 @@ function gg_wget {
     cd $out
 
     # should not re-download if file is the same
-    wget -nv -N $url
+    wget -nv -c -N $url
 
     cd $cwd
 }
@@ -270,7 +270,9 @@ function gg_run_ctest_with_model_debug {
     local model; model=$(gg_get_model)
     cd build-ci-debug
     set -e
+
     (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
     set +e
     cd ..
 }
@@ -281,7 +283,15 @@ function gg_run_ctest_with_model_release {
     local model; model=$(gg_get_model)
     cd build-ci-release
     set -e
+
     (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    # test memory leaks
+    #if [[ ! -z ${GG_BUILD_METAL} ]]; then
+    #    # TODO: this hangs for some reason ...
+    #    (time leaks -quiet -atExit -- ./bin/test-thread-safety -m $model --parallel 2 -t 2 -p "hello") 2>&1 | tee -a $OUT/${ci}-leaks.log
+    #fi
+
     set +e
     cd ..
 }
@@ -386,10 +396,10 @@ function gg_run_open_llama_7b_v2 {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -520,8 +530,8 @@ function gg_run_pythia_1_4b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -651,10 +661,10 @@ function gg_run_pythia_2_8b {
 
     (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -860,10 +870,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 fi
 
 ret=0
-if [ -z ${GG_BUILD_SYCL} ]; then
-    # SYCL build breaks with debug build flags
-    test $ret -eq 0 && gg_run ctest_debug
-fi
+test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -871,9 +878,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run test_scripts_debug
-        fi
+        test $ret -eq 0 && gg_run test_scripts_debug
         test $ret -eq 0 && gg_run test_scripts_release
     fi
 
@@ -884,9 +889,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
             test $ret -eq 0 && gg_run pythia_2_8b
             #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        if [ -z ${GG_BUILD_SYCL} ]; then
-            test $ret -eq 0 && gg_run ctest_with_model_debug
-        fi
+        test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi

common/arg.cpp

@@ -745,6 +745,124 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 
 #endif // LLAMA_USE_CURL
 
+//
+// Docker registry functions
+//
+
+static std::string common_docker_get_token(const std::string & repo) {
+    std::string url = "https://auth.docker.io/token?service=registry.docker.io&scope=repository:" + repo + ":pull";
+
+    common_remote_params params;
+    auto                 res = common_remote_get_content(url, params);
+
+    if (res.first != 200) {
+        throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
+    }
+
+    std::string            response_str(res.second.begin(), res.second.end());
+    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
+
+    if (!response.contains("token")) {
+        throw std::runtime_error("Docker registry token response missing 'token' field");
+    }
+
+    return response["token"].get<std::string>();
+}
+
+static std::string common_docker_resolve_model(const std::string & docker) {
+    // Parse ai/smollm2:135M-Q4_K_M
+    size_t      colon_pos = docker.find(':');
+    std::string repo, tag;
+    if (colon_pos != std::string::npos) {
+        repo = docker.substr(0, colon_pos);
+        tag  = docker.substr(colon_pos + 1);
+    } else {
+        repo = docker;
+        tag  = "latest";
+    }
+
+    // ai/ is the default
+    size_t      slash_pos = docker.find('/');
+    if (slash_pos == std::string::npos) {
+        repo.insert(0, "ai/");
+    }
+
+    LOG_INF("%s: Downloading Docker Model: %s:%s\n", __func__, repo.c_str(), tag.c_str());
+    try {
+        // --- helper: digest validation ---
+        auto validate_oci_digest = [](const std::string & digest) -> std::string {
+            // Expected: algo:hex ; start with sha256 (64 hex chars)
+            // You can extend this map if supporting other algorithms in future.
+            static const std::regex re("^sha256:([a-fA-F0-9]{64})$");
+            std::smatch m;
+            if (!std::regex_match(digest, m, re)) {
+                throw std::runtime_error("Invalid OCI digest format received in manifest: " + digest);
+            }
+            // normalize hex to lowercase
+            std::string normalized = digest;
+            std::transform(normalized.begin()+7, normalized.end(), normalized.begin()+7, [](unsigned char c){
+                return std::tolower(c);
+            });
+            return normalized;
+        };
+
+        std::string token = common_docker_get_token(repo);  // Get authentication token
+
+        // Get manifest
+        const std::string    url_prefix = "https://registry-1.docker.io/v2/" + repo;
+        std::string          manifest_url = url_prefix + "/manifests/" + tag;
+        common_remote_params manifest_params;
+        manifest_params.headers.push_back("Authorization: Bearer " + token);
+        manifest_params.headers.push_back(
+            "Accept: application/vnd.docker.distribution.manifest.v2+json,application/vnd.oci.image.manifest.v1+json");
+        auto manifest_res = common_remote_get_content(manifest_url, manifest_params);
+        if (manifest_res.first != 200) {
+            throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
+        }
+
+        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
+        std::string            gguf_digest;  // Find the GGUF layer
+        if (manifest.contains("layers")) {
+            for (const auto & layer : manifest["layers"]) {
+                if (layer.contains("mediaType")) {
+                    std::string media_type = layer["mediaType"].get<std::string>();
+                    if (media_type == "application/vnd.docker.ai.gguf.v3" ||
+                        media_type.find("gguf") != std::string::npos) {
+                        gguf_digest = layer["digest"].get<std::string>();
+                        break;
+                    }
+                }
+            }
+        }
+
+        if (gguf_digest.empty()) {
+            throw std::runtime_error("No GGUF layer found in Docker manifest");
+        }
+
+        // Validate & normalize digest
+        gguf_digest = validate_oci_digest(gguf_digest);
+        LOG_DBG("%s: Using validated digest: %s\n", __func__, gguf_digest.c_str());
+
+        // Prepare local filename
+        std::string model_filename = repo;
+        std::replace(model_filename.begin(), model_filename.end(), '/', '_');
+        model_filename += "_" + tag + ".gguf";
+        std::string local_path = fs_get_cache_file(model_filename);
+
+        const std::string blob_url = url_prefix + "/blobs/" + gguf_digest;
+        if (!common_download_file_single(blob_url, local_path, token, false)) {
+            throw std::runtime_error("Failed to download Docker Model");
+        }
+
+        LOG_INF("%s: Downloaded Docker Model to: %s\n", __func__, local_path.c_str());
+        return local_path;
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: Docker Model download failed: %s\n", __func__, e.what());
+        throw;
+    }
+}
+
 //
 // utils
 //
@@ -795,7 +913,9 @@ static handle_model_result common_params_handle_model(
     handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
-        if (!model.hf_repo.empty()) {
+        if (!model.docker_repo.empty()) {  // Handle Docker URLs by resolving them to local paths
+            model.path = common_docker_resolve_model(model.docker_repo);
+        } else if (!model.hf_repo.empty()) {
             // short-hand to avoid specifying --hf-file -> default it to --model
             if (model.hf_file.empty()) {
                 if (model.path.empty()) {
@@ -1106,7 +1226,7 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
     printf("\"\n\n");
 
     printf("    case \"$prev\" in\n");
-    printf("        --model)\n");
+    printf("        --model|-m)\n");
     printf("            COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
     printf("            return 0\n");
     printf("            ;;\n");
@@ -1184,7 +1304,7 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     } else {
         for (const auto & device : dev_names) {
             auto * dev = ggml_backend_dev_by_name(device.c_str());
-            if (!dev || ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+            if (!dev || ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
                 throw std::invalid_argument(string_format("invalid device: %s", device.c_str()));
             }
             devices.push_back(dev);
@@ -1194,7 +1314,7 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
-static void add_rpc_devices(std::string servers) {
+static void add_rpc_devices(const std::string & servers) {
     auto rpc_servers = string_split<std::string>(servers, ',');
     if (rpc_servers.empty()) {
         throw std::invalid_argument("no RPC servers specified");
@@ -1263,6 +1383,18 @@ static std::string list_builtin_chat_templates() {
     return msg.str();
 }
 
+static bool is_truthy(const std::string & value) {
+    return value == "on" || value == "enabled" || value == "1";
+}
+
+static bool is_falsey(const std::string & value) {
+    return value == "off" || value == "disabled" || value == "0";
+}
+
+static bool is_autoy(const std::string & value) {
+    return value == "auto" || value == "-1";
+}
+
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     // load dynamic backends
     ggml_backend_load_all();
@@ -1532,7 +1664,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--context-shift"},
-        string_format("enables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
+        string_format("enables context shift on infinite text generation (default: %s)", params.ctx_shift ? "enabled" : "disabled"),
         [](common_params & params) {
             params.ctx_shift = true;
         }
@@ -1544,13 +1676,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_chunks = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL}));
-    add_opt(common_arg(
-        {"-fa", "--flash-attn"},
-        string_format("enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled"),
-        [](common_params & params) {
-            params.flash_attn = true;
-        }
-    ).set_env("LLAMA_ARG_FLASH_ATTN"));
+    add_opt(common_arg({ "-fa", "--flash-attn" }, "[on|off|auto]",
+                       string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')",
+                                     llama_flash_attn_type_name(params.flash_attn_type)),
+                       [](common_params & params, const std::string & value) {
+                           if (is_truthy(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
+                           } else if (is_falsey(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
+                           } else if (is_autoy(value)) {
+                               params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
+                           } else {
+                               throw std::runtime_error(
+                                   string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
+                           }
+                       }).set_env("LLAMA_ARG_FLASH_ATTN"));
     add_opt(common_arg(
         {"-p", "--prompt"}, "PROMPT",
         "prompt to start generation with; for system message, use -sys",
@@ -1564,7 +1704,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.system_prompt = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_DIFFUSION}));
     add_opt(common_arg(
         {"--no-perf"},
         string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@@ -1755,7 +1895,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.warmup = false;
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -2254,9 +2394,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
     add_opt(common_arg(
         {"-dt", "--defrag-thold"}, "N",
-        string_format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold),
+        string_format("KV cache defragmentation threshold (DEPRECATED)"),
         [](common_params & params, const std::string & value) {
-            params.defrag_thold = std::stof(value);
+            GGML_UNUSED(params);
+            GGML_UNUSED(value);
+            LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
     add_opt(common_arg(
@@ -2374,24 +2516,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--list-devices"},
         "print list of available devices and exit",
         [](common_params &) {
-            std::vector<ggml_backend_dev_t> rpc_devices;
-            std::vector<ggml_backend_dev_t> all_devices;
+            std::vector<ggml_backend_dev_t> devices;
             for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
                 auto * dev = ggml_backend_dev_get(i);
-                if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
-                    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
-                    if (ggml_backend_reg_name(reg) == std::string("RPC")) {
-                        rpc_devices.push_back(dev);
-                    } else {
-                        all_devices.push_back(dev);
-                    }
+                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+                    devices.push_back(dev);
                 }
             }
-            // insert RPC devices in front
-            all_devices.insert(all_devices.begin(), rpc_devices.begin(), rpc_devices.end());
             printf("Available devices:\n");
-            for (size_t i = 0; i < all_devices.size(); ++i) {
-                auto * dev = all_devices[i];
+            for (auto * dev : devices) {
                 size_t free, total;
                 ggml_backend_dev_memory(dev, &free, &total);
                 printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
@@ -2415,7 +2548,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--cpu-moe", "-cmoe"},
         "keep all Mixture of Experts (MoE) weights in the CPU",
         [](common_params & params) {
-            params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
         }
     ).set_env("LLAMA_ARG_CPU_MOE"));
     add_opt(common_arg(
@@ -2428,7 +2561,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             for (int i = 0; i < value; ++i) {
                 // keep strings alive and avoid leaking memory by storing them in a static vector
                 static std::list<std::string> buft_overrides;
-                buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                buft_overrides.push_back(llm_ffn_exps_block_regex(i));
                 params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
             }
         }
@@ -2437,7 +2570,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--cpu-moe-draft", "-cmoed"},
         "keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
         [](common_params & params) {
-            params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+            params.speculative.tensor_buft_overrides.push_back(llm_ffn_exps_cpu_override());
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
     add_opt(common_arg(
@@ -2449,14 +2582,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
             for (int i = 0; i < value; ++i) {
                 static std::list<std::string> buft_overrides_draft;
-                buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
                 params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
             }
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
-        "number of layers to store in VRAM",
+        string_format("max. number of layers to store in VRAM (default: %d)", params.n_gpu_layers),
         [](common_params & params, int value) {
             params.n_gpu_layers = value;
             if (!llama_supports_gpu_offload()) {
@@ -2553,7 +2686,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & value) {
-            params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
+            params.lora_adapters.push_back({ std::string(value), 1.0, "", "", nullptr });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@@ -2561,7 +2694,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--lora-scaled"}, "FNAME", "SCALE",
         "path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
         [](common_params & params, const std::string & fname, const std::string & scale) {
-            params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
+            params.lora_adapters.push_back({ fname, std::stof(scale), "", "", nullptr });
         }
         // we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
     ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@@ -2614,6 +2747,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.url = value;
         }
     ).set_env("LLAMA_ARG_MODEL_URL"));
+    add_opt(common_arg(
+        { "-dr", "--docker-repo" }, "[<repo>/]<model>[:quant]",
+        "Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.\n"
+        "example: gemma3\n"
+        "(default: unused)",
+        [](common_params & params, const std::string & value) {
+            params.model.docker_repo = value;
+        }
+    ).set_env("LLAMA_ARG_DOCKER_REPO"));
     add_opt(common_arg(
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
@@ -2952,13 +3094,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.endpoint_metrics = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_METRICS"));
-    add_opt(common_arg(
-        {"--slots"},
-        string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
-        [](common_params & params) {
-            params.endpoint_slots = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
     add_opt(common_arg(
         {"--props"},
         string_format("enable changing global properties via POST /props (default: %s)", params.endpoint_props ? "enabled" : "disabled"),
@@ -2966,6 +3101,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.endpoint_props = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_PROPS"));
+    add_opt(common_arg(
+        {"--slots"},
+        string_format("enable slots monitoring endpoint (default: %s)", params.endpoint_slots ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.endpoint_slots = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ENDPOINT_SLOTS"));
     add_opt(common_arg(
         {"--no-slots"},
         "disables slots monitoring endpoint",
@@ -3124,13 +3266,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ));
-    add_opt(common_arg(
-        {"--log-colors"},
-        "Enable colored logging",
-        [](common_params &) {
-            common_log_set_colors(common_log_main(), true);
-        }
-    ).set_env("LLAMA_LOG_COLORS"));
+    add_opt(common_arg({ "--log-colors" }, "[on|off|auto]",
+                       "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
+                       "'auto' enables colors when output is to a terminal",
+                       [](common_params &, const std::string & value) {
+                           if (is_truthy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
+                           } else if (is_falsey(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
+                           } else if (is_autoy(value)) {
+                               common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
+                           } else {
+                               throw std::invalid_argument(
+                                   string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
+                           }
+                       }).set_env("LLAMA_LOG_COLORS"));
     add_opt(common_arg(
         {"-v", "--verbose", "--log-verbose"},
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
@@ -3457,8 +3607,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
             params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
             params.port = 8012;
-            params.n_gpu_layers = 99;
-            params.flash_attn = true;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
             params.n_ctx = 0;
@@ -3473,8 +3621,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
             params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
             params.port = 8012;
-            params.n_gpu_layers = 99;
-            params.flash_attn = true;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
             params.n_ctx = 0;
@@ -3489,8 +3635,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
             params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
             params.port = 8012;
-            params.n_gpu_layers = 99;
-            params.flash_attn = true;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
             params.n_ctx = 0;
@@ -3506,10 +3650,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
             params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
             params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
-            params.speculative.n_gpu_layers = 99;
             params.port = 8012;
-            params.n_gpu_layers = 99;
-            params.flash_attn = true;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
             params.n_ctx = 0;
@@ -3525,10 +3666,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
             params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
             params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
-            params.speculative.n_gpu_layers = 99;
             params.port = 8012;
-            params.n_gpu_layers = 99;
-            params.flash_attn = true;
+            params.n_ubatch = 1024;
+            params.n_batch = 1024;
+            params.n_ctx = 0;
+            params.n_cache_reuse = 256;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+
+    add_opt(common_arg(
+        {"--fim-qwen-30b-default"},
+        string_format("use default Qwen 3 Coder 30B A3B Instruct (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF";
+            params.model.hf_file = "qwen3-coder-30b-a3b-instruct-q8_0.gguf";
+            params.port = 8012;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
             params.n_ctx = 0;

common/chat.cpp

@@ -147,6 +147,7 @@ struct templates_params {
     json extra_context;
     bool add_bos;
     bool add_eos;
+    bool is_inference = true;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -162,6 +163,19 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
     throw std::runtime_error("Invalid tool_choice: " + tool_choice);
 }
 
+bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates) {
+    common_chat_templates_inputs dummy_inputs;
+    common_chat_msg msg;
+    msg.role = "user";
+    msg.content = "test";
+    dummy_inputs.messages = {msg};
+    dummy_inputs.enable_thinking = false;
+    const auto rendered_no_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
+    dummy_inputs.enable_thinking = true;
+    const auto rendered_with_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
+    return rendered_no_thinking.prompt != rendered_with_thinking.prompt;
+}
+
 template <>
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {
     std::vector<common_chat_msg> msgs;
@@ -617,10 +631,13 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
+        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1: return "DeepSeek V3.1";
         case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
         case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
         case COMMON_CHAT_FORMAT_GRANITE: return "Granite";
         case COMMON_CHAT_FORMAT_GPT_OSS: return "GPT-OSS";
+        case COMMON_CHAT_FORMAT_SEED_OSS: return "Seed-OSS";
+        case COMMON_CHAT_FORMAT_NEMOTRON_V2: return "Nemotron V2";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -682,11 +699,13 @@ static void parse_json_tool_calls(
         size_t from = std::string::npos;
         auto first = true;
         while (true) {
+            auto start_pos = builder.pos();
             auto res = function_regex_start_only && first
                 ? builder.try_consume_regex(*function_regex_start_only)
                 : function_regex
                     ? builder.try_find_regex(*function_regex, from)
                     : std::nullopt;
+
             if (res) {
                 std::string name;
                 if (get_function_name) {
@@ -721,6 +740,8 @@ static void parse_json_tool_calls(
                     return;
                 }
                 throw common_chat_msg_partial_exception("incomplete tool call");
+            } else {
+                builder.move_to(start_pos);
             }
             break;
         }
@@ -1182,6 +1203,67 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     });
     return data;
 }
+
+static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // Generate the prompt using the apply() function with the template
+    data.prompt = apply(tmpl, inputs);
+    data.format = COMMON_CHAT_FORMAT_NEMOTRON_V2;
+
+    // Handle thinking tags appropriately based on inputs.enable_thinking
+    if (string_ends_with(data.prompt, "<think>\n")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+
+    // When tools are present, build grammar for the <TOOLCALL> format, similar to CommandR, but without tool call ID
+    if (!inputs.tools.is_null() && inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = true;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            auto schemas = json::array();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                schemas.push_back({
+                    { "type",       "object"                                                   },
+                    { "properties",
+                        {
+                            { "name",
+                            {
+                                { "type", "string" },
+                                { "const", function.at("name") },
+                            } },
+                            { "arguments", function.at("parameters") },
+                        }                                                                        },
+                    { "required",   json::array({ "name", "arguments" }) },
+                });
+            });
+            auto schema = json{
+                        { "type",     "array"                                                         },
+                        { "items",    schemas.size() == 1 ? schemas[0] : json{ { "anyOf", schemas } } },
+                        { "minItems", 1                                                               },
+            };
+            if (!inputs.parallel_tool_calls) {
+                schema["maxItems"] = 1;
+            }
+            builder.add_rule("root",
+                                std::string(data.thinking_forced_open ? "( \"</think>\" space )? " : "") +
+                                    "\"<TOOLCALL>\" " + builder.add_schema("tool_calls", schema) +
+                                    " \"</TOOLCALL>\"");
+        });
+        data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+            // If thinking_forced_open, then we capture the </think> tag in the grammar,
+            // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
+            std::string(data.thinking_forced_open ?
+                            "[\\s\\S]*?(</think>\\s*)" :
+                            "(?:<think>[\\s\\S]*?</think>\\s*)?") +
+                "(<TOOLCALL>)[\\s\\S]*" });
+    }
+    return data;
+}
 static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
     if (!builder.syntax().parse_tool_calls) {
         builder.add_content(builder.consume_rest());
@@ -1311,6 +1393,71 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
     }
     return data;
 }
+
+static common_chat_params common_chat_params_init_deepseek_v3_1(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // Pass thinking context for DeepSeek V3.1 template
+    json additional_context = {
+        {"thinking", inputs.enable_thinking},
+    };
+
+    auto prompt = apply(tmpl, inputs,
+                       /* messages_override= */ inputs.messages,
+                       /* tools_override= */ std::nullopt,
+                       additional_context);
+    data.prompt = prompt;
+    data.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    if (string_ends_with(data.prompt, "<think>")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null();
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+                tool_rules.push_back(builder.add_rule(name + "-call",
+                    "( \"<｜tool▁call▁begin｜>\" )? \"" + name + "<｜tool▁sep｜>"
+                    "\" " + builder.add_schema(name + "-args", parameters) + " "
+                    "\"<｜tool▁call▁end｜>\""));
+            });
+            // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag,
+            // so we accept common variants (then it's all constrained)
+            builder.add_rule("root",
+                std::string(data.thinking_forced_open ? "( \"</think>\" space )? " : "") +
+                "( \"<｜tool▁calls▁begin｜>\" | \"<｜tool_calls_begin｜>\" | \"<｜tool calls begin｜>\" | \"<｜tool\\\\_calls\\\\_begin｜>\" | \"<｜tool▁calls｜>\" ) "
+                "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " "
+                "\"<｜tool▁calls▁end｜>\""
+                " space");
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                // If thinking_forced_open, then we capture the </think> tag in the grammar,
+                // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
+                std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)" : "(?:<think>[\\s\\S]*?</think>\\s*)?") +
+                    "(<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)[\\s\\S]*"
+            });
+            data.preserved_tokens = {
+                "<think>",
+                "</think>",
+                "<｜tool▁calls▁begin｜>",
+                "<｜tool▁call▁begin｜>",
+                "<｜tool▁sep｜>",
+                "<｜tool▁call▁end｜>",
+                "<｜tool▁calls▁end｜>",
+            };
+        });
+    }
+    return data;
+}
+
 static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
     builder.try_parse_reasoning("<think>", "</think>");
     if (!builder.syntax().parse_tool_calls) {
@@ -1332,10 +1479,81 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
         tool_calls_end);
 }
 
+static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) {
+    static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)");
+
+    static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>");
+    static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)");
+    static const common_regex tool_calls_end("<｜tool▁calls▁end｜>");
+
+    if (!builder.syntax().parse_tool_calls) {
+        LOG_DBG("%s: not parse_tool_calls\n", __func__);
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    LOG_DBG("%s: parse_tool_calls\n", __func__);
+
+    parse_json_tool_calls(
+        builder,
+        /* block_open= */ tool_calls_begin,
+        /* function_regex_start_only= */ std::nullopt,
+        function_regex,
+        close_regex,
+        tool_calls_end);
+}
+
+static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
+    // DeepSeek V3.1 outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
+    // First try to parse using the standard reasoning parsing method
+    LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
+
+    auto start_pos = builder.pos();
+    auto found_end_think = builder.try_find_literal("</think>");
+    builder.move_to(start_pos);
+
+    if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
+        LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else if (builder.try_parse_reasoning("<think>", "</think>")) {
+        // If reasoning was parsed successfully, the remaining content is regular content
+        LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
+        // </think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜>
+        common_chat_parse_deepseek_v3_1_content(builder);
+    } else {
+        if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
+          LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
+          common_chat_parse_deepseek_v3_1_content(builder);
+          return;
+        }
+        // If no reasoning tags found, check if we should treat everything as reasoning
+        if (builder.syntax().thinking_forced_open) {
+            // If thinking is forced open but no tags found, treat everything as reasoning
+            LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
+            builder.add_reasoning_content(builder.consume_rest());
+        } else {
+            LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
+            // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜>
+            common_chat_parse_deepseek_v3_1_content(builder);
+        }
+    }
+}
+
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     auto prompt = apply(tmpl, inputs);
 
+    // Check if we need to replace the return token with end token during
+    // inference and without generation prompt. For more details see:
+    // https://github.com/ggml-org/llama.cpp/issues/15417
+    if (inputs.is_inference && !inputs.add_generation_prompt) {
+        static constexpr std::string_view return_token = "<|return|>";
+        static constexpr std::string_view end_token    = "<|end|>";
+        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
+            prompt.replace(pos, return_token.length(), end_token);
+        }
+    }
+
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_GPT_OSS;
 
@@ -1349,6 +1567,26 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         "<|end|>",
     };
 
+    if (!inputs.json_schema.is_null()) {
+        data.grammar_lazy = false;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            auto schema = inputs.json_schema;
+            builder.resolve_refs(schema);
+
+            auto not_end = builder.add_rule("not-end",
+                "[^<] | \"<\" [^|] | \"<|\" [^e] | \"<|e\" [^n] | \"<|en\" [^d] | \"<|end\" [^|] | \"<|end|\" [^>]");
+            auto analysis = builder.add_rule("analysis",
+                "\"<|channel|>analysis<|message|>\" ( " + not_end + " )* \"<|end|>\"");
+            auto constraint = builder.add_rule("constraint", "\"<|constrain|>\"? [a-zA-Z0-9_-]+");
+            auto final = builder.add_rule("final",
+                "\"<|channel|>final\" ( \" \" " + constraint + " )? \"<|message|>\" " +
+                builder.add_schema("response", schema)
+            );
+
+            builder.add_rule("root", "( " + analysis + " \"<|start|>assistant\" )? " + final);
+        });
+    }
+
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@@ -1797,7 +2035,7 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
                 // If thinking_forced_open, then we capture the </think> tag in the grammar,
                 // (important for required tool choice) and in the trigger's first capture (decides what is sent to the grammar)
                 std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)" : "(?:<think>[\\s\\S]*?</think>\\s*)?") + (
-                    "(\\s*"
+                    "\\s*("
                     "(?:<tool_call>"
                     "|<function"
                     "|(?:```(?:json|xml)?\n\\s*)?(?:<function_call>|<tools>|<xml><json>|<response>)?"
@@ -2027,6 +2265,121 @@ static void common_chat_parse_granite(common_chat_msg_parser & builder) {
     }
 }
 
+static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) {
+    // Parse thinking tags
+    builder.try_parse_reasoning("<think>", "</think>");
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // Look for tool calls
+    static const common_regex tool_call_regex(regex_escape("<TOOLCALL>"));
+    if (auto res = builder.try_find_regex(tool_call_regex)) {
+        builder.move_to(res->groups[0].end);
+
+        // Expect JSON array of tool calls
+        auto tool_calls_data = builder.consume_json();
+        if (tool_calls_data.json.is_array()) {
+            if (!builder.try_consume_literal("</TOOLCALL>")) {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+            builder.add_tool_calls(tool_calls_data.json);
+        } else {
+            throw common_chat_msg_partial_exception("Incomplete tool call");
+        }
+    }
+    builder.add_content(builder.consume_rest());
+}
+
+static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) {
+    // Parse thinking tags first - this handles the main reasoning content
+    builder.try_parse_reasoning("<seed:think>", "</seed:think>");
+
+    if (!builder.syntax().parse_tool_calls) {
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    // Parse tool calls - Seed-OSS uses <seed:tool_call> format
+    static const common_regex tool_call_begin_regex("<seed:tool_call>");
+    static const common_regex tool_call_end_regex("</seed:tool_call>");
+    static const common_regex function_regex("<function=([^>]+)>");
+    static const common_regex param_regex("<parameter=([^>]+)>");
+
+    while (auto tool_res = builder.try_find_regex(tool_call_begin_regex)) {
+        builder.consume_spaces();  // Consume whitespace after <seed:tool_call>
+
+        // Look for function call inside tool call, ignore any content before it
+        if (auto func_res = builder.try_find_regex(function_regex, std::string::npos, false)) {
+            auto function_name = builder.str(func_res->groups[1]);
+
+            // Parse Seed-OSS parameters <parameter=name>value</parameter>
+            json args = json::object();
+            // Parse all parameters
+            while (auto param_res = builder.try_find_regex(param_regex, std::string::npos, false)) {
+                // again, ignore noise around parameters
+                auto param_name = builder.str(param_res->groups[1]);
+                builder.move_to(param_res->groups[0].end);
+                builder.consume_spaces();  // Consume whitespace after parameter
+                auto savedPos = builder.pos();
+                if (auto param_parse = builder.try_find_literal("</parameter>")) {
+                    auto param = param_parse->prelude;
+                    builder.move_to(savedPos);
+                    try {
+                        if (auto param_res = builder.try_consume_json()) {
+                            args[param_name] = param_res->json;
+                        } else {
+                            args[param_name] = param;
+                        }
+                    } catch (json::exception &) {
+                        args[param_name] = param;
+                    }
+                } else {
+                    throw common_chat_msg_partial_exception("Incomplete tool parameter");
+                }
+            }
+            // Look for closing function tag
+            auto end_func = builder.try_find_literal("</function>");
+            if (end_func) {
+                builder.move_to(end_func->groups[0].end);
+                builder.consume_spaces();  // Consume whitespace after </function>
+
+                // Add the tool call with parsed arguments, but only if we REALLY got the literal
+                auto eaten_fragment = builder.input().substr(end_func->groups[0].begin, end_func->groups[0].end);
+                auto funlen = std::string("</function>").length();
+                if (eaten_fragment.length() >= funlen && eaten_fragment.substr(0, funlen) == std::string("</function>")) {
+                    if (!builder.add_tool_call(function_name, "", args.dump())) {
+                        throw common_chat_msg_partial_exception("Incomplete tool call");
+                    }
+                } else {
+                    throw common_chat_msg_partial_exception("Incomplete tool call");
+                }
+            } else {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+            // Look for closing tool call tag
+            if (auto end_tool = builder.try_find_regex(tool_call_end_regex, std::string::npos, false)) {
+                builder.move_to(end_tool->groups[0].end);
+                builder.consume_spaces();  // Consume trailing whitespace after tool call
+            } else {
+                throw common_chat_msg_partial_exception("Incomplete tool call");
+            }
+        } else {
+            // No function found - don't consume content here, let it be handled at the end
+            break;
+        }
+    }
+
+    // Consume any remaining whitespace after all tool call processing
+    builder.consume_spaces();
+    auto remaining = builder.consume_rest();
+    // If there's any non-whitespace content remaining, add it as content
+    if (!string_strip(remaining).empty()) {
+        builder.add_content(remaining);
+    }
+}
+
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2043,8 +2396,62 @@ static common_chat_params common_chat_params_init_without_tools(const common_cha
     return data;
 }
 
+static common_chat_params common_chat_params_init_seed_oss(
+    const common_chat_template         & tmpl,
+    templates_params                   & params,
+    const common_chat_templates_inputs & inputs)
+{
+    common_chat_params data;
+    data.prompt = apply(tmpl, params);
+    data.format = COMMON_CHAT_FORMAT_SEED_OSS;
+    if (string_ends_with(data.prompt, "<seed:think>")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</seed:think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+
+    if (params.tools.is_array() && !params.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(params.tools, [&](const json & tool) {
+                const auto & function   = tool.at("function");
+                std::string  name       = function.at("name");
+                auto         parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                // Create rule for Seed-OSS function call format
+                std::string param_rules;
+                if (parameters.contains("properties")) {
+                    for (const auto & [key, value] : parameters.at("properties").items()) {
+                        param_rules += "\"<parameter=" + key + ">\"" + builder.add_schema(name + "-arg-" + key, value) +
+                                       "\"</parameter>\"";
+                    }
+                }
+
+                tool_rules.push_back(builder.add_rule(name + "-call",
+                                                      "\"<seed:tool_call>\" space \"<function=" + name + ">\" space " +
+                                                          param_rules +
+                                                          " \"</function>\" space \"</seed:tool_call>\""));
+            });
+
+            data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<seed:tool_call>" });
+
+            data.preserved_tokens = {
+                "<seed:think>", "</seed:think>", "<seed:tool_call>", "</seed:tool_call>",
+                "<function=",   "</function>",   "<parameter=",      "</parameter>",
+            };
+
+            builder.add_rule("root", string_join(tool_rules, " | "));
+        });
+    }
+    return data;
+}
+
 static common_chat_params common_chat_templates_apply_jinja(
-    const struct common_chat_templates * tmpls,
+    const struct common_chat_templates        * tmpls,
     const struct common_chat_templates_inputs & inputs)
 {
     templates_params params;
@@ -2088,6 +2495,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         }
     }
 
+    // DeepSeek V3.1: detect based on specific patterns in the template
+    if (src.find("message['prefix'] is defined and message['prefix'] and thinking") != std::string::npos &&
+        params.json_schema.is_null()) {
+        return common_chat_params_init_deepseek_v3_1(tmpl, params);
+    }
+
     // DeepSeek R1: use handler in all cases except json schema (thinking / tools).
     if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_deepseek_r1(tmpl, params);
@@ -2109,10 +2522,20 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // GPT-OSS
-    if (src.find("<|channel|>") != std::string::npos && params.json_schema.is_null()) {
+    if (src.find("<|channel|>") != std::string::npos) {
         return common_chat_params_init_gpt_oss(tmpl, params);
     }
 
+    // Seed-OSS
+    if (src.find("<seed:think>") != std::string::npos) {
+        return common_chat_params_init_seed_oss(tmpl, params, inputs);
+    }
+
+    // Nemotron v2
+    if (src.find("<SPECIAL_10>") != std::string::npos) {
+        return common_chat_params_init_nemotron_v2(tmpl, params);
+    }
+
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -2250,6 +2673,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
             common_chat_parse_deepseek_r1(builder);
             break;
+        case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1:
+            common_chat_parse_deepseek_v3_1(builder);
+            break;
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
             common_chat_parse_functionary_v3_2(builder);
             break;
@@ -2271,6 +2697,12 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_GPT_OSS:
             common_chat_parse_gpt_oss(builder);
             break;
+        case COMMON_CHAT_FORMAT_SEED_OSS:
+            common_chat_parse_seed_oss(builder);
+            break;
+        case COMMON_CHAT_FORMAT_NEMOTRON_V2:
+            common_chat_parse_nemotron_v2(builder);
+            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.h

@@ -107,10 +107,13 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
+    COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
     COMMON_CHAT_FORMAT_GRANITE,
     COMMON_CHAT_FORMAT_GPT_OSS,
+    COMMON_CHAT_FORMAT_SEED_OSS,
+    COMMON_CHAT_FORMAT_NEMOTRON_V2,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
@@ -197,6 +200,8 @@ common_chat_msg           common_chat_parse(const std::string & input, bool is_p
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
+bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
+
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 // T can be std::string containing JSON or nlohmann::ordered_json
 template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);

common/common.cpp

@@ -558,13 +558,6 @@ std::string string_from(const struct llama_context * ctx, const std::vector<llam
 
         auto detokenized = common_token_to_piece(ctx, token);
 
-        detokenized.erase(
-            std::remove_if(
-                detokenized.begin(),
-                detokenized.end(),
-                [](const unsigned char c) { return !std::isprint(c); }),
-            detokenized.end());
-
         buf << "'" << detokenized << "'"
             << ":" << std::to_string(token);
     }
@@ -589,13 +582,6 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
         auto detokenized = common_token_to_piece(ctx, batch.token[i]);
 
-        detokenized.erase(
-                std::remove_if(
-                    detokenized.begin(),
-                    detokenized.end(),
-                    [](const unsigned char c) { return !std::isprint(c); }),
-                detokenized.end());
-
         buf << "\n"          << std::to_string(i)
             << ", token '"   << detokenized << "'"
             << ", pos "      << std::to_string(batch.pos[i])
@@ -915,7 +901,8 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
         return iparams;
     }
 
@@ -925,7 +912,8 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
+            __func__, params.model.path.c_str());
         llama_model_free(model);
         return iparams;
     }
@@ -1002,7 +990,12 @@ struct common_init_result common_init_from_params(common_params & params) {
             return iparams;
         }
 
+        char buf[1024];
         la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
         iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
@@ -1166,11 +1159,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.attention_type    = params.attention_type;
-    cparams.defrag_thold      = params.defrag_thold;
+    cparams.flash_attn_type   = params.flash_attn_type;
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
-    cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;

common/common.h

@@ -34,6 +34,9 @@ struct common_adapter_lora_info {
     std::string path;
     float scale;
 
+    std::string task_name;
+    std::string prompt_prefix;
+
     struct llama_adapter_lora * ptr;
 };
 
@@ -190,10 +193,11 @@ struct common_params_sampling {
 };
 
 struct common_params_model {
-    std::string path    = ""; // model local path                                           // NOLINT
-    std::string url     = ""; // model url to download                                      // NOLINT
-    std::string hf_repo = ""; // HF repo                                                    // NOLINT
-    std::string hf_file = ""; // HF file                                                    // NOLINT
+    std::string path        = ""; // model local path                                       // NOLINT
+    std::string url         = ""; // model url to download                                  // NOLINT
+    std::string hf_repo     = ""; // HF repo                                                // NOLINT
+    std::string hf_file     = ""; // HF file                                                // NOLINT
+    std::string docker_repo = ""; // Docker repo                                            // NOLINT
 };
 
 struct common_params_speculative {
@@ -284,11 +288,10 @@ struct common_params {
     float   rope_freq_base        =  0.0f; // RoPE base frequency
     float   rope_freq_scale       =  0.0f; // RoPE frequency scaling factor
     float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor      =  1.0f; // YaRN magnitude scaling factor
-    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow        =  1.0f; // YaRN high correction dim
+    float   yarn_attn_factor      = -1.0f; // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = -1.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = -1.0f; // YaRN high correction dim
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
-    float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
@@ -310,6 +313,7 @@ struct common_params {
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
+    enum llama_flash_attn_type   flash_attn_type   = LLAMA_FLASH_ATTN_TYPE_AUTO; // whether to use Flash Attention
 
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
@@ -373,9 +377,8 @@ struct common_params {
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
-    bool flash_attn        = false; // flash attention
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = false;  // context shift on inifinite text generation
+    bool ctx_shift         = false;  // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
@@ -442,7 +445,7 @@ struct common_params {
 
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
-    bool endpoint_slots   = false;
+    bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
@@ -450,7 +453,7 @@ struct common_params {
 
     std::string slot_save_path;
 
-    float slot_prompt_similarity = 0.5f;
+    float slot_prompt_similarity = 0.1f;
 
     // batched-bench params
     bool is_pp_shared = false;
@@ -731,6 +734,20 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
 
+//
+// MoE utils
+//
+
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_exps";
+
+static std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
+}
+
+static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+    return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
+}
+
 //
 // training utils
 //

common/json-schema-to-grammar.cpp

@@ -257,12 +257,13 @@ std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
 };
 
 static bool is_reserved_name(const std::string & name) {
-    static std::unordered_set<std::string> RESERVED_NAMES;
-    if (RESERVED_NAMES.empty()) {
-        RESERVED_NAMES.insert("root");
-        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
-    }
+    static const std::unordered_set<std::string> RESERVED_NAMES = [] {
+        std::unordered_set<std::string> s;
+        s.insert("root");
+        for (const auto & p : PRIMITIVE_RULES) s.insert(p.first);
+        for (const auto & p : STRING_FORMAT_RULES) s.insert(p.first);
+        return s;
+    }();
     return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }
 
@@ -843,9 +844,10 @@ public:
                 _build_object_rule(
                     properties, required, name,
                     schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+        } else if ((schema_type.is_null() || schema_type == "object" || schema_type == "string") && schema.contains("allOf")) {
             std::unordered_set<std::string> required;
             std::vector<std::pair<std::string, json>> properties;
+            std::map<std::string, size_t> enum_values;
             std::string hybrid_name = name;
             std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
                 if (comp_schema.contains("$ref")) {
@@ -857,6 +859,14 @@ public:
                             required.insert(prop.key());
                         }
                     }
+                } else if (comp_schema.contains("enum")) {
+                    for (const auto & v : comp_schema["enum"]) {
+                        const auto rule = _generate_constant_rule(v);
+                        if (enum_values.find(rule) == enum_values.end()) {
+                            enum_values[rule] = 0;
+                        }
+                        enum_values[rule] += 1;
+                    }
                 } else {
                   // todo warning
                 }
@@ -870,6 +880,17 @@ public:
                     add_component(t, true);
                 }
             }
+            if (!enum_values.empty()) {
+                std::vector<std::string> enum_intersection;
+                for (const auto & p : enum_values) {
+                    if (p.second == schema["allOf"].size()) {
+                        enum_intersection.push_back(p.first);
+                    }
+                }
+                if (!enum_intersection.empty()) {
+                    return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ") space");
+                }
+            }
             return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
         } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
             json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];

common/log.cpp

@@ -4,17 +4,52 @@
 #include <condition_variable>
 #include <cstdarg>
 #include <cstdio>
+#include <cstdlib>
+#include <cstring>
 #include <mutex>
 #include <sstream>
 #include <thread>
 #include <vector>
 
+#if defined(_WIN32)
+#    include <io.h>
+#    include <windows.h>
+#    define isatty _isatty
+#    define fileno _fileno
+#else
+#    include <unistd.h>
+#endif // defined(_WIN32)
+
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
 void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
+// Auto-detect if colors should be enabled based on terminal and environment
+static bool common_log_should_use_colors_auto() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
+
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -353,6 +388,11 @@ struct common_log * common_log_init() {
 
 struct common_log * common_log_main() {
     static struct common_log log;
+    static std::once_flag    init_flag;
+    std::call_once(init_flag, [&]() {
+        // Set default to auto-detect colors
+        log.set_colors(common_log_should_use_colors_auto());
+    });
 
     return &log;
 }
@@ -380,8 +420,19 @@ void common_log_set_file(struct common_log * log, const char * file) {
     log->set_file(file);
 }
 
-void common_log_set_colors(struct common_log * log, bool colors) {
-    log->set_colors(colors);
+void common_log_set_colors(struct common_log * log, log_colors colors) {
+    if (colors == LOG_COLORS_AUTO) {
+        log->set_colors(common_log_should_use_colors_auto());
+        return;
+    }
+
+    if (colors == LOG_COLORS_DISABLED) {
+        log->set_colors(false);
+        return;
+    }
+
+    GGML_ASSERT(colors == LOG_COLORS_ENABLED);
+    log->set_colors(true);
 }
 
 void common_log_set_prefix(struct common_log * log, bool prefix) {

common/log.h

@@ -24,6 +24,12 @@
 #define LOG_DEFAULT_DEBUG 1
 #define LOG_DEFAULT_LLAMA 0
 
+enum log_colors {
+    LOG_COLORS_AUTO     = -1,
+    LOG_COLORS_DISABLED = 0,
+    LOG_COLORS_ENABLED  = 1,
+};
+
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
 extern int common_log_verbosity_thold;
@@ -65,10 +71,10 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
 //
 
-void common_log_set_file      (struct common_log * log, const char * file);       // not thread-safe
-void common_log_set_colors    (struct common_log * log,       bool   colors);     // not thread-safe
-void common_log_set_prefix    (struct common_log * log,       bool   prefix);     // whether to output prefix to each log
-void common_log_set_timestamps(struct common_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
+void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
+void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
+void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold

common/sampling.cpp

@@ -426,8 +426,29 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
 
 // helpers
 
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
-    return &gsmpl->cur_p;
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort) {
+    auto * res = &gsmpl->cur_p;
+
+    if (do_sort && !res->sorted) {
+        // remember the selected token before sorting
+        const llama_token id = res->data[res->selected].id;
+
+        std::sort(res->data, res->data + res->size, [](const llama_token_data & a, const llama_token_data & b) {
+            return a.p > b.p;
+        });
+
+        // restore the selected token after sorting
+        for (size_t i = 0; i < res->size; ++i) {
+            if (res->data[i].id == id) {
+                res->selected = i;
+                break;
+            }
+        }
+
+        res->sorted = true;
+    }
+
+    return res;
 }
 
 llama_token common_sampler_last(const struct common_sampler * gsmpl) {

common/sampling.h

@@ -86,7 +86,9 @@ uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 // helpers
 
 // access the internal list of current candidate tokens
-llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
+// if do_sort == true, the candidates are guaranteed to be sorted afterwards (in descending order of probability)
+// the .sorted flag of the result indicates whether the returned candidates are sorted
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl, bool do_sort);
 
 // get the last accepted token
 llama_token common_sampler_last(const struct common_sampler * gsmpl);

common/speculative.cpp

@@ -317,7 +317,7 @@ llama_tokens common_speculative_gen_draft(
 
         common_sampler_sample(smpl, ctx_dft, 0, true);
 
-        const auto * cur_p = common_sampler_get_candidates(smpl);
+        const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
         for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
             LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",

convert_hf_to_gguf.py

@@ -72,6 +72,7 @@ class ModelBase:
     endianess: gguf.GGUFEndian
     use_temp_file: bool
     lazy: bool
+    dry_run: bool
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
@@ -89,13 +90,16 @@ class ModelBase:
     block_count: int
     tensor_map: gguf.TensorNameMap
 
+    # Mistral format specifics
     is_mistral_format: bool = False
+    disable_mistral_community_chat_template: bool = False
 
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
-                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None):
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
+                 disable_mistral_community_chat_template: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -108,6 +112,7 @@ class ModelBase:
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
         self.use_temp_file = use_temp_file
         self.lazy = not eager or (remote_hf_model_id is not None)
+        self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
         if remote_hf_model_id is not None:
             self.is_safetensors = True
@@ -147,6 +152,9 @@ class ModelBase:
         self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                            split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
 
+        # Mistral specific
+        self.disable_mistral_community_chat_template = disable_mistral_community_chat_template
+
     @classmethod
     def add_prefix_to_filename(cls, path: Path, prefix: str) -> Path:
         stem, suffix = path.stem, path.suffix
@@ -294,10 +302,6 @@ class ModelBase:
                 # data = data_torch.squeeze().numpy()
                 data = data_torch.numpy()
 
-                # if data ends up empty, it means data_torch was a scalar tensor -> restore
-                if len(data.shape) == 0:
-                    data = data_torch.numpy()
-
                 n_dims = len(data.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
@@ -731,6 +735,9 @@ class TextModel(ModelBase):
         if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c":
             # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B
             res = "qwen2"
+        if chkhsh == "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273":
+            # ref: https://huggingface.co/alvarobartt/grok-2-tokenizer
+            res = "grok-2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -881,6 +888,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1e163ecab2e718a4c829d1148b6e86824ec36163bb71941c3dca9cd5ac25756":
             # ref: https://huggingface.co/JetBrains/Mellum-4b-base
             res = "mellum"
+        if chkhsh == "9b1be57e70d20d9501b2b3186e792d81181ae36ada3903c26f9fea418cf87206":
+            # ref: https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base
+            res = "llada-moe"
 
         if res is None:
             logger.warning("\n")
@@ -1210,6 +1220,55 @@ class TextModel(ModelBase):
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
 
+    def _set_vocab_interns1(self):
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
+        vocab_size = self.hparams.get("vocab_size", len(vocab))
+        assert max(vocab.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        added_tokens_decoder = tokenizer.added_tokens_decoder
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
+                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
+                    if not added_tokens_decoder[i].normalized:
+                        previous_token = token
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
+                        if previous_token != token:
+                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
+
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab._set_special_token("bos", 151643)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
 
 class MmprojModel(ModelBase):
     model_type = ModelType.MMPROJ
@@ -2011,8 +2070,17 @@ class LlamaModel(TextModel):
 
         template_dir = Path(__file__).parent / "models/templates/"
 
-        template = MistralModel.get_community_chat_template(vocab, template_dir)
-        self.gguf_writer.add_chat_template(template)
+        if not self.is_mistral_format or not self.disable_mistral_community_chat_template:
+            # Log only for Mistral format that the official tokenization and detokenization is via `mistral-common`.
+            if self.is_mistral_format:
+                logger.info(
+                    "Using a Mistral community chat template. These templates can be subject to errors in early days or weeks after a release. "
+                    "Mistral recommends to use `mistral-common` to perform tokenization and detokenization."
+                )
+            template = MistralModel.get_community_chat_template(vocab, template_dir, self.is_mistral_format)
+            self.gguf_writer.add_chat_template(template)
+        else:
+            logger.info("Not using a Mistral community chat template. Ensure to perform the tokenization and detokenization via `mistral-common`.")
 
     def set_vocab(self):
         if self.is_mistral_format:
@@ -2620,12 +2688,20 @@ class BitnetModel(TextModel):
         yield (new_name, data_torch)
 
 
-@ModelBase.register("GrokForCausalLM")
+@ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")
 class GrokModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GROK
 
     def set_vocab(self):
-        self._set_vocab_sentencepiece()
+        if (self.dir_model / 'tokenizer.model').is_file():
+            self._set_vocab_sentencepiece()
+            return
+
+        if not (self.dir_model / 'tokenizer.json').is_file() or not (self.dir_model / 'chat_template.jinja').is_file():
+            logger.error('Error: Missing vocab and chat template, download files from https://huggingface.co/alvarobartt/grok-2-tokenizer')
+            sys.exit(1)
+
+        self._set_vocab_gpt2()
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -2633,11 +2709,46 @@ class GrokModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-    _experts: list[dict[str, Tensor]] | None = None
+        self.gguf_writer.add_attn_logit_softcapping(self.hparams.get("attn_logit_softcapping", 30.0))
+        self.gguf_writer.add_router_logit_softcapping(self.hparams.get("router_logit_softcapping", 30.0))
+        if (final_logit_softcap := self.hparams.get("final_logit_softcapping")):
+            self.gguf_writer.add_final_logit_softcapping(final_logit_softcap)
+
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+
+        # Treat "original" as "yarn", seems to have been a mistake
+        if self.hparams.get("rope_type") in ("yarn", "original"):
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(self.hparams["scaling_factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["original_max_position_embeddings"])
+            self.gguf_writer.add_rope_scaling_yarn_ext_factor(self.hparams["extrapolation_factor"])
+            self.gguf_writer.add_rope_scaling_yarn_attn_factor(self.hparams["attn_factor"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_fast(self.hparams["beta_fast"])
+            self.gguf_writer.add_rope_scaling_yarn_beta_slow(self.hparams["beta_slow"])
+
+        if temp_len := self.hparams.get("attn_temperature_len"):
+            self.gguf_writer.add_attn_temperature_length(temp_len)
+
+        self.gguf_writer.add_attn_output_scale(self.hparams.get("attn_output_multiplier", rope_dim**-0.5))
+        self.gguf_writer.add_embedding_scale(self.hparams["embedding_multiplier_scale"])
+        self.gguf_writer.add_logit_scale(self.hparams["output_multiplier_scale"])
+
+    _experts: list[dict[str, list[Tensor]]] | None = None
+    _cur_expert = ""
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        tensors: list[tuple[str, Tensor]] = []
+        is_expert = ".moe." in name or ".block_sparse_moe.experts." in name
+
+        if not is_expert:
+            tensors.append((self.map_tensor_name(name), data_torch))
+
         # process the experts separately
-        if name.find(".moe.") != -1:
+        if is_expert or self._cur_expert:
             n_experts = self.hparams["num_local_experts"]
 
             assert bid is not None
@@ -2645,32 +2756,41 @@ class GrokModel(TextModel):
             if self._experts is None:
                 self._experts = [{} for _ in range(self.block_count)]
 
-            self._experts[bid][name] = data_torch
-
-            if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
-                # merge the experts into a single 3d tensor
-                for wid in ["linear", "linear_1", "linear_v"]:
-                    datas: list[Tensor] = []
-
-                    for xid in range(n_experts):
-                        ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
-                        datas.append(self._experts[bid][ename])
-                        del self._experts[bid][ename]
-
-                    data_torch = torch.stack(datas, dim=0)
-
-                    merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
-
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
-            else:
+            # concatenate split tensors
+            if name in self._experts[bid]:
+                self._cur_expert = name
+                self._experts[bid][name].append(data_torch)
                 return []
+            elif is_expert:
+                self._cur_expert = name
+                self._experts[bid][name] = [data_torch]
+                return []
+            else:
+                self._cur_expert = ""
 
-        return [(self.map_tensor_name(name), data_torch)]
+            for bid in range(self.block_count):
+                if len(self._experts[bid]) >= n_experts * 3:
+                    # merge the experts into a single 3d tensor
+                    for wid in [("linear", "w1", 0), ("linear_1", "w2", 1), ("linear_v", "w3", 0)]:
+                        datas: list[Tensor] = []
+
+                        for xid in range(n_experts):
+                            ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid[0]}.weight"
+                            if ename not in self._experts[bid]:
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid[1]}.weight"
+                            tensor_list = self._experts[bid][ename]
+                            datas.append(torch.cat(tensor_list, dim=wid[2]) if len(tensor_list) > 1 else tensor_list[0])
+                            del self._experts[bid][ename]
+
+                        data_torch = torch.stack(datas, dim=0)
+
+                        merged_name = f"transformer.decoder_layer.{bid}.moe.{wid[0]}.weight"
+
+                        new_name = self.map_tensor_name(merged_name)
+
+                        yield (new_name, data_torch)
+
+        yield from tensors
 
 
 @ModelBase.register("DbrxForCausalLM")
@@ -2917,7 +3037,8 @@ class Qwen2Model(TextModel):
         if "language_model." in name:
             name = name.replace("language_model.", "") # for InternVL
         if name.startswith("mlp") or name.startswith("multi_modal_projector") \
-                or name.startswith("vision_model") or name.startswith("audio_tower"):
+                or name.startswith("vision_model") or name.startswith("audio_tower") \
+                or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
             # skip vision and audio tensors
             return []
         yield from super().modify_tensors(data_torch, name, bid)
@@ -3094,7 +3215,7 @@ class LLaDAModel(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Ernie4_5_ForCausalLM")
+@ModelBase.register("Ernie4_5_ForCausalLM", "Ernie4_5ForCausalLM")
 class Ernie4_5Model(TextModel):
     model_arch = gguf.MODEL_ARCH.ERNIE4_5
 
@@ -3589,6 +3710,19 @@ class Qwen2MoeModel(TextModel):
 class Qwen3Model(Qwen2Model):
     model_arch = gguf.MODEL_ARCH.QWEN3
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
+        self.origin_hf_arch = hparams.get('architectures', [None])[0]
+
+    def set_vocab(self):
+        # deal with intern-s1-mini
+        if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
+            self._set_vocab_interns1()
+            return
+
+        super().set_vocab()
+
 
 @ModelBase.register("Qwen3MoeForCausalLM")
 class Qwen3MoeModel(Qwen2MoeModel):
@@ -3605,73 +3739,7 @@ class Qwen3MoeModel(Qwen2MoeModel):
             self._set_vocab_interns1()
             return
 
-        try:
-            self._set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_gpt2()
-
-    def _set_vocab_interns1(self):
-        tokens: list[str] = []
-        toktypes: list[int] = []
-
-        from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
-        vocab_size = self.hparams.get("vocab_size", len(vocab))
-        assert max(vocab.values()) < vocab_size
-
-        tokpre = self.get_vocab_base_pre(tokenizer)
-
-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()
-
-        added_tokens_decoder = tokenizer.added_tokens_decoder
-
-        for i in range(vocab_size):
-            if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.UNUSED)
-            else:
-                token: str = reverse_vocab[i]
-                if token in added_vocab:
-                    # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
-                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
-                    if not added_tokens_decoder[i].normalized:
-                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
-                        if previous_token != token:
-                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
-
-                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
-                        toktypes.append(gguf.TokenType.CONTROL)
-                    else:
-                        toktypes.append(gguf.TokenType.USER_DEFINED)
-                else:
-                    toktypes.append(gguf.TokenType.NORMAL)
-                tokens.append(token)
-
-        self.gguf_writer.add_tokenizer_model("gpt2")
-        self.gguf_writer.add_tokenizer_pre(tokpre)
-        self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_types(toktypes)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        special_tokens_map_file = self.dir_model / 'special_tokens_map.json'
-        additional_special_tokens = []
-        if special_tokens_map_file.is_file():
-            with open(special_tokens_map_file, encoding = 'utf-8') as f:
-                additional_special_tokens = json.load(f).get('additional_special_tokens', [])
-        tokenizer_cfg_file = self.dir_model / 'special_tokens_map.json'
-        if tokenizer_cfg_file.is_file():
-            with open(tokenizer_cfg_file, encoding = 'utf-8') as f:
-                added_tokens_decoder = json.load(f).get('added_tokens_decoder', {})
-                token2ids_map = {data['content'] : int(token) for token, data in added_tokens_decoder.items() if data['special']}
-                for token in additional_special_tokens:
-                    if token in token2ids_map:
-                        special_vocab._set_special_token(token, token2ids_map[token])
-        special_vocab._set_special_token('eos', 151645)
-        special_vocab._set_special_token("bos", 151643)
-        special_vocab.add_to_gguf(self.gguf_writer)
+        super().set_vocab()
 
 
 @ModelBase.register("GPT2LMHeadModel")
@@ -4859,11 +4927,35 @@ class NeoBert(BertModel):
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
+    _lora_files = {}
+    _lora_names = []
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, False)
+
+        if lora_names := hparams.get("lora_adaptations"):
+            self._lora_names = lora_names
+            self.model_arch = gguf.MODEL_ARCH.JINA_BERT_V3
+
+        super().__init__(dir_model, ftype, fname_out, hparams=hparams, **kwargs)
         self._xlmroberta_tokenizer_init()
 
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        if self._lora_names:
+            for name in self._lora_names:
+                fname = self.add_prefix_to_filename(self.fname_out, f"lora-{name}-")
+                self._lora_files[name] = gguf.GGUFWriter(fname, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file, dry_run=self.dry_run)
+
+        return super().generate_extra_tensors()
+
+    def set_type(self):
+        for lora_writer in self._lora_files.values():
+            lora_writer.add_type(gguf.GGUFType.ADAPTER)
+            lora_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
+        super().set_type()
+
     def set_vocab(self):
         self._xlmroberta_set_vocab()
 
@@ -4873,13 +4965,62 @@ class XLMRobertaModel(BertModel):
         if name.startswith("roberta."):
             name = name[8:]
 
+        # jina-embeddings-v3
+        if ".parametrizations." in name:
+            name = name.replace(".parametrizations.", ".")
+            if name.endswith(".original"):
+                name = name[:-9]
+
         # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
         if name == "embeddings.position_embeddings.weight":
             if self._position_offset is not None:
                 data_torch = data_torch[self._position_offset:,:]
 
+        if name.endswith(".0.lora_A") or name.endswith(".0.lora_B"):
+            if name.startswith("pooler.dense"):
+                return []
+
+            num_loras = data_torch.size(0)
+            assert num_loras == len(self._lora_names)
+
+            # Split out each LoRA in their own GGUF
+            for i, lora_writer in enumerate(self._lora_files.values()):
+                new_name = self.map_tensor_name(name[:-9]) + name[-7:].lower()
+                data = data_torch[i, :, :]
+                # Transpose/flip token_embd/types into correct shape
+                if new_name == "token_embd.weight.lora_b":
+                    data = data.T
+                elif new_name.startswith("token_types.weight."):
+                    new_name = new_name[:-1] + ("a" if new_name[-1:] == "b" else "b")
+                lora_writer.add_tensor(new_name, data.float().numpy(), raw_dtype=gguf.GGMLQuantizationType.F32)
+
+            return []
+
         return super().modify_tensors(data_torch, name, bid)
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # jina-embeddings-v3
+        if rotary_emb_base := self.hparams.get("rotary_emb_base"):
+            self.gguf_writer.add_rope_freq_base(rotary_emb_base)
+        lora_alpha = self.hparams.get("lora_alpha")
+        if lora_prompt_prefixes := self.hparams.get("task_instructions"):
+            assert self._lora_files and all(lora_name in lora_prompt_prefixes for lora_name in self._lora_files.keys())
+        for lora_name, lora_writer in self._lora_files.items():
+            lora_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, lora_alpha if lora_alpha is not None else 1.0)
+            lora_writer.add_string(gguf.Keys.Adapter.LORA_TASK_NAME, lora_name)
+            if lora_prompt_prefixes:
+                lora_writer.add_string(gguf.Keys.Adapter.LORA_PROMPT_PREFIX, lora_prompt_prefixes[lora_name])
+
+    def write(self):
+        super().write()
+        for lora_writer in self._lora_files.values():
+            lora_writer.write_header_to_file()
+            lora_writer.write_kv_data_to_file()
+            lora_writer.write_tensors_to_file(progress=True)
+            lora_writer.close()
+
 
 @ModelBase.register("GemmaForCausalLM")
 class GemmaModel(TextModel):
@@ -5039,6 +5180,29 @@ class Gemma3Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Gemma3TextModel")
+class EmbeddingGemma(Gemma3Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDING
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Override the sliding window size as it gets adjusted by the Gemma3TextConfig
+        # constructor. We want to use the value from the original model's config.json.
+        # ref: https://github.com/huggingface/transformers/pull/40700
+        with open(self.dir_model / "config.json", "r", encoding="utf-8") as f:
+            config = json.load(f)
+            orig_sliding_window = config.get("sliding_window")
+            if orig_sliding_window is None:
+                raise ValueError("sliding_window not found in model config - this is required for the model")
+
+            logger.info(f"Using original sliding_window from config: {orig_sliding_window} "
+                        f"instead of {self.hparams['sliding_window']}")
+            self.gguf_writer.add_sliding_window(orig_sliding_window)
+
+        self._try_set_pooling_type()
+
+
 @ModelBase.register("Gemma3ForConditionalGeneration")
 class Gemma3VisionModel(MmprojModel):
     def set_gguf_parameters(self):
@@ -5839,10 +6003,40 @@ class OlmoModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("SeedOssForCausalLM")
+class SeedOssModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.SEED_OSS
+
+
 @ModelBase.register("Olmo2ForCausalLM")
+@ModelBase.register("Olmo3ForCausalLM")
 class Olmo2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.OLMO2
 
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_attn_factors(rope_scaling["attention_factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+        if "sliding_window" in self.hparams:
+            self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+
+            sliding_window_pattern = []
+            if "layer_types" in self.hparams:
+                sliding_window_pattern = [t == "sliding_attention" for t in self.hparams["layer_types"]]
+            else:
+                # Olmo2 does not use sliding window attention.
+                # Olmo3 defaults to using sliding window for all layers except every 4th.
+                for i in range(self.hparams["num_hidden_layers"]):
+                    sliding_window_pattern.append((i + 1) % 4 != 0)
+
+            self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+
 
 @ModelBase.register("OlmoeForCausalLM")
 class OlmoeModel(TextModel):
@@ -6237,9 +6431,11 @@ class DeepseekModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("DeepseekV2ForCausalLM")
-@ModelBase.register("DeepseekV3ForCausalLM")
-@ModelBase.register("KimiVLForConditionalGeneration")
+@ModelBase.register(
+    "DeepseekV2ForCausalLM",
+    "DeepseekV3ForCausalLM",
+    "KimiVLForConditionalGeneration",
+)
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
@@ -6588,6 +6784,8 @@ class T5Model(TextModel):
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
         self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        if (dec_n_layer := self.hparams.get("num_decoder_layers")) is not None:
+            self.gguf_writer.add_decoder_block_count(dec_n_layer)
         self.gguf_writer.add_head_count(self.hparams["num_heads"])
         self.gguf_writer.add_key_length(self.hparams["d_kv"])
         self.gguf_writer.add_value_length(self.hparams["d_kv"])
@@ -7452,9 +7650,13 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         ]
 
         # n_group and d_inner are used during reshape_tensors for mamba2
-        self.d_model = self.find_hparam(["hidden_size", "d_model"])
-        self.n_group = self.find_hparam(["n_groups"])
-        self.d_inner = self.find_hparam(["expand"]) * self.d_model
+        # NOTE: Explicitly include hparam prefix prefix for d_model to
+        #   disambiguate with top-level head_dim
+        # NOTE 2: If needed for future models, this can be isolated in a method
+        #   to separate the prefix setting and teh keys used
+        self.d_model = self.find_hparam([f"{self.hparam_prefixes[0]}_head_dim", "hidden_size", "d_model"])
+        self.n_group = self.find_hparam(["n_groups", "num_groups"])
+        self.d_inner = self.find_hparam(["expand", "num_heads"]) * self.d_model
 
     def get_attn_layers(self):
         # Explicit list of layer type names
@@ -7515,12 +7717,12 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
 
         ## Mamba mixer params ##
         self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"]))
-        self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state"]))
+        self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state", "state_dim", "ssm_state_size"]))
         self.gguf_writer.add_ssm_group_count(self.n_group)
         self.gguf_writer.add_ssm_inner_size(self.d_inner)
         # NOTE: The mamba_dt_rank is _not_ the right field for how this is used
         #   in llama.cpp
-        self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"]))
+        self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads", "num_heads"]))
 
         ## Attention params ##
         head_count_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
@@ -7547,6 +7749,55 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
         Mamba2Model.set_vocab(self)
 
 
+@ModelBase.register("NemotronHForCausalLM")
+class NemotronHModel(GraniteHybridModel):
+    """Hybrid mamba2/attention model from NVIDIA"""
+    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Save the top-level head_dim for later
+        self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))
+        assert self.head_dim is not None, "Could not find the attention head dim in config"
+
+        # Don't use expand to calculate d_inner
+        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model
+
+        # Update the ssm / attn / mlp layers
+        # M: Mamba2, *: Attention, -: MLP
+        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
+        self._ssm_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "M"]
+        self._mlp_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "-"]
+
+    def get_attn_layers(self):
+        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
+        assert len(hybrid_override_pattern) == self.block_count, "Mismatch between hybrid override and num_hidden_layers!"
+        return [i for i, val in enumerate(hybrid_override_pattern) if val == "*"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_key_length(self.head_dim)
+        self.gguf_writer.add_value_length(self.head_dim)
+
+        # Set feed_forward_length
+        # NOTE: This will trigger an override warning. This is preferrable to
+        #   duplicating all the parent logic
+        n_ff = self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"])
+        self.gguf_writer.add_feed_forward_length([
+            n_ff if i in self._mlp_layers else 0 for i in range(self.block_count)
+        ])
+
+    def set_vocab(self):
+        super().set_vocab()
+
+        # The tokenizer _does_ add a BOS token (via post_processor type
+        # TemplateProcessing) but does not set add_bos_token to true in the
+        # config, so we need to explicitly override it here.
+        self.gguf_writer.add_add_bos_token(True)
+
+
 @ModelBase.register("BailingMoeForCausalLM")
 class BailingMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE
@@ -8016,6 +8267,76 @@ class HunYuanMoEModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("LLaDAMoEModel", "LLaDAMoEModelLM")
+class LLaDAMoEModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LLADA_MOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+
+        if (expert_intermediate_size := self.hparams.get("expert_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(expert_intermediate_size)
+
+        # number of experts used per token (top-k)
+        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+            self.gguf_writer.add_expert_used_count(n_experts_used)
+
+        self.gguf_writer.add_mask_token_id(156895)
+        self.gguf_writer.add_causal_attention(False)
+        self.gguf_writer.add_diffusion_shift_logits(False)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    # Copied from: Qwen2MoeModel
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # Copied from: Qwen2MoeModel
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("HunYuanDenseV1ForCausalLM")
 class HunYuanModel(TextModel):
     model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
@@ -8422,7 +8743,7 @@ class MistralModel(LlamaModel):
     undo_permute = False
 
     @staticmethod
-    def get_community_chat_template(vocab: MistralVocab, templates_dir: Path):
+    def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
         assert TokenizerVersion is not None, "mistral_common is not installed"
         assert isinstance(vocab.tokenizer, (Tekkenizer, SentencePieceTokenizer)), (
             f"Expected Tekkenizer or SentencePieceTokenizer, got {type(vocab.tokenizer)}"
@@ -8443,7 +8764,13 @@ class MistralModel(LlamaModel):
         elif vocab.tokenizer.version == TokenizerVersion.v13:
             template_file = "unsloth-mistral-Devstral-Small-2507.jinja"
         else:
-            raise ValueError(f"Unknown tokenizer type: {vocab.tokenizer_type} and version {vocab.tokenizer.version}")
+            err_message = f"Unknown tokenizer type: {vocab.tokenizer_type} and version {vocab.tokenizer.version}"
+            if is_mistral_format:
+                err_message += (
+                    " . Please pass --disable-mistral-community-chat-template argument to the CLI "
+                    "if you want to skip this error and use the Mistral official `mistral-common` pre-processing library."
+                )
+            raise ValueError(err_message)
 
         template_path = templates_dir / template_file
         if not template_path.exists():
@@ -8484,6 +8811,43 @@ class PixtralModel(LlavaVisionModel):
             return "mm.2.weight"
         return super().map_tensor_name(name, try_suffixes)
 
+
+@ModelBase.register("KimiVLForConditionalGeneration")
+class KimiVLModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["image_size"] = 64 * 14 # for compatibility
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIVL)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_projector_scale_factor(2)
+        # eps is the same as pytorch's default value
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-5))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            if "pos_emb.weight" in name:
+                data_torch = data_torch.view(data_torch.shape[0] * data_torch.shape[1], data_torch.shape[2])
+            elif "wqkv" in name:
+                split_dim = 0 if "weight" in name else -1
+                wq, wk, wv = data_torch.chunk(3, dim=split_dim)
+                return [
+                    (self.map_tensor_name(name.replace("wqkv", "wq")), wq),
+                    (self.map_tensor_name(name.replace("wqkv", "wk")), wk),
+                    (self.map_tensor_name(name.replace("wqkv", "wv")), wv)
+                ]
+
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
 ###### CONVERSION LOGIC ######
 
 
@@ -8638,6 +9002,13 @@ def parse_args() -> argparse.Namespace:
         "--mistral-format", action="store_true",
         help="Whether the model is stored following the Mistral format.",
     )
+    parser.add_argument(
+        "--disable-mistral-community-chat-template", action="store_true",
+        help=(
+            "Whether to disable usage of Mistral community chat templates. If set, use the Mistral official `mistral-common` library for tokenization and detokenization of Mistral models. "
+            "Using `mistral-common` ensure correctness and zero-day support of tokenization for models converted from the Mistral format but requires to manually setup the tokenization server."
+        )
+    )
 
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
@@ -8744,6 +9115,7 @@ def main() -> None:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     is_mistral_format = args.mistral_format
+    disable_mistral_community_chat_template = args.disable_mistral_community_chat_template
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
@@ -8770,7 +9142,7 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id,
+                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template
                                      )
 
         if args.vocab_only:

convert_hf_to_gguf_update.py

@@ -139,6 +139,7 @@ models = [
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
     {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
+    {"name": "llada-moe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/LLaDA-MoE-7B-A1B-Base", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -158,6 +159,7 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
     {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
     {"name": "qwen2",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "chkhsh": "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c"},
+    {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
 ]
 
 

convert_lora_to_gguf.py

@@ -12,7 +12,7 @@ import json
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
-from transformers import AutoConfig
+from transformers import AutoConfig, AutoTokenizer
 
 import torch
 
@@ -26,6 +26,8 @@ import gguf
 # reuse model definitions from convert_hf_to_gguf.py
 from convert_hf_to_gguf import LazyTorchTensor, ModelBase
 
+from gguf.constants import GGUFValueType
+
 logger = logging.getLogger("lora-to-gguf")
 
 
@@ -369,7 +371,31 @@ if __name__ == '__main__':
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
+                logger.debug("GGUF KV: %s = %d", gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
+                alora_invocation_tokens = lparams.get("alora_invocation_tokens")
+                invocation_string = lparams.get("invocation_string")
+                if invocation_string and not alora_invocation_tokens:
+                    logger.debug("Tokenizing invocation_string -> alora_invocation_tokens")
+                    base_model_path_or_id = hparams.get("_name_or_path")
+                    try:
+                        tokenizer = AutoTokenizer.from_pretrained(base_model_path_or_id)
+                    except ValueError:
+                        logger.error("Unable to load tokenizer from %s", base_model_path_or_id)
+                        raise
+                    # NOTE: There's an off-by-one with the older aLoRAs where
+                    # the invocation string includes the "<|start_of_turn|>"
+                    # token, but the adapters themselves were trained to
+                    # activate _after_ that first token, so we drop it here.
+                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
+                if alora_invocation_tokens:
+                    logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
+                    self.gguf_writer.add_key_value(
+                        gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS,
+                        alora_invocation_tokens,
+                        GGUFValueType.ARRAY,
+                        GGUFValueType.UINT32,
+                    )
 
             def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                 # Never add extra tensors (e.g. rope_freqs) for LoRA adapters

docs/backend/CANN.md

@@ -293,17 +293,14 @@ We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers fr
 
 ## Environment variable setup
 
-### GGML_CANN_ASYNC_MODE
-
-Enables asynchronous operator submission. Disabled by default.
-
 ### GGML_CANN_MEM_POOL
 
-Specifies the memory pool management strategy:
+Specifies the memory pool management strategy, Default is vmm.
 
 - vmm: Utilizes a virtual memory manager pool. If hardware support for VMM is unavailable, falls back to the legacy (leg) memory pool.
 
 - prio: Employs a priority queue-based memory pool management.
+
 - leg: Uses a fixed-size buffer pool.
 
 ### GGML_CANN_DISABLE_BUF_POOL_CLEAN
@@ -312,5 +309,16 @@ Controls automatic cleanup of the memory pool. This option is only effective whe
 
 ### GGML_CANN_WEIGHT_NZ
 
-Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.
+Converting the matmul weight format from ND to NZ to improve performance. Enabled by default.
 
+### GGML_CANN_ACL_GRAPH
+
+Operators are executed using ACL graph execution, rather than in op-by-op (eager) mode. Enabled by default.
+
+### GGML_CANN_GRAPH_CACHE_CAPACITY
+
+Maximum number of compiled CANN graphs kept in the LRU cache, default is 12. When the number of cached graphs exceeds this capacity, the least recently used graph will be evicted.
+
+### GGML_CANN_PREFILL_USE_GRAPH
+
+Enable ACL graph execution during the prefill stage, default is false. This option is only effective when FA is enabled.

docs/build-s390x.md

@@ -42,18 +42,6 @@ cmake --build build --config Release -j $(nproc)
     cmake --build build --config Release -j $(nproc)
     ```
 
--   By default, NNPA is disabled by default. To enable it:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_NNPA=ON
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
 -   For debug builds:
 
     ```bash
@@ -164,15 +152,11 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 Only available in IBM z15/LinuxONE 3 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
 
-### 2. NNPA Vector Intrinsics Acceleration
-
-Only available in IBM z16/LinuxONE 4 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
-
-### 3. zDNN Accelerator (WIP)
+### 2. zDNN Accelerator (WIP)
 
 Only available in IBM z17/LinuxONE 5 or later system with the `-DGGML_ZDNN=ON` compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs will default back to CPU routines.
 
-### 4. Spyre Accelerator
+### 3. Spyre Accelerator
 
 _Only available with IBM z17 / LinuxONE 5 or later system. No support currently available._
 
@@ -230,10 +214,6 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
     CXXFLAGS="-include cstdint" pip3 install -r requirements.txt
     ```
 
-5. `-DGGML_NNPA=ON` generates gibberish output
-
-    Answer: We are aware of this as detailed in [this issue](https://github.com/ggml-org/llama.cpp/issues/14877). Please either try reducing the number of threads, or disable the compile option using `-DGGML_NNPA=OFF`.
-
 ## Getting Help on IBM Z & LinuxONE
 
 1. **Bugs, Feature Requests**
@@ -258,37 +238,38 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 ## Appendix B: SIMD Support Matrix
 
-|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
-| ---------- | ----------- | ---- | ---- | ----- |
-| FP32       | ✅          | ✅   | ✅   | ❓    |
-| FP16       | ✅          | ✅   | ❓   | ❓    |
-| BF16       | 🚫          | 🚫   | ❓   | ❓    |
-| Q4_0       | ✅          | ✅   | ❓   | ❓    |
-| Q4_1       | ✅          | ✅   | ❓   | ❓    |
-| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
-| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
-| Q8_0       | ✅          | ✅   | ❓   | ❓    |
-| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
-| Q3_K       | ✅          | ✅   | ❓   | ❓    |
-| Q4_K       | ✅          | ✅   | ❓   | ❓    |
-| Q5_K       | ✅          | ✅   | ❓   | ❓    |
-| Q6_K       | ✅          | ✅   | ❓   | ❓    |
-| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
-| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
-| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
-| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
-| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
-| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
-| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
-| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+|            | VX/VXE/VXE2 | zDNN | Spyre |
+|------------|-------------|------|-------|
+| FP32       | ✅           | ✅    | ❓     |
+| FP16       | ✅           | ✅    | ❓     |
+| BF16       | 🚫           | ✅    | ❓     |
+| Q4_0       | ✅           | ❓    | ❓     |
+| Q4_1       | ✅           | ❓    | ❓     |
+| MXFP4      | 🚫           | ❓    | ❓     |
+| Q5_0       | ✅           | ❓    | ❓     |
+| Q5_1       | ✅           | ❓    | ❓     |
+| Q8_0       | ✅           | ❓    | ❓     |
+| Q2_K       | 🚫           | ❓    | ❓     |
+| Q3_K       | ✅           | ❓    | ❓     |
+| Q4_K       | ✅           | ❓    | ❓     |
+| Q5_K       | ✅           | ❓    | ❓     |
+| Q6_K       | ✅           | ❓    | ❓     |
+| TQ1_0      | 🚫           | ❓    | ❓     |
+| TQ2_0      | 🚫           | ❓    | ❓     |
+| IQ2_XXS    | 🚫           | ❓    | ❓     |
+| IQ2_XS     | 🚫           | ❓    | ❓     |
+| IQ2_S      | 🚫           | ❓    | ❓     |
+| IQ3_XXS    | 🚫           | ❓    | ❓     |
+| IQ3_S      | 🚫           | ❓    | ❓     |
+| IQ1_S      | 🚫           | ❓    | ❓     |
+| IQ1_M      | 🚫           | ❓    | ❓     |
+| IQ4_NL     | ✅           | ❓    | ❓     |
+| IQ4_XS     | ✅           | ❓    | ❓     |
+| FP32->FP16 | 🚫           | ❓    | ❓     |
+| FP16->FP32 | 🚫           | ❓    | ❓     |
 
 -   ✅ - acceleration available
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 31, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.

docs/build.md

@@ -59,8 +59,6 @@ cmake --build build --config Release
     cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
     cmake --build build-arm64-windows-llvm-release
     ```
-    Building for arm64 can also be done with the MSVC compiler with the build-arm64-windows-MSVC preset, or the standard CMake build instructions. However, note that the MSVC compiler does not support inline ARM assembly code, used e.g. for the accelerated Q4_0_N_M CPU kernels.
-
     For building with ninja generator and clang compiler as default:
       -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
       ```bash
@@ -197,13 +195,12 @@ The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enab
 
 The following compilation options are also available to tweak performance:
 
-| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
-|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
-| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
-| GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
+| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                                                                                                                      |
+|-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
+| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for CDNA and RDNA4) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000).                                            |
+| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                                                                                                                  |
+| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                                                                                                           |
 
 ## MUSA
 

docs/function-calling.md

@@ -21,6 +21,8 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Use `--chat-template-file` to override the template when appropriate (see examples below)
   - Generic support may consume more tokens and be less efficient than a model's native format.
 
+- Multiple/parallel tool calling is supported on some models but disabled by default, enable it by passing `"parallel_tool_calls": true` in the completion endpoint payload.
+
 <details>
 <summary>Show some common templates and which format handler they use</summary>
 

docs/multimodal/MobileVLM.md

@@ -194,7 +194,7 @@ llama_print_timings:       total time =   44411.01 ms /   377 tokens
 ## Orin compile and run
 ### compile
 ```sh
-make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
+make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 -j 32
 ```
 ### run on Orin
 ### case 1

docs/multimodal/minicpmv4.0.md

@@ -6,7 +6,7 @@ Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model
 
 
 ### Build llama.cpp
-Readme modification time: 20250206
+Readme modification time: 20250731
 
 If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
 

docs/multimodal/minicpmv4.5.md

@@ -0,0 +1,47 @@
+## MiniCPM-V 4.5
+
+### Prepare models and code
+
+Download [MiniCPM-V-4_5](https://huggingface.co/openbmb/MiniCPM-V-4_5) PyTorch model from huggingface to "MiniCPM-V-4_5" folder.
+
+
+### Build llama.cpp
+Readme modification time: 20250826
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-V 4
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-4_5-gguf) by us)
+
+```bash
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4_5
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4_5 --minicpmv-projector ../MiniCPM-V-4_5/minicpmv.projector --output-dir ../MiniCPM-V-4_5/ --minicpmv_version 6
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4_5/model
+
+# quantize int4 version
+./build/bin/llama-quantize ../MiniCPM-V-4_5/model/ggml-model-f16.gguf ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4_5/mmproj-model-f16.gguf
+```

docs/ops.md

@@ -18,6 +18,7 @@ Legend:
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -26,6 +27,7 @@ Legend:
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
@@ -49,9 +51,11 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
@@ -61,7 +65,9 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
 |                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -98,6 +104,7 @@ Legend:
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |

examples/CMakeLists.txt

@@ -34,6 +34,7 @@ else()
     add_subdirectory(gen-docs)
     add_subdirectory(training)
     add_subdirectory(diffusion)
+    add_subdirectory(model-conversion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/batched.swift/README.md

@@ -1,4 +1,5 @@
 This is a swift clone of `examples/batched`.
 
-$ `make`
-$ `./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]`
+```bash
+$ ./llama-batched-swift MODEL_PATH [PROMPT] [PARALLEL]
+```

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -333,17 +333,17 @@ static void print_params(struct my_llama_hparams * params) {
 }
 
 static void print_tensor_info(const struct ggml_context * ctx) {
-    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+    for (auto * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
         LOG_INF("%s: Allocating ", __func__);
         int64_t total = 1;
         int i = 0;
         for (; i < ggml_n_dims(t); ++i) {
-            if (i > 0) LOG("x ");
-            LOG("[%" PRId64 "] ", t->ne[i]);
+            if (i > 0) { LOG_INF("x "); }
+            LOG_INF("[%" PRId64 "] ", t->ne[i]);
             total *= t->ne[i];
         }
-        if (i > 1) LOG("= [%" PRId64 "] ", total);
-        LOG("float space for %s\n", ggml_get_name(t));
+        if (i > 1) { LOG_INF("= [%" PRId64 "] ", total); }
+        LOG_INF("float space for %s\n", ggml_get_name(t));
     }
 }
 

examples/diffusion/diffusion-cli.cpp

@@ -510,19 +510,27 @@ static void diffusion_generate(llama_context *          ctx,
     n_generated = params.max_length;
 }
 
-static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
+static std::string format_input_text(const std::string & prompt, const std::string & system_prompt, bool use_chat_template, llama_model * model) {
     if (!use_chat_template) {
         return prompt;
     }
 
     auto chat_templates = common_chat_templates_init(model, "");
-
     common_chat_templates_inputs inputs;
-    common_chat_msg              user_msg;
-    user_msg.role                = "user";
-    user_msg.content             = prompt;
-    inputs.add_generation_prompt = true;
+    common_chat_msg system_msg;
+
+    if (!system_prompt.empty()) {
+        system_msg.role = "system";
+        system_msg.content = system_prompt;
+        inputs.messages.push_back(system_msg);
+    }
+
+    common_chat_msg user_msg;
+    user_msg.role = "user";
+    user_msg.content = prompt;
+
     inputs.messages.push_back(user_msg);
+    inputs.add_generation_prompt = true;
 
     auto result = common_chat_templates_apply(chat_templates.get(), inputs);
 
@@ -564,7 +572,7 @@ int main(int argc, char ** argv) {
     ctx_params.n_ctx                = params.n_ctx;
     ctx_params.n_batch              = params.n_batch;
     ctx_params.n_ubatch             = params.n_ubatch;
-    ctx_params.flash_attn           = params.flash_attn;
+    ctx_params.flash_attn_type      = params.flash_attn_type;
     ctx_params.no_perf              = params.no_perf;
     ctx_params.type_k               = params.cache_type_k;
     ctx_params.type_v               = params.cache_type_v;
@@ -579,7 +587,8 @@ int main(int argc, char ** argv) {
     llama_set_n_threads(ctx, params.cpuparams.n_threads, params.cpuparams_batch.n_threads);
 
     const llama_vocab * vocab            = llama_model_get_vocab(model);
-    std::string         formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
+
+    std::string         formatted_prompt = format_input_text(params.prompt, params.system_prompt, params.enable_chat_template, model);
 
     std::vector<llama_token> input_tokens = common_tokenize(vocab,
                                                             formatted_prompt,
@@ -596,6 +605,7 @@ int main(int argc, char ** argv) {
     }
 
     llama_token mask_token_id = llama_vocab_mask(vocab);
+
     GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
 
     bool visual_mode = params.diffusion.visual_mode;

examples/eval-callback/eval-callback.cpp

@@ -28,9 +28,51 @@ static std::string ggml_ne_string(const ggml_tensor * t) {
     return str;
 }
 
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+static float ggml_get_float_value(uint8_t * data, ggml_type type, const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
+    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+    float v;
+    if (type == GGML_TYPE_F16) {
+        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
+    } else if (type == GGML_TYPE_F32) {
+        v = *(float *) &data[i];
+    } else if (type == GGML_TYPE_I64) {
+        v = (float) *(int64_t *) &data[i];
+    } else if (type == GGML_TYPE_I32) {
+        v = (float) *(int32_t *) &data[i];
+    } else if (type == GGML_TYPE_I16) {
+        v = (float) *(int16_t *) &data[i];
+    } else if (type == GGML_TYPE_I8) {
+        v = (float) *(int8_t *) &data[i];
+    } else if (type == GGML_TYPE_BF16) {
+        v = ggml_compute_bf16_to_fp32(*(ggml_bf16_t *) &data[i]);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+    return v;
+}
+
 static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
     GGML_ASSERT(n > 0);
     float sum = 0;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
+                    sum += v;
+                }
+            }
+        }
+    }
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
         LOG("                                     [\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
@@ -50,25 +92,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         LOG("..., ");
                         i0 = ne[0] - n;
                     }
-                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
-                    float v;
-                    if (type == GGML_TYPE_F16) {
-                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
-                    } else if (type == GGML_TYPE_F32) {
-                        v = *(float *) &data[i];
-                    } else if (type == GGML_TYPE_I64) {
-                        v = (float) *(int64_t *) &data[i];
-                    } else if (type == GGML_TYPE_I32) {
-                        v = (float) *(int32_t *) &data[i];
-                    } else if (type == GGML_TYPE_I16) {
-                        v = (float) *(int16_t *) &data[i];
-                    } else if (type == GGML_TYPE_I8) {
-                        v = (float) *(int8_t *) &data[i];
-                    } else {
-                        GGML_ABORT("fatal error");
-                    }
+                    const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
                     LOG("%12.4f", v);
-                    sum += v;
                     if (i0 < ne[0] - 1) LOG(", ");
                 }
                 LOG("],\n");

examples/json_schema_to_grammar.py

@@ -586,9 +586,10 @@ class SchemaConverter:
             properties = list(schema.get('properties', {}).items())
             return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
 
-        elif schema_type in (None, 'object') and 'allOf' in schema:
+        elif schema_type in (None, 'object', 'string') and 'allOf' in schema:
             required = set()
             properties = []
+            enum_sets = []
             hybrid_name = name
             def add_component(comp_schema, is_required):
                 if (ref := comp_schema.get('$ref')) is not None:
@@ -600,6 +601,9 @@ class SchemaConverter:
                         if is_required:
                             required.add(prop_name)
 
+                if 'enum' in comp_schema:
+                    enum_sets.append(set(comp_schema['enum']))
+
             for t in schema['allOf']:
                 if 'anyOf' in t:
                     for tt in t['anyOf']:
@@ -607,6 +611,15 @@ class SchemaConverter:
                 else:
                     add_component(t, is_required=True)
 
+            if enum_sets:
+                enum_intersection = enum_sets[0]
+                for s in enum_sets[1:]:
+                    enum_intersection &= s
+
+                if enum_intersection:
+                    rule = '(' + ' | '.join((self._generate_constant_rule(v) for v in sorted(enum_intersection))) + ') space'
+                    return self._add_rule(rule_name, rule)
+
             return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=None))
 
         elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):

examples/llama.vim

@@ -17,7 +17,7 @@
 "
 " start the llama.cpp server with a FIM-compatible model. for example:
 "
-"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa --ubatch-size 512 --batch-size 1024 --cache-reuse 256
 "
 "   --batch-size [512, model max context]
 "

examples/model-conversion/.gitignore

@@ -0,0 +1,3 @@
+.model_name
+data
+ppl

examples/model-conversion/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-logits)
+add_executable(${TARGET} logits.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/model-conversion/Makefile

@@ -0,0 +1,206 @@
+MAKEFLAGS += --no-print-directory
+
+define validate_model_path
+	@if [ -z "$(MODEL_PATH)" ]; then \
+		echo "Error: MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define validate_embedding_model_path
+	@if [ -z "$(EMBEDDING_MODEL_PATH)" ]; then \
+		echo "Error: EMBEDDING_MODEL_PATH must be provided either as:"; \
+		echo "  1. Environment variable: export EMBEDDING_MODEL_PATH=/path/to/model"; \
+		echo "  2. Command line argument: make $(1) EMBEDDING_MODEL_PATH=/path/to/model"; \
+		exit 1; \
+	fi
+endef
+
+define quantize_model
+	@CONVERTED_MODEL="$(1)" QUANTIZED_TYPE="$(QUANTIZED_TYPE)" \
+	TOKEN_EMBD_TYPE="$(TOKEN_EMBD_TYPE)" OUTPUT_TYPE="$(OUTPUT_TYPE)" \
+	./scripts/utils/quantize.sh "$(1)" "$(QUANTIZED_TYPE)" "$(TOKEN_EMBD_TYPE)" "$(OUTPUT_TYPE)"
+	@echo "Export the quantized model path to $(2) variable in your environment"
+endef
+
+###
+### Casual Model targets/recipes
+###
+causal-convert-model-bf16: OUTTYPE=bf16
+causal-convert-model-bf16: causal-convert-model
+
+causal-convert-model:
+	$(call validate_model_path,causal-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+causal-convert-mm-model-bf16: OUTTYPE=bf16
+causal-convert-mm-model-bf16: MM_OUTTYPE=f16
+causal-convert-mm-model-bf16: causal-convert-mm-model
+
+causal-convert-mm-model:
+	$(call validate_model_path,causal-convert-mm-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh
+
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(MM_OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/causal/convert-model.sh --mmproj
+
+causal-run-original-model:
+	$(call validate_model_path,causal-run-original-model)
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/run-org-model.py
+
+causal-run-converted-model:
+	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh
+
+causal-verify-logits: causal-run-original-model causal-run-converted-model
+	@./scripts/causal/compare-logits.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/utils/check-nmse.py -m ${MODEL_PATH}
+
+causal-run-original-embeddings:
+	@./scripts/causal/run-casual-gen-embeddings-org.py
+
+causal-run-converted-embeddings:
+	@./scripts/causal/run-converted-model-embeddings-logits.sh
+
+causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-embeddings
+	@./scripts/causal/compare-embeddings-logits.sh
+
+causal-inspect-original-model:
+	@./scripts/utils/inspect-org-model.py
+
+causal-inspect-converted-model:
+	@./scripts/utils/inspect-converted-model.sh
+
+causal-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_MODEL}
+
+causal-curl-embedding-endpoint: causal-run-original-embeddings
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/causal/compare-embeddings-logits.sh
+
+causal-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+causal-quantize-Q8_0: causal-quantize-model
+
+causal-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-Q4_0: causal-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+causal-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+causal-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+causal-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+causal-quantize-qat-Q4_0: causal-quantize-model
+
+causal-quantize-model:
+	$(call quantize_model,$(CONVERTED_MODEL),QUANTIZED_MODEL)
+
+causal-run-quantized-model:
+	@QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/causal/run-converted-model.sh ${QUANTIZED_MODEL}
+
+
+###
+### Embedding Model targets/recipes
+###
+
+embedding-convert-model-bf16: OUTTYPE=bf16
+embedding-convert-model-bf16: embedding-convert-model
+
+embedding-convert-model:
+	$(call validate_embedding_model_path,embedding-convert-model)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/embedding/convert-model.sh
+
+embedding-run-original-model:
+	$(call validate_embedding_model_path,embedding-run-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
+
+embedding-run-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
+	@./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-inspect-original-model:
+	$(call validate_embedding_model_path,embedding-inspect-original-model)
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+
+embedding-inspect-converted-model:
+	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-start-embedding-server:
+	@./scripts/utils/run-embedding-server.sh ${CONVERTED_EMBEDDING_MODEL}
+
+embedding-curl-embedding-endpoint:
+	@./scripts/utils/curl-embedding-server.sh | ./scripts/embedding/compare-embeddings-logits.sh
+
+embedding-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
+embedding-quantize-Q8_0: embedding-quantize-model
+
+embedding-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-Q4_0: embedding-quantize-model
+
+# For Quantization Aware Trained (QAT) models in Q4_0 we explicitly set the
+# token embedding and output types to Q8_0 instead of the default Q6_K.
+embedding-quantize-qat-Q4_0: QUANTIZED_TYPE = Q4_0
+embedding-quantize-qat-Q4_0: TOKEN_EMBD_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: OUTPUT_TYPE = Q8_0
+embedding-quantize-qat-Q4_0: embedding-quantize-model
+
+embedding-quantize-model:
+	$(call quantize_model,$(CONVERTED_EMBEDDING_MODEL),QUANTIZED_EMBEDDING_MODEL)
+
+embedding-run-quantized-model:
+	@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
+
+###
+### Perplexity targets/recipes
+###
+perplexity-data-gen:
+	CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/utils/perplexity-gen.sh
+
+perplexity-run-full:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" LOOGITS_FILE="$(LOGITS_FILE)" \
+	./scripts/utils/perplexity-run.sh
+
+perplexity-run:
+	QUANTIZED_MODEL="$(QUANTIZED_MODEL)" ./scripts/utils/perplexity-run-simple.sh
+
+###
+### HuggingFace targets/recipes
+###
+
+hf-create-model:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}"
+
+hf-create-model-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -d
+
+hf-create-model-embedding:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e
+
+hf-create-model-embedding-dry-run:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -e -d
+
+hf-create-model-private:
+	@./scripts/utils/hf-create-model.py -m "${MODEL_NAME}" -ns "${NAMESPACE}" -b "${ORIGINAL_BASE_MODEL}" -p
+
+hf-upload-gguf-to-model:
+	@./scripts/utils/hf-upload-gguf-model.py -m "${MODEL_PATH}" -r "${REPO_ID}" -o "${NAME_IN_REPO}"
+
+hf-create-collection:
+	@./scripts/utils/hf-create-collection.py -n "${NAME}" -d "${DESCRIPTION}" -ns "${NAMESPACE}"
+
+hf-add-model-to-collection:
+	@./scripts/utils/hf-add-model-to-collection.py -c "${COLLECTION}" -m "${MODEL}"
+
+
+.PHONY: clean
+clean:
+	@${RM} -rf data .converted_embedding_model.txt .converted_model.txt .embedding_model_name.txt .model_name.txt
+

examples/model-conversion/README.md

@@ -0,0 +1,367 @@
+# Model Conversion Example
+This directory contains scripts and code to help in the process of converting
+HuggingFace PyTorch models to GGUF format.
+
+The motivation for having this is that the conversion process can often be an
+iterative process, where the original model is inspected, converted, updates
+made to llama.cpp, converted again, etc. Once the model has been converted it
+needs to be verified against the original model, and then optionally quantified,
+and in some cases perplexity checked of the quantized model. And finally the
+model/models need to the ggml-org on Hugging Face. This tool/example tries to
+help with this process.
+
+### Overview
+The idea is that the makefile targets and scripts here can be used in the
+development/conversion process assisting with things like:
+
+* inspect/run the original model to figure out how it works
+* convert the original model to GGUF format
+* inspect/run the converted model
+* verify the logits produced by the original model and the converted model
+* quantize the model to GGUF format
+* run perplexity evaluation to verify that the quantized model is performing
+  as expected
+* upload the model to HuggingFace to make it available for others
+
+## Setup
+Create virtual python environment
+```console
+$ python3.11 -m venv venv
+$ source venv/bin/activate
+(venv) $ pip install -r requirements.txt
+```
+
+## Causal Language Model Conversion
+This section describes the steps to convert a causal language model to GGUF and
+to verify that the conversion was successful.
+
+### Download the original model
+First, clone the original model to some local directory:
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+### Set the MODEL_PATH
+The path to the downloaded model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export MODEL_PATH=~/work/ai/models/some_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+In cases where the transformer implementation for the model has not been released
+yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
+will then cause the transformer implementation to be loaded explicitely and not
+use AutoModelForCausalLM:
+```
+export UNRELEASED_MODEL_NAME=SomeNewModel
+```
+
+### Inspecting the original tensors
+```console
+# Using environment variable
+(venv) $ make causal-inspect-original-model
+
+# Or using command line argument
+(venv) $ make causal-inspect-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Running the original model
+This is mainly to verify that the original model works, and to compare the output
+from the converted model.
+```console
+# Using environment variable
+(venv) $ make causal-run-original-model
+
+# Or using command line argument
+(venv) $ make causal-run-original-model MODEL_PATH=~/work/ai/models/some_model
+```
+This command will save two files to the `data` directory, one is a binary file
+containing logits which will be used for comparison with the converted model
+later, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model, the model can be converted to GGUF format using the following command:
+```console
+# Using environment variable
+(venv) $ make causal-convert-model
+
+# Or using command line argument
+(venv) $ make causal-convert-model MODEL_PATH=~/work/ai/models/some_model
+```
+
+### Inspecting the converted model
+The converted model can be inspected using the following command:
+```console
+(venv) $ make inspect-converted-model
+```
+
+### Running the converted model
+```console
+(venv) $ make run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model and
+compare the logits:
+```console
+(venv) $ make causal-verify-logits
+```
+
+### Quantizing the model
+The causal model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make causal-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_MODEL` environment variable:
+```console
+export QUANTIZED_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make causal-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make causal-quantize-qat-Q4_0
+```
+
+
+## Embedding Language Model Conversion
+
+### Download the original model
+```console
+$ mkdir models && cd models
+$ git clone https://huggingface.co/user/model_name
+$ cd model_name
+$ git lfs install
+$ git lfs pull
+```
+
+The path to the embedding model can be provided in two ways:
+
+**Option 1: Environment variable (recommended for iterative development)**
+```console
+export EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+**Option 2: Command line argument (for one-off tasks)**
+```console
+make embedding-convert-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+
+Command line arguments take precedence over environment variables when both are provided.
+
+### Running the original model
+This is mainly to verify that the original model works and to compare the output
+with the output from the converted model.
+```console
+# Using environment variable
+(venv) $ make embedding-run-original-model
+
+# Or using command line argument
+(venv) $ make embedding-run-original-model EMBEDDING_MODEL_PATH=~/path/to/embedding_model
+```
+This command will save two files to the `data` directory, one is a binary
+file containing logits which will be used for comparison with the converted
+model, and the other is a text file which allows for manual visual inspection.
+
+### Model conversion
+After updates have been made to [gguf-py](../../gguf-py) to add support for the
+new model the model can be converted to GGUF format using the following command:
+```console
+(venv) $ make embedding-convert-model
+```
+
+### Run the converted model
+```console
+(venv) $ make embedding-run-converted-model
+```
+
+### Model logits verfication
+The following target will run the original model and the converted model (which
+was done manually in the previous steps) and compare the logits:
+```console
+(venv) $ make embedding-verify-logits
+```
+
+### llama-server verification
+To verify that the converted model works with llama-server, the following
+command can be used:
+```console
+(venv) $ make embedding-start-embedding-server
+```
+Then open another terminal and set the `EMBEDDINGS_MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make embedding-curl-embedding-endpoint
+```
+This will call the `embedding` endpoing and the output will be piped into
+the same verification script as used by the target `embedding-verify-logits`.
+
+The causal model can also be used to produce embeddings and this can be verified
+using the following commands:
+```console
+(venv) $ make causal-start-embedding-server
+```
+Then open another terminal and set the `MODEL_PATH` environment
+variable as this will not be inherited by the new terminal:
+```console
+(venv) $ make casual-curl-embedding-endpoint
+```
+
+### Quantizing the model
+The embedding model can be quantized to GGUF format using the following command:
+```console
+(venv) $ make embedding-quantize-Q8_0
+Quantized model saved to: /path/to/quantized/model-Q8_0.gguf
+Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment
+```
+This will show the path to the quantized model in the terminal, which can then
+be used to set the `QUANTIZED_EMBEDDING_MODEL` environment variable:
+```console
+export QUANTIZED_EMBEDDING_MODEL=/path/to/quantized/model-Q8_0.gguf
+```
+Then the quantized model can be run using the following command:
+```console
+(venv) $ make embedding-run-quantized-model
+```
+
+### Quantizing QAT (Quantization Aware Training) models
+When quantizing to `Q4_0`, the default data type for the token embedding weights
+will be `Q6_K`. For models that are going to be uploaded to ggml-org it is
+recommended to use `Q8_0` instead for the embeddings and output tensors.
+The reason is that although `Q6_K` is smaller in size, it requires more compute
+to unpack, which can hurt performance during output generation when the entire
+embedding matrix must be dequantized to compute vocabulary logits. `Q8_0`
+provides practically full quality with better computational efficiency.
+```console
+(venv) $ make embedding-quantize-qat-Q4_0
+```
+
+## Perplexity Evaluation
+
+### Simple perplexity evaluation
+This allows to run the perplexity evaluation without having to generate a
+token/logits file:
+```console
+(venv) $ make perplexity-run QUANTIZED_MODEL=~/path/to/quantized/model.gguf
+```
+This will use the wikitext dataset to run the perplexity evaluation and
+output the perplexity score to the terminal. This value can then be compared
+with the perplexity score of the unquantized model.
+
+### Full perplexity evaluation
+First use the converted, non-quantized, model to generate the perplexity evaluation
+dataset using the following command:
+```console
+$ make perplexity-data-gen CONVERTED_MODEL=~/path/to/converted/model.gguf
+```
+This will generate a file in the `data` directory named after the model and with
+a `.kld` suffix which contains the tokens and the logits for the wikitext dataset.
+
+After the dataset has been generated, the perplexity evaluation can be run using
+the quantized model:
+```console
+$ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LOGITS_FILE=data/model.gguf.ppl
+```
+
+> 📝 **Note:** The `LOGITS_FILE` is the file generated by the previous command
+> can be very large, so make sure you have enough disk space available.
+
+## HuggingFace utilities
+The following targets are useful for creating collections and model repositories
+on Hugging Face in the the ggml-org. These can be used when preparing a relase
+to script the process for new model releases.
+
+For the following targets a `HF_TOKEN` environment variable is required.
+
+> 📝 **Note:** Don't forget to logout from Hugging Face after running these
+> commands, otherwise you might have issues pulling/cloning repositories as
+> the token will still be in use:
+> $ huggingface-cli logout
+> $ unset HF_TOKEN
+
+### Create a new Hugging Face Model (model repository)
+This will create a new model repsository on Hugging Face with the specified
+model name.
+```console
+(venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+Repository ID:  danbev/TestModel-GGUF
+Repository created: https://huggingface.co/danbev/TestModel-GGUF
+```
+Note that we append a `-GGUF` suffix to the model name to ensure a consistent
+naming convention for GGUF models.
+
+An embedding model can be created using the following command:
+```console
+(venv) $ make hf-create-model-embedding MODEL_NAME='TestEmbeddingModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"
+```
+The only difference is that the model card for an embedding model will be different
+with regards to the llama-server command and also how to access/call the embedding
+endpoint.
+
+### Upload a GGUF model to model repository
+The following target uploads a model to an existing Hugging Face model repository.
+```console
+(venv) $ make hf-upload-gguf-to-model MODEL_PATH=dummy-model1.gguf REPO_ID=danbev/TestModel-GGUF
+📤 Uploading dummy-model1.gguf to danbev/TestModel-GGUF/dummy-model1.gguf
+✅ Upload successful!
+🔗 File available at: https://huggingface.co/danbev/TestModel-GGUF/blob/main/dummy-model1.gguf
+```
+This command can also be used to update an existing model file in a repository.
+
+### Create a new Collection
+```console
+(venv) $ make hf-new-collection NAME=TestCollection DESCRIPTION="Collection for testing scripts" NAMESPACE=danbev
+🚀 Creating Hugging Face Collection
+Title: TestCollection
+Description: Collection for testing scripts
+Namespace: danbev
+Private: False
+✅ Authenticated as: danbev
+📚 Creating collection: 'TestCollection'...
+✅ Collection created successfully!
+📋 Collection slug: danbev/testcollection-68930fcf73eb3fc200b9956d
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Collection created successfully!
+Use this slug to add models: danbev/testcollection-68930fcf73eb3fc200b9956d
+```
+
+### Add model to a Collection
+```console
+(venv) $ make hf-add-model-to-collection COLLECTION=danbev/testcollection-68930fcf73eb3fc200b9956d MODEL=danbev/TestModel-GGUF
+✅ Authenticated as: danbev
+🔍 Checking if model exists: danbev/TestModel-GGUF
+✅ Model found: danbev/TestModel-GGUF
+📚 Adding model to collection...
+✅ Model added to collection successfully!
+🔗 Collection URL: https://huggingface.co/collections/danbev/testcollection-68930fcf73eb3fc200b9956d
+
+🎉 Model added successfully!
+
+```

examples/model-conversion/logits.cpp

@@ -0,0 +1,210 @@
+#include "llama.h"
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <vector>
+#include <ctype.h>
+#include <filesystem>
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    std::string model_path;
+    std::string prompt = "Hello, my name is";
+    int ngl = 0;
+    bool embedding_mode = false;
+
+    {
+        int i = 1;
+        for (; i < argc; i++) {
+            if (strcmp(argv[i], "-m") == 0) {
+                if (i + 1 < argc) {
+                    model_path = argv[++i];
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-ngl") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        ngl = std::stoi(argv[++i]);
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-embd-mode") == 0) {
+                if (i + 1 < argc) {
+                    try {
+                        embedding_mode = true;
+                    } catch (...) {
+                        print_usage(argc, argv);
+                        return 1;
+                    }
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else {
+                // prompt starts here
+                break;
+            }
+        }
+
+        if (model_path.empty()) {
+            print_usage(argc, argv);
+            return 1;
+        }
+
+        if (i < argc) {
+            prompt = argv[i++];
+            for (; i < argc; i++) {
+                prompt += " ";
+                prompt += argv[i];
+            }
+        }
+    }
+
+    ggml_backend_load_all();
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = ngl;
+
+    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // Extract basename from model_path
+    const char * basename = strrchr(model_path.c_str(), '/');
+    basename = (basename == NULL) ? model_path.c_str() : basename + 1;
+
+    char model_name[256];
+    strncpy(model_name, basename, 255);
+    model_name[255] = '\0';
+
+    char * dot = strrchr(model_name, '.');
+    if (dot != NULL && strcmp(dot, ".gguf") == 0) {
+        *dot = '\0';
+    }
+    printf("Model name: %s\n", model_name);
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+
+    std::vector<llama_token> prompt_tokens(n_prompt);
+    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
+        fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
+        return 1;
+    }
+
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = n_prompt;
+    ctx_params.n_batch = n_prompt;
+    ctx_params.no_perf = false;
+    if (embedding_mode) {
+        ctx_params.embeddings = true;
+        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        ctx_params.n_ubatch = ctx_params.n_batch;
+    }
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    printf("Input prompt: \"%s\"\n", prompt.c_str());
+    printf("Tokenized prompt (%d tokens): ", n_prompt);
+    for (auto id : prompt_tokens) {
+        char buf[128];
+        int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
+            return 1;
+        }
+        std::string s(buf, n);
+        printf("%s", s.c_str());
+    }
+    printf("\n");
+
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+
+    if (llama_decode(ctx, batch)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    float * logits;
+    int n_logits;
+    const char * type;
+
+    if (embedding_mode) {
+        logits = llama_get_embeddings(ctx);
+        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        type = "-embeddings";
+        printf("Embeddings size: %d\n", n_logits);
+    } else {
+        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        n_logits = llama_vocab_n_tokens(vocab);
+        type = "";
+        printf("Vocab size: %d\n", n_logits);
+    }
+
+    std::filesystem::create_directory("data");
+
+    // Save logits to binary file
+    char bin_filename[512];
+    snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
+    printf("Saving logits to %s\n", bin_filename);
+
+    FILE * f = fopen(bin_filename, "wb");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
+        return 1;
+    }
+    fwrite(logits, sizeof(float), n_logits, f);
+    fclose(f);
+
+    // Also save as text for debugging
+    char txt_filename[512];
+    snprintf(txt_filename, sizeof(txt_filename), "data/llamacpp-%s%s.txt", model_name, type);
+    f = fopen(txt_filename, "w");
+    if (f == NULL) {
+        fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
+        return 1;
+    }
+    for (int i = 0; i < n_logits; i++) {
+        fprintf(f, "%d: %.6f\n", i, logits[i]);  // Added index and changed format
+    }
+    fclose(f);
+
+    // Print first and last 10 logits for quick verification
+    printf("First 10 logits: ");
+    for (int i = 0; i < 10 && i < n_logits; i++) {
+        printf("%.6f ", logits[i]);
+    }
+    printf("\n");
+
+    printf("Last 10 logits: ");
+    for (int i = n_logits - 10; i < n_logits; i++) {
+        if (i >= 0) printf("%.6f ", logits[i]);
+    }
+    printf("\n\n");
+
+    printf("Logits saved to %s\n", bin_filename);
+    printf("Logits saved to %s\n", txt_filename);
+
+    llama_free(ctx);
+    llama_model_free(model);
+
+    return 0;
+}

examples/model-conversion/requirements.txt

@@ -0,0 +1,6 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch
+torchvision
+transformers
+huggingface-hub
+accelerate

examples/model-conversion/scripts/causal/compare-embeddings-logits.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+
+set -e
+
+MODEL_PATH="${1:-"$MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today" \
+    --causal
+

examples/model-conversion/scripts/causal/compare-logits.py

@@ -0,0 +1,88 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+from pathlib import Path
+
+def quick_logits_check(pytorch_file, llamacpp_file):
+    """Lightweight sanity check before NMSE"""
+
+    try:
+        pytorch_logits = np.fromfile(pytorch_file, dtype=np.float32)
+        llamacpp_logits = np.fromfile(llamacpp_file, dtype=np.float32)
+    except Exception as e:
+        print(f"❌ NOK: Failed to load files - {e}")
+        return False
+
+    # Check shapes match
+    if pytorch_logits.shape != llamacpp_logits.shape:
+        print(f"❌ NOK: Shape mismatch - PyTorch: {pytorch_logits.shape}, llama.cpp: {llamacpp_logits.shape}")
+        return False
+
+    # Calculate key metrics
+    diff = pytorch_logits - llamacpp_logits
+    abs_diff = np.abs(diff)
+    max_diff = np.max(abs_diff)
+
+    # Get top 10 predictions from both models
+    pytorch_top10 = np.argsort(pytorch_logits)[-10:][::-1]
+    llamacpp_top10 = np.argsort(llamacpp_logits)[-10:][::-1]
+    print(f"Top 10 PyTorch logits: {pytorch_logits[pytorch_top10]}")
+    print(f"Top 10 llama.cpp logits: {llamacpp_logits[llamacpp_top10]}")
+    print(f"Max absolute difference: {max_diff:.4f}")
+
+    if max_diff > 1.0:
+        print(f"❌ NOK: Large differences detected - max diff: {max_diff:.4f}")
+        return False
+
+    return True
+
+def main():
+    model_path = os.getenv('MODEL_PATH')
+    if not model_path:
+        print("Error: MODEL_PATH environment variable not set")
+        sys.exit(1)
+
+    if not os.path.exists(model_path):
+        print(f"Error: Model file not found: {model_path}")
+        sys.exit(1)
+
+    model_name = os.path.splitext(os.path.basename(model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    if not pytorch_file.exists():
+        print(f"Error: PyTorch logits file not found: {pytorch_file}")
+        print("Please run scripts/run-org-model.sh first to generate this file.")
+        sys.exit(1)
+
+    if not llamacpp_file.exists():
+        print(f"Error: llama.cpp logits file not found: {llamacpp_file}")
+        print("Please run scripts/run-converted-model.sh first to generate this file.")
+        sys.exit(1)
+
+    print("Checked all required files were found. Proceeding...\n")
+
+
+    print("🔍 GGML Model Validation for model ", model_name)
+    print("=" * 40)
+    print(f"PyTorch logits  : {pytorch_file}")
+    print(f"llama.cpp logits: {llamacpp_file}")
+    print()
+
+    success = quick_logits_check(pytorch_file, llamacpp_file)
+
+    # Exit with appropriate code
+    if success:
+        print("✅ OK: Lightweight model check successful!")
+        print("       Ok to proceed with NMSE check...")
+        sys.exit(0)
+    else:
+        print(f"❌ NOK: Top 10 predictions don't match - generation will differ")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/causal/convert-model.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env bash
+
+set -e
+
+# Parse command line arguments
+MMPROJ=""
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --mmproj)
+            MMPROJ="--mmproj"
+            shift
+            ;;
+        *)
+            shift
+            ;;
+    esac
+done
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+echo "Metadata override: ${METADATA_OVERRIDE}"
+
+CMD_ARGS=("python" "../../convert_hf_to_gguf.py" "--verbose")
+CMD_ARGS+=("${MODEL_PATH}")
+CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
+CMD_ARGS+=("--outtype" "${TYPE}")
+[[ -n "$METADATA_OVERRIDE" ]] && CMD_ARGS+=("--metadata" "${METADATA_OVERRIDE}")
+[[ -n "$MMPROJ" ]] && CMD_ARGS+=("${MMPROJ}")
+
+"${CMD_ARGS[@]}"
+
+echo ""
+echo "The environment variable CONVERTED_MODEL can be set to this path using:"
+echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"
+if [[ -n "$MMPROJ" ]]; then
+    mmproj_file="${OUTPUT_DIR}/mmproj-$(basename "${CONVERTED_MODEL}")"
+    echo "The mmproj model was created in $(realpath "$mmproj_file")"
+fi

examples/model-conversion/scripts/causal/modelcard.template

@@ -0,0 +1,13 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF -c 0 -fa
+```
+
+Then, access http://localhost:8080

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.py

@@ -0,0 +1,114 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+import torch
+import numpy as np
+
+from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM
+from pathlib import Path
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        print("Falling back to AutoModelForCausalLM")
+        model = AutoModelForCausalLM.from_pretrained(model_path)
+else:
+    model = AutoModelForCausalLM.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+
+model_name = os.path.basename(model_path)
+print(f"Model name: {model_name}")
+
+prompt = "Hello world today"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids, output_hidden_states=True)
+
+    # Extract hidden states from the last layer
+    # outputs.hidden_states is a tuple of (num_layers + 1) tensors
+    # Index -1 gets the last layer, shape: [batch_size, seq_len, hidden_size]
+    last_hidden_states = outputs.hidden_states[-1]
+
+    # Get embeddings for all tokens
+    token_embeddings = last_hidden_states[0].cpu().numpy()  # Remove batch dimension
+
+    print(f"Hidden states shape: {last_hidden_states.shape}")
+    print(f"Token embeddings shape: {token_embeddings.shape}")
+    print(f"Hidden dimension: {token_embeddings.shape[-1]}")
+    print(f"Number of tokens: {token_embeddings.shape[0]}")
+
+    # Save raw token embeddings
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all token embeddings as binary
+    print(token_embeddings)
+    token_embeddings.astype(np.float32).tofile(bin_filename)
+
+    # Save as text for inspection
+    with open(txt_filename, "w") as f:
+        for i, embedding in enumerate(token_embeddings):
+            for j, val in enumerate(embedding):
+                f.write(f"{i} {j} {val:.6f}\n")
+
+    # Print embeddings per token in the requested format
+    print("\nToken embeddings:")
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
+    for i, embedding in enumerate(token_embeddings):
+        # Format: show first few values, ..., then last few values
+        if len(embedding) > 10:
+            # Show first 3 and last 3 values with ... in between
+            first_vals = " ".join(f"{val:8.6f}" for val in embedding[:3])
+            last_vals = " ".join(f"{val:8.6f}" for val in embedding[-3:])
+            print(f"embedding {i}: {first_vals}  ... {last_vals}")
+        else:
+            # If embedding is short, show all values
+            vals = " ".join(f"{val:8.6f}" for val in embedding)
+            print(f"embedding {i}: {vals}")
+
+    # Also show token info for reference
+    print(f"\nToken reference:")
+    for i, token in enumerate(tokens):
+        print(f"  Token {i}: {repr(token)}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m $CONVERTED_MODEL -embd-mode "Hello world today"

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" "Hello, my name is"

examples/model-conversion/scripts/causal/run-org-model.py

@@ -0,0 +1,231 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+import torch
+import numpy as np
+
+### If you want to dump RoPE activations, apply this monkey patch to the model
+### class from Transformers that you are running (replace apertus.modeling_apertus
+### with the proper package and class for your model
+### === START ROPE DEBUG ===
+# from transformers.models.apertus.modeling_apertus import apply_rotary_pos_emb
+
+# orig_rope = apply_rotary_pos_emb
+# torch.set_printoptions(threshold=float('inf'))
+# torch.set_printoptions(precision=6, sci_mode=False)
+
+# def debug_rope(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+#     # log inputs
+#     summarize(q, "RoPE.q_in")
+#     summarize(k, "RoPE.k_in")
+
+#     # call original
+#     q_out, k_out = orig_rope(q, k, cos, sin, position_ids, unsqueeze_dim)
+
+#     # log outputs
+#     summarize(q_out, "RoPE.q_out")
+#     summarize(k_out, "RoPE.k_out")
+
+#     return q_out, k_out
+
+# # Patch it
+# import transformers.models.apertus.modeling_apertus as apertus_mod  # noqa: E402
+# apertus_mod.apply_rotary_pos_emb = debug_rope
+### == END ROPE DEBUG ===
+
+
+def summarize(tensor: torch.Tensor, name: str, max_seq: int = 3, max_vals: int = 3):
+    """
+    Print a tensor in llama.cpp debug style.
+
+    Supports:
+    - 2D tensors (seq, hidden)
+    - 3D tensors (batch, seq, hidden)
+    - 4D tensors (batch, seq, heads, dim_per_head) via flattening heads × dim_per_head
+
+    Shows first and last max_vals of each vector per sequence position.
+    """
+    t = tensor.detach().to(torch.float32).cpu()
+
+    # Determine dimensions
+    if t.ndim == 3:
+        _, s, _ = t.shape
+    elif t.ndim == 2:
+        _, s = 1, t.shape[0]
+        t = t.unsqueeze(0)
+    elif t.ndim == 4:
+        _, s, _, _ = t.shape
+    else:
+        print(f"Skipping tensor due to unsupported dimensions: {t.ndim}")
+        return
+
+    ten_shape = t.shape
+
+    print(f"ggml_debug: {name} = (f32)  ... = {{{ten_shape}}}")
+    print("                                     [")
+    print("                                      [")
+
+    # Determine indices for first and last sequences
+    first_indices = list(range(min(s, max_seq)))
+    last_indices = list(range(max(0, s - max_seq), s))
+
+    # Check if there's an overlap between first and last indices or if we're at the edge case of s = 2 * max_seq
+    has_overlap = bool(set(first_indices) & set(last_indices)) or (max_seq * 2 == s)
+
+    # Combine indices
+    if has_overlap:
+        # If there's overlap, just use the combined unique indices
+        indices = sorted(list(set(first_indices + last_indices)))
+        separator_index = None
+    else:
+        # If no overlap, we'll add a separator between first and last sequences
+        indices = first_indices + last_indices
+        separator_index = len(first_indices)
+
+    for i, si in enumerate(indices):
+        # Add separator if needed
+        if separator_index is not None and i == separator_index:
+            print("                                       ...")
+
+        # Extract appropriate slice
+        vec = t[0, si]
+        if vec.ndim == 2:  # 4D case: flatten heads × dim_per_head
+            flat = vec.flatten().tolist()
+        else:  # 2D or 3D case
+            flat = vec.tolist()
+
+        # First and last slices
+        first = flat[:max_vals]
+        last = flat[-max_vals:] if len(flat) >= max_vals else flat
+        first_str = ", ".join(f"{v:12.4f}" for v in first)
+        last_str = ", ".join(f"{v:12.4f}" for v in last)
+
+        print(f"                                       [{first_str}, ..., {last_str}]")
+
+    print("                                      ],")
+    print("                                     ]")
+    print(f"                                     sum = {t.sum().item():.6f}\n")
+
+
+def debug_hook(name):
+    def fn(_m, input, output):
+        if isinstance(input, torch.Tensor):
+            summarize(input, name + "_in")
+        elif isinstance(input, (tuple, list)) and isinstance(input[0], torch.Tensor):
+            summarize(input[0], name + "_in")
+        if isinstance(output, torch.Tensor):
+            summarize(output, name + "_out")
+        elif isinstance(output, (tuple, list)) and isinstance(output[0], torch.Tensor):
+            summarize(output[0], name + "_out")
+
+    return fn
+
+
+unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
+
+parser = argparse.ArgumentParser(description="Process model with specified path")
+parser.add_argument("--model-path", "-m", help="Path to the model")
+args = parser.parse_args()
+
+model_path = os.environ.get("MODEL_PATH", args.model_path)
+if model_path is None:
+    parser.error(
+        "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
+    )
+
+config = AutoConfig.from_pretrained(model_path)
+
+print("Model type:       ", config.model_type)
+print("Vocab size:       ", config.vocab_size)
+print("Hidden size:      ", config.hidden_size)
+print("Number of layers: ", config.num_hidden_layers)
+print("BOS token id:     ", config.bos_token_id)
+print("EOS token id:     ", config.eos_token_id)
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+config = AutoConfig.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = (
+        f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    )
+    class_name = f"{unreleased_model_name}ForCausalLM"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(
+            importlib.import_module(unreleased_module_path), class_name
+        )
+        model = model_class.from_pretrained(
+            model_path
+        )  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, device_map="auto", offload_folder="offload"
+    )
+
+for name, module in model.named_modules():
+    if len(list(module.children())) == 0:  # only leaf modules
+        module.register_forward_hook(debug_hook(name))
+
+model_name = os.path.basename(model_path)
+# Printing the Model class to allow for easier debugging. This can be useful
+# when working with models that have not been publicly released yet and this
+# migth require that the concrete class is imported and used directly instead
+# of using AutoModelForCausalLM.
+print(f"Model class: {model.__class__.__name__}")
+
+prompt = "Hello, my name is"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+print(f"Input tokens: {input_ids}")
+print(f"Input text: {repr(prompt)}")
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")
+
+with torch.no_grad():
+    outputs = model(input_ids)
+    logits = outputs.logits
+
+    # Extract logits for the last token (next token prediction)
+    last_logits = logits[0, -1, :].cpu().numpy()
+
+    print(f"Logits shape: {logits.shape}")
+    print(f"Last token logits shape: {last_logits.shape}")
+    print(f"Vocab size: {len(last_logits)}")
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}.txt"
+
+    # Save to file for comparison
+    last_logits.astype(np.float32).tofile(bin_filename)
+
+    # Also save as text file for easy inspection
+    with open(txt_filename, "w") as f:
+        for i, logit in enumerate(last_logits):
+            f.write(f"{i}: {logit:.6f}\n")
+
+    # Print some sample logits for quick verification
+    print(f"First 10 logits: {last_logits[:10]}")
+    print(f"Last 10 logits: {last_logits[-10:]}")
+
+    # Show top 5 predicted tokens
+    top_indices = np.argsort(last_logits)[-5:][::-1]
+    print("Top 5 predictions:")
+    for idx in top_indices:
+        token = tokenizer.decode([idx])
+        print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
+
+    print(f"Saved bin logits to: {bin_filename}")
+    print(f"Saved txt logist to: {txt_filename}")

examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh

@@ -0,0 +1,42 @@
+#!/usr/bin/env bash
+
+set -e
+
+MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
+MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+
+if [ -t 0 ]; then
+    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+else
+    # Process piped JSON data and convert to binary (matching logits.cpp format)
+    TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)
+    python3 -c "
+import json
+import sys
+import struct
+
+data = json.load(sys.stdin)
+
+# Flatten all embeddings completely
+flattened = []
+for item in data:
+    embedding = item['embedding']
+    for token_embedding in embedding:
+        flattened.extend(token_embedding)
+
+print(f'Total embedding values: {len(flattened)}', file=sys.stderr)
+
+# Write as binary floats - matches logitc.cpp fwrite format
+with open('$TEMP_FILE', 'wb') as f:
+    for value in flattened:
+        f.write(struct.pack('f', value))
+"
+    CPP_EMBEDDINGS="$TEMP_FILE"
+    trap "rm -f $TEMP_FILE" EXIT
+fi
+
+python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
+    --cpp-embeddings $CPP_EMBEDDINGS \
+    --prompt "Hello world today"
+

examples/model-conversion/scripts/embedding/convert-model.sh

@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+
+set -e
+
+MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
+OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
+TYPE="${OUTTYPE:-f16}"
+METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
+CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+
+echo "Model path: ${EMBEDDING_MODEL_PATH}"
+echo "Model name: ${MODEL_NAME}"
+echo "Data  type: ${TYPE}"
+echo "Converted model path:: ${CONVERTED_MODEL}"
+python ../../convert_hf_to_gguf.py --verbose \
+    ${EMBEDDING_MODEL_PATH} \
+    --outfile ${CONVERTED_MODEL} \
+    --outtype ${TYPE}
+
+echo ""
+echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"
+echo "export CONVERTED_EMBEDDING_MODEL=$(realpath ${CONVERTED_MODEL})"

examples/model-conversion/scripts/embedding/modelcard.template

@@ -0,0 +1,48 @@
+---
+base_model:
+- {base_model}
+---
+# {model_name} GGUF
+
+Recommended way to run this model:
+
+```sh
+llama-server -hf {namespace}/{model_name}-GGUF --embeddings
+```
+
+Then the endpoint can be accessed at http://localhost:8080/embedding, for
+example using `curl`:
+```console
+curl --request POST \
+    --url http://localhost:8080/embedding \
+    --header "Content-Type: application/json" \
+    --data '{{"input": "Hello embeddings"}}' \
+    --silent
+```
+
+Alternatively, the `llama-embedding` command line tool can be used:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF --verbose-prompt -p "Hello embeddings"
+```
+
+#### embd_normalize
+When a model uses pooling, or the pooling method is specified using `--pooling`,
+the normalization can be controlled by the `embd_normalize` parameter.
+
+The default value is `2` which means that the embeddings are normalized using
+the Euclidean norm (L2). Other options are:
+* -1 No normalization
+*  0 Max absolute
+*  1 Taxicab
+*  2 Euclidean/L2
+* \>2 P-Norm
+
+This can be passed in the request body to `llama-server`, for example:
+```sh
+    --data '{{"input": "Hello embeddings", "embd_normalize": -1}}' \
+```
+
+And for `llama-embedding`, by passing `--embd-normalize <value>`, for example:
+```sh
+llama-embedding -hf {namespace}/{model_name}-GGUF  --embd-normalize -1 -p "Hello embeddings"
+```

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+
+set -e
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_EMBEDDING_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-logits -j8
+
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import numpy as np
+import importlib
+from pathlib import Path
+
+from transformers import AutoTokenizer, AutoConfig, AutoModel
+import torch
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
+
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+if unreleased_model_name:
+    model_name_lower = unreleased_model_name.lower()
+    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+    class_name = f"{unreleased_model_name}Model"
+    print(f"Importing unreleased model module: {unreleased_module_path}")
+
+    try:
+        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+    except (ImportError, AttributeError) as e:
+        print(f"Failed to import or load model: {e}")
+        exit(1)
+else:
+    model = AutoModel.from_pretrained(model_path)
+print(f"Model class: {type(model)}")
+#print(f"Model file: {type(model).__module__}")
+config = AutoConfig.from_pretrained(model_path)
+
+model_name = os.path.basename(model_path)
+
+texts = [ "Hello world today" ]
+
+encoded = tokenizer(
+    texts,
+    padding=True,
+    truncation=True,
+    return_tensors="pt"
+)
+
+tokens = encoded['input_ids'][0]
+token_strings = tokenizer.convert_ids_to_tokens(tokens)
+for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+    print(f"{token_id:6d} -> '{token_str}'")
+
+with torch.no_grad():
+    outputs = model(**encoded)
+    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
+    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+
+    print(f"Hidden states shape: {hidden_states.shape}")
+    print(f"All embeddings shape: {all_embeddings.shape}")
+    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+
+    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
+    n_embd = all_embeddings.shape[1]
+    n_embd_count = all_embeddings.shape[0]
+
+    print()  # Empty line to match C++ output
+
+    for j in range(n_embd_count):
+        embedding = all_embeddings[j]
+        print(f"embedding {j}: ", end="")
+
+        # Print first 3 values
+        for i in range(min(3, n_embd)):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print(" ... ", end="")
+
+        # Print last 3 values
+        for i in range(n_embd - 3, n_embd):
+            print(f"{embedding[i]:9.6f} ", end="")
+
+        print()  # New line
+
+    print()  # Final empty line to match C++ output
+
+    data_dir = Path("data")
+    data_dir.mkdir(exist_ok=True)
+    bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
+    txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
+
+    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
+    flattened_embeddings = all_embeddings.flatten()
+    flattened_embeddings.astype(np.float32).tofile(bin_filename)
+
+    with open(txt_filename, "w") as f:
+        f.write(f"# Model class: {model_name}\n")
+        f.write(f"# Tokens: {token_strings}\n")
+        f.write(f"# Shape: {all_embeddings.shape}\n")
+        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
+
+        for j in range(n_embd_count):
+            f.write(f"# Token {j} ({token_strings[j]}):\n")
+            for i, value in enumerate(all_embeddings[j]):
+                f.write(f"{j}_{i}: {value:.6f}\n")
+            f.write("\n")
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
+    print("")
+    print(f"Saved bin embeddings to: {bin_filename}")
+    print(f"Saved txt embeddings to: {txt_filename}")

examples/model-conversion/scripts/utils/check-nmse.py

@@ -0,0 +1,174 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import sys
+import os
+import argparse
+from pathlib import Path
+
+def calculate_nmse(reference, test):
+    mse = np.mean((test - reference) ** 2)
+    ref_var = np.var(reference)
+    if ref_var == 0:
+        nmse = float('inf') if mse > 0 else 0.0
+        return mse, mse, ref_var
+
+    nmse = mse / ref_var
+
+    return nmse, mse, ref_var
+
+def load_logits(file_path):
+    if not os.path.exists(file_path):
+        raise FileNotFoundError(f"File not found: {file_path}")
+
+    if file_path.suffix == '.npy':
+        return np.load(file_path)
+    elif file_path.suffix == '.bin':
+        return np.fromfile(file_path, dtype=np.float32)
+    else:
+        # Try to load as text file
+        try:
+            # If it has index format "0: value", extract just values
+            data = []
+            with open(file_path, 'r') as f:
+                for line in f:
+                    if ':' in line:
+                        # Format: "index: value"
+                        value = float(line.split(':')[1].strip())
+                    else:
+                        # Just the value
+                        value = float(line.strip())
+                    data.append(value)
+            return np.array(data, dtype=np.float32)
+        except:
+            return np.loadtxt(file_path, dtype=np.float32)
+
+def interpret_nmse(nmse):
+    """Provide interpretation of NMSE value"""
+    if nmse == 0:
+        return "Perfect match", "🎉"
+    elif nmse < 1e-6:
+        return "Essentially identical", "✅"
+    elif nmse < 1e-4:
+        return "Excellent match", "✅"
+    elif nmse < 1e-3:
+        return "Very good match", "👍"
+    elif nmse < 1e-2:
+        return "Good match", "👍"
+    elif nmse < 0.1:
+        return "Acceptable match", "⚠️"
+    elif nmse < 1.0:
+        return "Poor match", "❌"
+    else:
+        return "Very poor match (worse than noise)", "❌"
+
+def main():
+    parser = argparse.ArgumentParser(description='Validate model logits')
+    parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
+    args = parser.parse_args()
+
+    model_name = os.path.splitext(os.path.basename(args.model_path))[0]
+    data_dir = Path("data")
+
+    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
+    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    print(f"Model name: {model_name}")
+    print(f"PyTorch logits file: {pytorch_file}")
+    print(f"llama.cpp logits file: {llamacpp_file}")
+
+    reference_file = pytorch_file
+    test_file = llamacpp_file
+
+    print("📊 NMSE Check for Model Comparison")
+    print("=" * 50)
+    print(f"Reference (ground truth): {reference_file}")
+    print(f"Test (to evaluate):       {test_file}")
+    print()
+
+    try:
+        print("Loading reference logits...")
+        reference = load_logits(reference_file)
+        print(f"  Shape: {reference.shape}, Type: {reference.dtype}")
+
+        print("Loading test logits...")
+        test = load_logits(test_file)
+        print(f"  Shape: {test.shape}, Type: {test.dtype}")
+
+        # Check shapes match
+        if reference.shape != test.shape:
+            print(f"\n❌ Error: Shape mismatch!")
+            print(f"  Reference: {reference.shape}")
+            print(f"  Test: {test.shape}")
+            sys.exit(1)
+
+        print(f"\n✅ Shapes match: {reference.shape}")
+
+        nmse, mse, ref_var = calculate_nmse(reference, test)
+
+        # Additional metrics
+        max_abs_error = np.max(np.abs(test - reference))
+        mean_abs_error = np.mean(np.abs(test - reference))
+
+        # Results
+        print(f"\n📈 METRICS")
+        print("=" * 30)
+        print(f"MSE (Mean Squared Error):     {mse:.6e}")
+        print(f"Reference Variance:           {ref_var:.6e}")
+        print(f"NMSE:                         {nmse:.6e}")
+        print(f"Max Absolute Error:           {max_abs_error:.6f}")
+        print(f"Mean Absolute Error:          {mean_abs_error:.6f}")
+
+        # NMSE in dB (common in signal processing)
+        if nmse > 0:
+            nmse_db = 10 * np.log10(nmse)
+            print(f"NMSE (dB):                    {nmse_db:.2f} dB")
+
+        # Interpretation
+        interpretation, emoji = interpret_nmse(nmse)
+        print(f"\n🎯 INTERPRETATION")
+        print("=" * 30)
+        print(f"{emoji} {interpretation}")
+
+        # Detailed guidance
+        print(f"\n📋 GUIDANCE")
+        print("=" * 30)
+        if nmse < 1e-3:
+            print("✅ EXCELLENT: Your GGML conversion is working very well!")
+            print("   The differences are negligible for practical use.")
+        elif nmse < 1e-2:
+            print("👍 GOOD: Your GGML conversion is working well.")
+            print("   Small differences are likely due to precision/quantization.")
+        elif nmse < 0.1:
+            print("⚠️  ACCEPTABLE: Conversion is working but with some differences.")
+            print("   Check if you're using quantization (Q4, Q8, etc.)")
+            print("   Test generation quality to see if it's acceptable.")
+        else:
+            print("❌ PROBLEMATIC: Large differences detected.")
+            print("   Check your conversion process for potential issues.")
+            print("   Verify you're using the same model weights.")
+
+        # NMSE benchmarks
+        print(f"\n📚 NMSE BENCHMARKS")
+        print("=" * 30)
+        print("< 1e-6:  Essentially identical")
+        print("< 1e-4:  Excellent (typical for good conversions)")
+        print("< 1e-3:  Very good")
+        print("< 1e-2:  Good (acceptable for most use cases)")
+        print("< 0.1:   Acceptable (may need verification)")
+        print("> 1.0:   Poor (worse than random)")
+
+        # Exit code based on NMSE
+        if nmse < 1e-2:
+            print(f"\n✅ RESULT: PASS (NMSE = {nmse:.2e})")
+            sys.exit(0)
+        else:
+            print(f"\n❌ RESULT: NEEDS REVIEW (NMSE = {nmse:.2e})")
+            sys.exit(1)
+
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/create-collection-add-model.sh

@@ -0,0 +1,8 @@
+
+#!/usr/bin/env bash
+
+COLLECTION_SLUG=$(python ./create_collection.py --return-slug)
+echo "Created collection: $COLLECTION_SLUG"
+
+# Use it in the next command
+python add_model_to_collection.py "$COLLECTION_SLUG" "username/my-model"

examples/model-conversion/scripts/utils/curl-embedding-server.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+curl --request POST \
+    --url http://localhost:8080/embedding \
+    --header "Content-Type: application/json" \
+    --data '{"input": "Hello world today"}' \
+    --silent

examples/model-conversion/scripts/utils/hf-add-model-to-collection.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import sys
+
+def add_model_to_collection(collection_slug, model_id, note=""):
+    """
+    Add a model to an existing collection
+
+    Args:
+        collection_slug: The slug of the collection (e.g., "username/collection-name-12345")
+        model_id: The model repository ID (e.g., "username/model-name")
+        note: Optional note about the model
+
+    Returns:
+        True if successful, False if failed
+    """
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        user_info = api.whoami()
+        print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Verify the model exists
+        print(f"🔍 Checking if model exists: {model_id}")
+        try:
+            model_info = api.model_info(model_id)
+        except Exception as e:
+            print(f"❌ Model not found or not accessible: {model_id}")
+            print(f"Error: {e}")
+            return False
+
+        print(f"📚 Adding model to collection...")
+        api.add_collection_item(
+            collection_slug=collection_slug,
+            item_id=model_id,
+            item_type="model",
+            note=note
+        )
+
+        print(f"✅ Model added to collection successfully!")
+        print(f"🔗 Collection URL: https://huggingface.co/collections/{collection_slug}")
+
+        return True
+
+    except Exception as e:
+        print(f"❌ Error adding model to collection: {e}")
+        return False
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Add model to a Huggingface Collection')
+    parser.add_argument('--collection', '-c', help='The collection slug username/collection-hash', required=True)
+    parser.add_argument('--model', '-m', help='The model to add to the Collection', required=True)
+    parser.add_argument('--note', '-n', help='An optional note/description', required=False)
+    args = parser.parse_args()
+
+    collection = args.collection
+    model = args.model
+    note = args.note
+
+    success = add_model_to_collection(
+        collection_slug=collection,
+        model_id=model,
+        note=note
+    )
+
+    if success:
+        print("\n🎉 Model added successfully!")
+    else:
+        print("\n❌ Failed to add model to collection")
+        sys.exit(1)
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-collection.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+import sys
+
+
+def create_collection(title, description, private=False, namespace=None, return_slug=False):
+    """
+    Create a new collection on Hugging Face
+
+    Args:
+        title: Collection title
+        description: Collection description
+        private: Whether the collection should be private (default: False)
+        namespace: Optional namespace (defaults to your username)
+
+    Returns:
+        Collection object if successful, None if failed
+    """
+
+    # Check if HF_TOKEN is available
+    token = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_HUB_TOKEN")
+    if not token:
+        print("❌ No HF_TOKEN or HUGGINGFACE_HUB_TOKEN found in environment variables")
+        print("Please set your Hugging Face token as an environment variable")
+        return None
+
+    # Initialize API
+    api = HfApi()
+
+    try:
+        # Test authentication first
+        user_info = api.whoami()
+        if not return_slug:
+            print(f"✅ Authenticated as: {user_info['name']}")
+
+        # Create the collection
+        if not return_slug:
+            print(f"📚 Creating collection: '{title}'...")
+        collection = api.create_collection(
+            title=title,
+            description=description,
+            private=private,
+            namespace=namespace
+        )
+
+        if not return_slug:
+            print(f"✅ Collection created successfully!")
+            print(f"📋 Collection slug: {collection.slug}")
+            print(f"🔗 Collection URL: https://huggingface.co/collections/{collection.slug}")
+
+        return collection
+
+    except Exception as e:
+        print(f"❌ Error creating collection: {e}")
+        return None
+
+def main():
+    # This script requires that the environment variable HF_TOKEN is set with your
+    # Hugging Face API token.
+    api = HfApi()
+
+    parser = argparse.ArgumentParser(description='Create a Huggingface Collection')
+    parser.add_argument('--name', '-n', help='The name/title of the Collection', required=True)
+    parser.add_argument('--description', '-d', help='The description for the Collection', required=True)
+    parser.add_argument('--namespace', '-ns', help='The namespace to add the Collection to', required=True)
+    parser.add_argument('--private', '-p', help='Create a private Collection', action='store_true')  # Fixed
+    parser.add_argument('--return-slug', '-s', help='Only output the collection slug', action='store_true')  # Fixed
+
+    args = parser.parse_args()
+
+    name = args.name
+    description = args.description
+    private = args.private
+    namespace = args.namespace
+    return_slug = args.return_slug
+
+    if not return_slug:
+        print("🚀 Creating Hugging Face Collection")
+        print(f"Title: {name}")
+        print(f"Description: {description}")
+        print(f"Namespace: {namespace}")
+        print(f"Private: {private}")
+
+    collection = create_collection(
+        title=name,
+        description=description,
+        private=private,
+        namespace=namespace,
+        return_slug=return_slug
+    )
+
+    if collection:
+        if return_slug:
+            print(collection.slug)
+        else:
+            print("\n🎉 Collection created successfully!")
+            print(f"Use this slug to add models: {collection.slug}")
+    else:
+        print("\n❌ Failed to create collection")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/hf-create-model.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+def load_template_and_substitute(template_path, **kwargs):
+    try:
+        with open(template_path, 'r', encoding='utf-8') as f:
+            template_content = f.read()
+
+        return template_content.format(**kwargs)
+    except FileNotFoundError:
+        print(f"Template file '{template_path}' not found!")
+        return None
+    except KeyError as e:
+        print(f"Missing template variable: {e}")
+        return None
+
+parser = argparse.ArgumentParser(description='Create a new Hugging Face model repository')
+parser.add_argument('--model-name', '-m', help='Name for the model', required=True)
+parser.add_argument('--namespace', '-ns', help='Namespace to add the model to', required=True)
+parser.add_argument('--org-base-model', '-b', help='Original Base model name', default="")
+parser.add_argument('--no-card', action='store_true', help='Skip creating model card')
+parser.add_argument('--private', '-p', action='store_true', help='Create private model')
+parser.add_argument('--embedding', '-e', action='store_true', help='Use embedding model card template')
+parser.add_argument('--dry-run', '-d', action='store_true', help='Print repository info and template without creating repository')
+
+args = parser.parse_args()
+
+repo_id = f"{args.namespace}/{args.model_name}-GGUF"
+print("Repository ID: ", repo_id)
+
+repo_url = None
+if not args.dry_run:
+    repo_url = api.create_repo(
+        repo_id=repo_id,
+        repo_type="model",
+        private=args.private,
+        exist_ok=False
+    )
+
+if not args.no_card:
+    if args.embedding:
+        template_path = "scripts/embedding/modelcard.template"
+    else:
+        template_path = "scripts/causal/modelcard.template"
+
+    print("Template path: ", template_path)
+
+    model_card_content = load_template_and_substitute(
+        template_path,
+        model_name=args.model_name,
+        namespace=args.namespace,
+        base_model=args.org_base_model,
+    )
+
+    if args.dry_run:
+        print("\nTemplate Content:\n")
+        print(model_card_content)
+    else:
+        if model_card_content:
+            api.upload_file(
+                path_or_fileobj=model_card_content.encode('utf-8'),
+                path_in_repo="README.md",
+                repo_id=repo_id
+            )
+            print("Model card created successfully.")
+        else:
+            print("Failed to create model card.")
+
+if not args.dry_run and repo_url:
+    print(f"Repository created: {repo_url}")
+
+

examples/model-conversion/scripts/utils/hf-upload-gguf-model.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+
+from huggingface_hub import HfApi
+import argparse
+import os
+
+def upload_gguf_file(local_file_path, repo_id, filename_in_repo=None):
+    """
+    Upload a GGUF file to a Hugging Face model repository
+
+    Args:
+        local_file_path: Path to your local GGUF file
+        repo_id: Your repository ID (e.g., "username/model-name")
+        filename_in_repo: Optional custom name for the file in the repo
+    """
+
+    if not os.path.exists(local_file_path):
+        print(f"❌ File not found: {local_file_path}")
+        return False
+
+    if filename_in_repo is None:
+        filename_in_repo = os.path.basename(local_file_path)
+
+    if filename_in_repo is None or filename_in_repo == "":
+        filename_in_repo = os.path.basename(local_file_path)
+
+    print(f"📤 Uploading {local_file_path} to {repo_id}/{filename_in_repo}")
+
+    api = HfApi()
+
+    try:
+        api.upload_file(
+            path_or_fileobj=local_file_path,
+            path_in_repo=filename_in_repo,
+            repo_id=repo_id,
+            repo_type="model",
+            commit_message=f"Upload {filename_in_repo}"
+        )
+
+        print("✅ Upload successful!")
+        print(f"🔗 File available at: https://huggingface.co/{repo_id}/blob/main/{filename_in_repo}")
+        return True
+
+    except Exception as e:
+        print(f"❌ Upload failed: {e}")
+        return False
+
+# This script requires that the environment variable HF_TOKEN is set with your
+# Hugging Face API token.
+api = HfApi()
+
+parser = argparse.ArgumentParser(description='Upload a GGUF model to a Huggingface model repository')
+parser.add_argument('--gguf-model-path', '-m', help='The GGUF model file to upload', required=True)
+parser.add_argument('--repo-id', '-r', help='The repository to upload to', required=True)
+parser.add_argument('--name', '-o', help='The name in the model repository', required=False)
+args = parser.parse_args()
+
+upload_gguf_file(args.gguf_model_path, args.repo_id, args.name)

examples/model-conversion/scripts/utils/inspect-converted-model.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+../../gguf-py/gguf/scripts/gguf_dump.py $CONVERTED_MODEL

examples/model-conversion/scripts/utils/inspect-org-model.py

@@ -0,0 +1,67 @@
+#!/usr/bin/env python3
+
+import argparse
+import os
+import json
+from safetensors import safe_open
+from collections import defaultdict
+
+parser = argparse.ArgumentParser(description='Process model with specified path')
+parser.add_argument('--model-path', '-m', help='Path to the model')
+args = parser.parse_args()
+
+model_path = os.environ.get('MODEL_PATH', args.model_path)
+if model_path is None:
+    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+
+# Check if there's an index file (multi-file model)
+index_path = os.path.join(model_path, "model.safetensors.index.json")
+single_file_path = os.path.join(model_path, "model.safetensors")
+
+if os.path.exists(index_path):
+    # Multi-file model
+    print("Multi-file model detected")
+
+    with open(index_path, 'r') as f:
+        index_data = json.load(f)
+
+    # Get the weight map (tensor_name -> file_name)
+    weight_map = index_data.get("weight_map", {})
+
+    # Group tensors by file for efficient processing
+    file_tensors = defaultdict(list)
+    for tensor_name, file_name in weight_map.items():
+        file_tensors[file_name].append(tensor_name)
+
+    print("Tensors in model:")
+
+    # Process each shard file
+    for file_name, tensor_names in file_tensors.items():
+        file_path = os.path.join(model_path, file_name)
+        print(f"\n--- From {file_name} ---")
+
+        with safe_open(file_path, framework="pt") as f:  # type: ignore
+            for tensor_name in sorted(tensor_names):
+                tensor = f.get_tensor(tensor_name)
+                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+elif os.path.exists(single_file_path):
+    # Single file model (original behavior)
+    print("Single-file model detected")
+
+    with safe_open(single_file_path, framework="pt") as f:  # type: ignore
+        keys = f.keys()
+        print("Tensors in model:")
+        for key in sorted(keys):
+            tensor = f.get_tensor(key)
+            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+
+else:
+    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
+    print("Available files:")
+    if os.path.exists(model_path):
+        for item in sorted(os.listdir(model_path)):
+            print(f"  {item}")
+    else:
+        print(f"  Directory {model_path} does not exist")
+    exit(1)

examples/model-conversion/scripts/utils/perplexity-gen.sh

@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+mkdir -p ppl
+OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
+echo "Model: $CONVERTED_MODEL"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+    -f ppl/wikitext-2-raw/wiki.test.raw \
+    --kl-divergence-base $OUTPUTFILE
+
+echo "Generated logits in $OUTPUTFILE"
+

examples/model-conversion/scripts/utils/perplexity-run-simple.sh

@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+# Check if data/wikitext-2-raw directory exists
+if [ ! -d "ppl/wikitext-2-raw" ]; then
+    echo "ppl/wikitext-2-raw directory does not exist. Downloading..." >&2
+    mkdir -p ppl
+    pushd ppl
+    ./../../../scripts/get-wikitext-2.sh
+    popd
+fi
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+
+

examples/model-conversion/scripts/utils/perplexity-run.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+set -e
+
+QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+
+if [ -z "$QUANTIZED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. QUANTIZED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ ! -f ${LOGITS_FILE} ]; then
+    echo "Error: logits file '${LOGITS_FILE} was not found"
+    echo "Did you run the perplexity-gen.sh script?"
+    exit 1
+fi
+
+echo "Model: $QUANTIZED_MODEL"
+echo "Data file: $LOGITS_FILE"
+
+cmake --build ../../build --target llama-perplexity -j8
+
+../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+    --kl-divergence-base $LOGITS_FILE \
+    --kl-divergence

examples/model-conversion/scripts/utils/quantize.sh

@@ -0,0 +1,48 @@
+#!/usr/bin/env bash
+
+set -e
+
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
+TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
+OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
+QUANTIZED_MODEL=$CONVERTED_MODEL
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+if [ -z "$QUANTIZED_TYPE" ]; then
+    echo "Error: QUANTIZED_TYPE is required" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+# Process the quantized model filename
+if [[ "$QUANTIZED_MODEL" == *.gguf ]]; then
+    # Remove .gguf suffix, add quantized type, then add .gguf back
+    BASE_NAME="${QUANTIZED_MODEL%.gguf}"
+    QUANTIZED_MODEL="${BASE_NAME}-${QUANTIZED_TYPE}.gguf"
+else
+    echo "Error: QUANTIZED_MODEL must end with .gguf extension" >&2
+    exit 1
+fi
+
+cmake --build ../../build --target llama-quantize -j8
+
+echo $TOKEN_EMBD_TYPE
+echo $OUTPUT_TYPE
+
+CMD_ARGS=("../../build/bin/llama-quantize")
+[[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
+[[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
+CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")
+
+"${CMD_ARGS[@]}"
+
+echo "Quantized model saved to: $QUANTIZED_MODEL"

examples/model-conversion/scripts/utils/run-embedding-server.sh

@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+
+set -e
+#
+# First try command line argument, then environment variable, then file
+CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+
+# Final check if we have a model path
+if [ -z "$CONVERTED_MODEL" ]; then
+    echo "Error: Model path must be provided either as:" >&2
+    echo "  1. Command line argument" >&2
+    echo "  2. CONVERTED_MODEL environment variable" >&2
+    exit 1
+fi
+
+echo $CONVERTED_MODEL
+
+cmake --build ../../build --target llama-server
+
+../../build/bin/llama-server -m $CONVERTED_MODEL \
+    --embedding \
+    --pooling none

examples/model-conversion/scripts/utils/semantic_check.py

@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+
+import numpy as np
+import argparse
+import os
+import importlib
+
+from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
+
+unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
+
+def cosine_similarity(a, b=None):
+    a = np.asarray(a)
+    if b is None:
+        b = a
+    else:
+        b = np.asarray(b)
+
+    if a.ndim == 1:
+        a = a.reshape(1, -1)
+    if b.ndim == 1:
+        b = b.reshape(1, -1)
+
+    a_norms = np.linalg.norm(a, axis=1, keepdims=True)
+    b_norms = np.linalg.norm(b, axis=1, keepdims=True)
+
+    a_norms = np.where(a_norms == 0, 1e-8, a_norms)
+    b_norms = np.where(b_norms == 0, 1e-8, b_norms)
+
+    a_normalized = a / a_norms
+    b_normalized = b / b_norms
+
+    # Compute cosine similarity
+    return np.dot(a_normalized, b_normalized.T)
+
+def load_embeddings_from_file(filename, n_tokens, n_embd):
+    embeddings = np.fromfile(filename, dtype=np.float32)
+    return embeddings.reshape(n_tokens, n_embd)
+
+def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
+    np.set_printoptions(suppress=True, precision=6)
+    print("pytorch embeddings:");
+    print(python_emb)
+    print("llama.cpp embeddings:");
+    print(cpp_emb)
+    print(f"\n=== Prompt: '{prompt}' ===")
+    print(f"Tokens: {tokens}")
+    print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
+
+    n_tokens = len(tokens)
+
+    # 1. Direct embedding comparison
+    print(f"\n1. Raw Embedding Magnitude Comparison:")
+    # Check if the distance of each token embedding from the origin and compare
+    # if the vectors are on the same "sphere". This does not tell us about
+    # direction (meaning of the token embedding), just magnitude.
+    for i in range(n_tokens):
+        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+        ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
+        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+
+    # 2. Cosine similarity between tokens within each model
+    # Here we check the direction of token embeddings to see if the have the
+    # same meaning (similarity). This is done by calculating cosine similarity
+    # of a pair of token embeddings within each model.
+    print(f"\n2. Within-Model Token Similarities:")
+    print("   Python model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    print("   llama.cpp model:")
+    for i in range(n_tokens):
+        for j in range(i+1, n_tokens):
+            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+
+    # 3. Cross-model similarity (same token position)
+    print(f"\n3. Cross-Model Same-Token Similarities:")
+    for i in range(n_tokens):
+        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+
+    # 4. Similarity matrix comparison
+    print(f"\n4. Similarity Matrix Differences:")
+    py_sim_matrix = cosine_similarity(python_emb)
+    cpp_sim_matrix = cosine_similarity(cpp_emb)
+    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+
+    print(f"   Max difference: {np.max(diff_matrix):.4f}")
+    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+
+    return {
+        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+        'similarity_matrix_diff': diff_matrix,
+        'max_diff': np.max(diff_matrix),
+        'mean_diff': np.mean(diff_matrix),
+        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+    }
+
+def main():
+    parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
+    parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
+    parser.add_argument('--python-embeddings', '-pe', help='Path to pytorch embeddings "logits" binary file')
+    parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
+    parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
+    parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
+
+    args = parser.parse_args()
+
+    print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
+    print("=" * 70)
+
+    # Single prompt detailed comparison
+    print(f"\nTesting with prompt: '{args.prompt}'")
+
+    # Load the python model to get configuration information and also to load the tokenizer.
+    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    config = AutoConfig.from_pretrained(args.model_path)
+
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+        if args.causal:
+            class_name = f"{unreleased_model_name}ForCausalLM"
+        else:
+            class_name = f"{unreleased_model_name}Model"
+        print(f"Model class: {class_name}")
+        print(f"Importing unreleased model module: {unreleased_module_path}")
+
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(args.model_path)
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        if args.causal:
+            model = AutoModelForCausalLM.from_pretrained(args.model_path)
+        else:
+            model = AutoModel.from_pretrained(args.model_path)
+
+    encoded = tokenizer(args.prompt, return_tensors="pt")
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
+    n_tokens = len(tokens)
+    print(f"n_tokens: {n_tokens}");
+    print(f"hidden_size: {model.config.hidden_size}")
+
+    # Load binary embeddings from data directory.
+    llamacpp_embeddings = load_embeddings_from_file(args.cpp_embeddings, n_tokens, model.config.hidden_size)
+    python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)
+
+    # Run comparison
+    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
+
+    # Summary
+    print(f"\n=== SUMMARY ===")
+    avg_cross_sim = np.mean(results['cross_model_similarities'])
+    print(f"Average cross-model similarity: {avg_cross_sim:.4f}")
+    print(f"Similarity matrix RMS difference: {results['rms_diff']:.4f}")
+
+    # Quality assessment
+    if avg_cross_sim > 0.95:
+        print("✅ EXCELLENT: Models are highly similar")
+    elif avg_cross_sim > 0.90:
+        print("✅ VERY GOOD: Models are very similar")
+    elif avg_cross_sim > 0.80:
+        print("⚠️  GOOD: Models are reasonably similar")
+    elif avg_cross_sim > 0.70:
+        print("⚠️  FAIR: Models have some differences")
+    else:
+        print("❌ POOR: Models are significantly different")
+
+if __name__ == "__main__":
+    main()

examples/passkey/README.md

@@ -11,5 +11,5 @@ See the following PRs for more info:
 ### Usage
 
 ```bash
-make -j && ./llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
+llama-passkey -m ./models/llama-7b-v2/ggml-model-f16.gguf --junk 250
 ```

examples/retrieval/README.md

@@ -15,7 +15,7 @@ https://github.com/ggml-org/llama.cpp/pull/6193
 `retrieval` example can be tested as follows:
 
 ```bash
-make -j && ./llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+llama-retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
 ```
 
 This chunks and embeds all given files and starts a loop requesting query inputs:

examples/simple/simple.cpp

@@ -145,6 +145,20 @@ int main(int argc, char ** argv) {
 
     llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
 
+    if (llama_model_has_encoder(model)) {
+        if (llama_encode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return 1;
+        }
+
+        llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
+        if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
+            decoder_start_token_id = llama_vocab_bos(vocab);
+        }
+
+        batch = llama_batch_get_one(&decoder_start_token_id, 1);
+    }
+
     // main loop
 
     const auto t_main_start = ggml_time_us();

examples/speculative/speculative.cpp

@@ -244,7 +244,7 @@ int main(int argc, char ** argv) {
                     // stochastic verification
                     common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
-                    auto & dist_tgt = *common_sampler_get_candidates(smpl);
+                    auto & dist_tgt = *common_sampler_get_candidates(smpl, true);
 
                     float p_tgt = 0.0f;
                     float p_dft = 0.0f;
@@ -493,7 +493,7 @@ int main(int argc, char ** argv) {
 
                 common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
 
-                const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl);
+                const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl, true);
 
                 for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
                     LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",

examples/sycl/win-build-sycl.bat

@@ -18,8 +18,6 @@ if %errorlevel% neq 0 goto ERROR
 ::  for FP32
 cmake -G "Ninja" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
-::  build example/main only
-::  make main
 
 ::  build all binary
 cmake --build . -j

ggml/CMakeLists.txt

@@ -1,5 +1,5 @@
 cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
-project("ggml" C CXX)
+project("ggml" C CXX ASM)
 include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -129,10 +129,11 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
-option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
+option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
+option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
+option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
-option(GGML_NNPA             "ggml: enable nnpa"             OFF)  # temp disabled by default, see: https://github.com/ggml-org/llama.cpp/issues/14877
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
@@ -158,7 +159,6 @@ option(GGML_CUDA                            "ggml: use CUDA"
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)

ggml/include/ggml-backend.h

@@ -132,6 +132,8 @@ extern "C" {
         GGML_BACKEND_DEVICE_TYPE_CPU,
         // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
+        // integrated GPU device using host memory
+        GGML_BACKEND_DEVICE_TYPE_IGPU,
         // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
         GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
@@ -150,11 +152,21 @@ extern "C" {
 
     // all the device properties
     struct ggml_backend_dev_props {
+        // device name
         const char * name;
+        // device description
         const char * description;
+        // device free memory in bytes
         size_t memory_free;
+        // device total memory in bytes
         size_t memory_total;
+        // device type
         enum ggml_backend_dev_type type;
+        // device id
+        //   for PCI devices, this should be the PCI bus id formatted as "domain:bus:device.function" (e.g. "0000:01:00.0")
+        //   if the id is unknown, this should be NULL
+        const char * device_id;
+        // device capabilities
         struct ggml_backend_dev_caps caps;
     };
 
@@ -307,6 +319,9 @@ extern "C" {
     GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
     GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
+    // Split graph without allocating it
+    GGML_API void                 ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+
     // Allocate and compute graph on the backend scheduler
     GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // returns success
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);

ggml/include/ggml-cpu.h

@@ -101,7 +101,6 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
-    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -135,6 +134,7 @@ extern "C" {
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp32_to_i32 (const float *,     int32_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

ggml/include/ggml-metal.h

@@ -43,14 +43,8 @@ GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
-GGML_DEPRECATED(
-        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
-
 GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf

ggml/include/ggml-zdnn.h

@@ -7,8 +7,6 @@
 extern "C" {
 #endif
 
-GGML_BACKEND_API ggml_backend_t ggml_backend_zdnn_init(void);
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
 
 #ifdef __cplusplus

ggml/include/ggml.h

@@ -244,6 +244,13 @@
 #define GGML_MROPE_SECTIONS   4
 
 #define GGML_UNUSED(x) (void)(x)
+#ifdef __CUDACC__
+template<typename... Args>
+__host__ __device__ constexpr inline void ggml_unused_vars_impl(Args&&...) noexcept {}
+#define GGML_UNUSED_VARS(...) ggml_unused_vars_impl(__VA_ARGS__)
+#else
+#define GGML_UNUSED_VARS(...) do { (void)sizeof((__VA_ARGS__, 0)); } while(0)
+#endif // __CUDACC__
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
@@ -504,7 +511,9 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_IM2COL_3D,
         GGML_OP_CONV_2D,
+        GGML_OP_CONV_3D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -1395,6 +1404,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // note: casting from f32 to i32 will discard the fractional part
     GGML_API struct ggml_tensor * ggml_cast(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1519,7 +1529,11 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // supports 3D: a->ne[2] == b->ne[1]
+    // supports 4D a:
+    // a     [n_embd, ne1, ne2, ne3]
+    // b I32 [n_rows, ne2, ne3, 1]
+    //
+    // return [n_embd, n_rows, ne2, ne3]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,  // data
@@ -1862,6 +1876,41 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    GGML_API struct ggml_tensor * ggml_im2col_3d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int64_t               IC,
+            int                   s0, // stride width
+            int                   s1, // stride height
+            int                   s2, // stride depth
+            int                   p0, // padding width
+            int                   p1, // padding height
+            int                   p2, // padding depth
+            int                   d0, // dilation width
+            int                   d1, // dilation height
+            int                   d2, // dilation depth
+            enum ggml_type        dst_type);
+
+    // a: [OC*IC, KD, KH, KW]
+    // b: [N*IC, ID, IH, IW]
+    // result: [N*OC, OD, OH, OW]
+    GGML_API struct ggml_tensor * ggml_conv_3d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                int64_t               IC,
+                int                   s0, // stride width
+                int                   s1, // stride height
+                int                   s2, // stride depth
+                int                   p0, // padding width
+                int                   p1, // padding height
+                int                   p2, // padding depth
+                int                   d0, // dilation width
+                int                   d1, // dilation height
+                int                   d2  // dilation depth
+        );
+
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
     // padding is zero
@@ -1933,6 +1982,23 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
+    GGML_API struct ggml_tensor * ggml_conv_3d_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // kernel [KW, KH, KD, IC * OC]
+            struct ggml_tensor  * b,   // input  [W, H, D, C * N]
+            int                   s0,  // stride
+            int                   s1,
+            int                   s2,
+            int                   p0,  // padding
+            int                   p1,
+            int                   p2,
+            int                   d0,  // dilation
+            int                   d1,
+            int                   d2,
+            int                   n_channels,
+            int                   n_batch,
+            int                   n_channels_out);
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -2023,6 +2089,19 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    GGML_API struct ggml_tensor * ggml_pad_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                  lp0,
+            int                  rp0,
+            int                  lp1,
+            int                  rp1,
+            int                  lp2,
+            int                  rp2,
+            int                  lp3,
+            int                  rp3
+            );
+
     // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
     GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
             struct ggml_context * ctx,

ggml/src/ggml-backend-impl.h

@@ -8,7 +8,7 @@
 extern "C" {
 #endif
 
-    #define GGML_BACKEND_API_VERSION 1
+    #define GGML_BACKEND_API_VERSION 2
 
     //
     // Backend buffer type
@@ -114,6 +114,9 @@ extern "C" {
         void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
         // wait for an event on on a different stream
         void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
+
+        // (optional) sort/optimize the nodes in the graph
+        void                      (*optimize_graph)    (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     };
 
     struct ggml_backend {

ggml/src/ggml-backend-reg.cpp

@@ -400,9 +400,8 @@ ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const
 
 ggml_backend_t ggml_backend_init_best(void) {
     ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
-    if (!dev) {
-        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    }
+    dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    dev = dev ? dev : ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     if (!dev) {
         return nullptr;
     }

ggml/src/ggml-backend.cpp

@@ -19,9 +19,8 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include <string>
-#include <vector>
 #include <algorithm>
+#include <vector>
 
 #ifdef __APPLE__
 #include <sys/types.h>
@@ -32,6 +31,7 @@
 // backend buffer type
 
 const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(buft);
     return buft->iface.get_name(buft);
 }
 
@@ -41,14 +41,17 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
         return ggml_backend_buffer_init(buft, {}, NULL, 0);
     }
 
+    GGML_ASSERT(buft);
     return buft->iface.alloc_buffer(buft, size);
 }
 
 size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(buft);
     return buft->iface.get_alignment(buft);
 }
 
 size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(buft);
     // get_max_size is optional, defaults to SIZE_MAX
     if (buft->iface.get_max_size) {
         return buft->iface.get_max_size(buft);
@@ -57,6 +60,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
 }
 
 size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) {
+    GGML_ASSERT(buft);
     // get_alloc_size is optional, defaults to ggml_nbytes
     if (buft->iface.get_alloc_size) {
         size_t size = buft->iface.get_alloc_size(buft, tensor);
@@ -67,6 +71,7 @@ size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const s
 }
 
 bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(buft);
     if (buft->iface.is_host) {
         return buft->iface.is_host(buft);
     }
@@ -74,6 +79,7 @@ bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
 }
 
 ggml_backend_dev_t ggml_backend_buft_get_device(ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(buft);
     return buft->device;
 }
 
@@ -111,10 +117,12 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
 }
 
 size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     return buffer->size;
 }
 
 void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     // get_base is optional if the buffer is zero-sized
     if (buffer->size == 0) {
         return NULL;
@@ -128,6 +136,7 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
 }
 
 enum ggml_status ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    GGML_ASSERT(buffer);
     // init_tensor is optional
     if (buffer->iface.init_tensor) {
         return buffer->iface.init_tensor(buffer, tensor);
@@ -136,6 +145,7 @@ enum ggml_status ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, s
 }
 
 void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    GGML_ASSERT(buffer);
     // clear is optional if the buffer is zero-sized
     if (buffer->size == 0) {
         return;
@@ -161,6 +171,7 @@ bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
 }
 
 void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(buffer);
     buffer->usage = usage;
 
     // FIXME: add a generic callback to the buffer interface
@@ -170,14 +181,17 @@ void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backe
 }
 
 enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     return buffer->usage;
 }
 
 ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     return buffer->buft;
 }
 
 void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     if (buffer->iface.reset) {
         buffer->iface.reset(buffer);
     }
@@ -216,6 +230,7 @@ void ggml_backend_free(ggml_backend_t backend) {
 }
 
 ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
+    GGML_ASSERT(backend);
     return ggml_backend_dev_buffer_type(backend->device);
 }
 
@@ -232,6 +247,8 @@ size_t ggml_backend_get_max_size(ggml_backend_t backend) {
 }
 
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(backend);
+    GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
@@ -243,6 +260,8 @@ void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor *
 }
 
 void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(backend);
+    GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
@@ -284,6 +303,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
 }
 
 void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -299,6 +319,7 @@ void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size
 }
 
 void ggml_backend_synchronize(ggml_backend_t backend) {
+    GGML_ASSERT(backend);
     if (backend->iface.synchronize == NULL) {
         return;
     }
@@ -307,18 +328,21 @@ void ggml_backend_synchronize(ggml_backend_t backend) {
 }
 
 ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend);
     GGML_ASSERT(backend->iface.graph_plan_create != NULL);
 
     return backend->iface.graph_plan_create(backend, cgraph);
 }
 
 void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend);
     GGML_ASSERT(backend->iface.graph_plan_free != NULL);
 
     backend->iface.graph_plan_free(backend, plan);
 }
 
 enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend);
     GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
 
     return backend->iface.graph_plan_compute(backend, plan);
@@ -331,22 +355,27 @@ enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_
 }
 
 enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend);
     return backend->iface.graph_compute(backend, cgraph);
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    GGML_ASSERT(backend);
     return ggml_backend_dev_supports_op(backend->device, op);
 }
 
 bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(backend);
     return ggml_backend_dev_supports_buft(backend->device, buft);
 }
 
 bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    GGML_ASSERT(backend);
     return ggml_backend_dev_offload_op(backend->device, op);
 }
 
 ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
+    GGML_ASSERT(backend);
     return backend->device;
 }
 
@@ -382,6 +411,7 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
         return;
     }
 
+    GGML_ASSERT(backend_dst);
     if (backend_dst->iface.cpy_tensor_async != NULL) {
         if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
             return;
@@ -413,38 +443,52 @@ void ggml_backend_event_free(ggml_backend_event_t event) {
 }
 
 void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend) {
+    GGML_ASSERT(backend);
     GGML_ASSERT(backend->iface.event_record != NULL);
 
     backend->iface.event_record(backend, event);
 }
 
 void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event);
     GGML_ASSERT(event->device->iface.event_synchronize);
 
     event->device->iface.event_synchronize(event->device, event);
 }
 
 void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend);
     GGML_ASSERT(backend->iface.event_wait != NULL);
 
     backend->iface.event_wait(backend, event);
 }
 
+static void ggml_backend_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend);
+    if (backend->iface.optimize_graph != NULL) {
+        backend->iface.optimize_graph(backend, cgraph);
+    }
+}
+
 // Backend device
 
 const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     return device->iface.get_name(device);
 }
 
 const char * ggml_backend_dev_description(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     return device->iface.get_description(device);
 }
 
 void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
+    GGML_ASSERT(device);
     device->iface.get_memory(device, free, total);
 }
 
 enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     return device->iface.get_type(device);
 }
 
@@ -454,18 +498,22 @@ void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_d
 }
 
 ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     return device->reg;
 }
 
 ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params) {
+    GGML_ASSERT(device);
     return device->iface.init_backend(device, params);
 }
 
 ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     return device->iface.get_buffer_type(device);
 }
 
 ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) {
+    GGML_ASSERT(device);
     if (device->iface.get_host_buffer_type == NULL) {
         return NULL;
     }
@@ -474,18 +522,22 @@ ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t
 }
 
 ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size) {
+    GGML_ASSERT(device);
     return device->iface.buffer_from_host_ptr(device, ptr, size, max_tensor_size);
 }
 
 bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    GGML_ASSERT(device);
     return device->iface.supports_op(device, op);
 }
 
 bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(device);
     return device->iface.supports_buft(device, buft);
 }
 
 bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    GGML_ASSERT(device);
     if (device->iface.offload_op != NULL) {
         return device->iface.offload_op(device, op);
     }
@@ -496,18 +548,22 @@ bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_te
 // Backend (reg)
 
 const char * ggml_backend_reg_name(ggml_backend_reg_t reg) {
+    GGML_ASSERT(reg);
     return reg->iface.get_name(reg);
 }
 
 size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg) {
+    GGML_ASSERT(reg);
     return reg->iface.get_device_count(reg);
 }
 
 ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(reg);
     return reg->iface.get_device(reg, index);
 }
 
 void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_ASSERT(reg);
     if (!reg->iface.get_proc_address) {
         return NULL;
     }
@@ -522,6 +578,7 @@ struct ggml_backend_multi_buffer_context {
 };
 
 static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
         ggml_backend_buffer_free(ctx->buffers[i]);
@@ -532,6 +589,7 @@ static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer)
 }
 
 static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    GGML_ASSERT(buffer);
     ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
         ggml_backend_buffer_clear(ctx->buffers[i], value);
@@ -567,10 +625,12 @@ ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer
 }
 
 bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     return buffer->iface.free_buffer == ggml_backend_multi_buffer_free_buffer;
 }
 
 void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(buffer);
     GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
     ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
     for (size_t i = 0; i < ctx->n_buffers; i++) {
@@ -598,7 +658,7 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
-#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#define GGML_SCHED_MAX_SPLIT_INPUTS 30
 #endif
 
 #ifndef GGML_SCHED_MAX_COPIES
@@ -849,7 +909,7 @@ static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, stru
 }
 
 // assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
-static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     // reset splits
     sched->n_splits = 0;
     sched->n_graph_inputs = 0;
@@ -1245,6 +1305,10 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
 
+        // Optimize this split of the graph. This needs to happen before we make graph_copy,
+        // so they are in sync.
+        ggml_backend_optimize_graph(sched->backends[split->backend_id], &split->graph);
+
         // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
             assert(graph_copy->size > (graph_copy->n_nodes + 1));
@@ -1350,17 +1414,22 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 }
 
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     struct ggml_backend_sched_split * splits = sched->splits;
 
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
+    ggml_tensor * prev_ids_tensor = nullptr;
+    std::vector<int32_t> ids;
+    std::vector<ggml_bitset_t> used_ids;
+
+    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
+        struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
         // copy the input tensors to the split backend
-        for (int j = 0; j < split->n_inputs; j++) {
-            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
-            struct ggml_tensor * input = split->inputs[j];
+        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
+            struct ggml_tensor * input = split->inputs[input_id];
             struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
 
             if (input->flags & GGML_TENSOR_FLAG_INPUT) {
@@ -1378,16 +1447,104 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
-                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
-                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+
+                // when offloading MoE weights, we can reduce the amount of data copied by copying only the experts that are used
+                ggml_tensor * node = split->graph.nodes[0];
+                if (split->graph.n_nodes > 0 &&
+                    ggml_backend_buffer_get_usage(input->buffer) == GGML_BACKEND_BUFFER_USAGE_WEIGHTS &&
+                    ggml_backend_buffer_is_host(input->buffer) && (
+                    (node->src[0] == input_cpy && node->op == GGML_OP_MUL_MAT_ID)
+                    //|| (node->src[1] == input_cpy && node->op == GGML_OP_ADD_ID) /* GGML_OP_ADD_ID weights are small and not worth splitting */
+                    )) {
+
+                    const int64_t n_expert   = node->op == GGML_OP_MUL_MAT_ID ? input->ne[2] : input->ne[1];
+                    const size_t expert_size = node->op == GGML_OP_MUL_MAT_ID ? input->nb[2] : input->nb[1];
+
                     ggml_backend_synchronize(input_backend);
-                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                    } else {
-                        ggml_backend_synchronize(split_backend);
+
+                    // get the ids
+                    ggml_tensor * ids_tensor = node->src[2];
+                    ggml_backend_t ids_backend = split_backend;
+
+                    // if the ids tensor is also an input of the split, it may not have been copied yet to the split backend
+                    // in that case, we use the original ids tensor
+                    for (int i = input_id + 1; i < split->n_inputs; i++) {
+                        if (ids_tensor == tensor_copy(split->inputs[i], split_backend_id, sched->cur_copy)) {
+                            ids_tensor = split->inputs[i];
+                            ids_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[i]);
+                            break;
+                        }
+                    }
+
+                    if (ids_tensor != prev_ids_tensor) {
+                        ids.resize(ggml_nbytes(ids_tensor) / sizeof(int32_t));
+                        ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
+                        ggml_backend_synchronize(ids_backend);
+
+                        // find the used experts
+                        used_ids.clear();
+                        used_ids.resize(ggml_bitset_size(n_expert));
+                        for (int64_t i1 = 0; i1 < ids_tensor->ne[1]; i1++) {
+                            for (int64_t i0 = 0; i0 < ids_tensor->ne[0]; i0++) {
+                                int32_t id = ids[i1 * ids_tensor->nb[1]/sizeof(int32_t) + i0 * ids_tensor->nb[0]/sizeof(int32_t)];
+                                GGML_ASSERT(id >= 0 && id < n_expert);
+                                ggml_bitset_set(used_ids.data(), id);
+                            }
+                        }
+
+                        prev_ids_tensor = ids_tensor;
+                    }
+
+                    // group consecutive experts and copy them together
+                    auto copy_experts = [&](int32_t first_id, int32_t last_id) {
+                        const size_t expert_offset = first_id * expert_size;
+                        const size_t expert_size_copy =  (last_id - first_id + 1) * expert_size;
+                        const size_t padding = std::min<size_t>(expert_size, 512);
+                        const size_t padding_end = last_id < n_expert - 1 ? padding : 0;
+
+                        ggml_backend_tensor_set_async(split_backend,
+                            input_cpy,
+                            (const uint8_t *)input->data + expert_offset, expert_offset,
+                            // copy a bit extra at the to ensure there are no NaNs in the padding of the last expert
+                            // this is necessary for MMQ in the CUDA backend
+                            expert_size_copy + padding_end);
+                    };
+
+                    int id = 0;
+                    while (!ggml_bitset_get(used_ids.data(), id)) {
+                        id++;
+                    }
+                    int32_t first_id = id;
+                    int32_t last_id = first_id;
+
+                    for (++id; id < n_expert; ++id) {
+                        if (!ggml_bitset_get(used_ids.data(), id)) {
+                            continue;
+                        }
+
+                        if (id == last_id + 1) {
+                            last_id = id;
+                            continue;
+                        }
+
+                        copy_experts(first_id, last_id);
+
+                        first_id = id;
+                        last_id = id;
+                    }
+                    copy_experts(first_id, last_id);
+                } else {
+                    // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                    // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                    if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                        ggml_backend_synchronize(input_backend);
+                        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                            ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                        } else {
+                            ggml_backend_synchronize(split_backend);
+                        }
+                        ggml_backend_tensor_copy(input, input_cpy);
                     }
-                    ggml_backend_tensor_copy(input, input_cpy);
                 }
             }
         }
@@ -1526,6 +1683,7 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
 }
 
 void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     // reset state for the next run
     if (!sched->is_reset) {
         ggml_hash_set_reset(&sched->hash_set);
@@ -1537,8 +1695,11 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
 }
 
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
 
+    ggml_backend_sched_reset(sched);
+
     ggml_backend_sched_synchronize(sched);
 
     ggml_backend_sched_split_graph(sched, measure_graph);
@@ -1553,6 +1714,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 }
 
 bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
     GGML_ASSERT(!sched->is_alloc);
 
@@ -1577,6 +1739,7 @@ enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, st
 }
 
 enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT(sched);
     if (!sched->is_reset && !sched->is_alloc) {
         ggml_backend_sched_reset(sched);
     }
@@ -1591,6 +1754,7 @@ enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sch
 }
 
 void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     for (int i = 0; i < sched->n_backends; i++) {
         ggml_backend_synchronize(sched->backends[i]);
     }
@@ -1603,28 +1767,34 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
+    GGML_ASSERT(sched);
     sched->callback_eval = callback;
     sched->callback_eval_user_data = user_data;
 }
 
 int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     return sched->n_splits;
 }
 
 int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     return sched->n_copies;
 }
 
 int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    GGML_ASSERT(sched);
     return sched->n_backends;
 }
 
 ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(sched);
     GGML_ASSERT(i >= 0 && i < sched->n_backends);
     return sched->backends[i];
 }
 
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    GGML_ASSERT(sched);
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
 
@@ -1632,6 +1802,7 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
 }
 
 void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    GGML_ASSERT(sched);
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     tensor_backend_id(node) = backend_index;
@@ -1640,6 +1811,7 @@ void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct gg
 }
 
 ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    GGML_ASSERT(sched);
     int backend_index = tensor_backend_id(node);
     if (backend_index == -1) {
         return NULL;
@@ -1650,6 +1822,7 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
 // utils
 
 enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->buffer == NULL);
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
@@ -1661,6 +1834,7 @@ enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor) {
 }
 
 enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
+    GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->buffer == NULL);
     GGML_ASSERT(tensor->data == NULL);
     GGML_ASSERT(tensor->view_src == NULL);
@@ -1734,6 +1908,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_
 }
 
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
+    GGML_ASSERT(graph);
     struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
     struct ggml_tensor ** node_copies = (ggml_tensor **) calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
     bool * node_init = (bool *) calloc(hash_set.size, sizeof(node_init[0]));
@@ -1878,6 +2053,7 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 // CPU backend - buffer
 
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     uintptr_t data = (uintptr_t)buffer->context;
 
     // align the buffer
@@ -1889,28 +2065,33 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
 }
 
 static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    GGML_ASSERT(buffer);
     ggml_aligned_free(buffer->context, buffer->size);
 }
 
 static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     memset((char *)tensor->data + offset, value, size);
 
     GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     memcpy((char *)tensor->data + offset, data, size);
 
     GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     memcpy(data, (const char *)tensor->data + offset, size);
 
     GGML_UNUSED(buffer);
 }
 
 static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(src);
     if (ggml_backend_buffer_is_host(src->buffer)) {
         memcpy(dst->data, src->data, ggml_nbytes(src));
         return true;
@@ -1921,6 +2102,7 @@ static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
 }
 
 static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    GGML_ASSERT(buffer);
     memset(buffer->context, value, buffer->size);
 }
 

ggml/src/ggml-blas/ggml-blas.cpp

@@ -270,6 +270,7 @@ static struct ggml_backend_i blas_backend_i = {
     /* .graph_compute           = */ ggml_backend_blas_graph_compute,
     /* .event_record            = */ NULL,
     /* .event_wait              = */ NULL,
+    /* .optimize_graph          = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_blas_guid(void) {

ggml/src/ggml-cann/common.h

@@ -38,6 +38,7 @@
 #include <unistd.h>
 #include <functional>
 #include <optional>
+#include <list>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -106,6 +107,7 @@ int32_t ggml_cann_get_device();
 
 std::optional<std::string> get_env(const std::string& name);
 bool parse_bool(const std::string& value);
+int parse_integer(const std::string& value);
 
 /**
  * @brief Abstract base class for memory pools used by CANN.
@@ -350,7 +352,7 @@ struct ggml_graph_node_properties {
 struct ggml_cann_graph {
     ~ggml_cann_graph() {
         if (graph != nullptr) {
-            aclmdlRIDestroy(graph);
+            ACL_CHECK(aclmdlRIDestroy(graph));
         }
     }
 
@@ -358,8 +360,105 @@ struct ggml_cann_graph {
 
     std::vector<ggml_graph_node_properties> ggml_graph_properties;
 };
+
+/**
+ * @brief LRU cache for managing ggml_cann_graph objects.
+ *
+ * This class maintains a list of shared_ptr to ggml_cann_graph objects
+ * and enforces a maximum capacity. It provides methods to push new graphs,
+ * move existing graphs to the front (most recently used), and clear the cache.
+ */
+struct ggml_cann_graph_lru_cache {
+    size_t capacity;  /**< Maximum number of graphs in the cache. */
+
+    std::list<ggml_cann_graph*> cache_list; /**< List storing cached graphs as raw pointers. */
+
+    ggml_cann_graph_lru_cache() {
+        capacity = parse_integer(get_env("GGML_CANN_GRAPH_CACHE_CAPACITY").value_or("12"));
+    }
+
+    /**
+     * @brief Push a new graph to the front of the cache.
+     * If the cache exceeds capacity, the least recently used graph is deleted.
+     * @param new_node Pointer to the new ggml_cann_graph to cache.
+     *        Ownership is transferred to the cache (cache will delete it).
+     */
+    void push(ggml_cann_graph* new_node) {
+        if (cache_list.size() >= capacity) {
+            ggml_cann_graph* old = cache_list.back();
+            cache_list.pop_back();
+            delete old; // free the old graph
+        }
+        cache_list.push_front(new_node);
+    }
+
+    /**
+     * @brief Move an existing graph to the front of the cache.
+     * @param node Pointer to the ggml_cann_graph to move.
+     */
+    void move_to_front(ggml_cann_graph* node) {
+        cache_list.remove(node);
+        cache_list.push_front(node);
+    }
+
+    /**
+     * @brief Clear all graphs from the cache (also frees memory).
+     */
+    void clear() {
+        for (auto ptr : cache_list) {
+            delete ptr;
+        }
+        cache_list.clear();
+    }
+
+    /**
+     * @brief Destructor that clears the cache and frees all cached graphs.
+     */
+    ~ggml_cann_graph_lru_cache() {
+        clear();
+    }
+};
 #endif  // USE_ACL_GRAPH
 
+struct ggml_cann_rope_cache {
+    ~ggml_cann_rope_cache() {
+        if(theta_scale_cache != nullptr) {
+            ACL_CHECK(aclrtFree(theta_scale_cache));
+        }
+        if(sin_cache != nullptr) {
+            ACL_CHECK(aclrtFree(sin_cache));
+        }
+        if(cos_cache != nullptr) {
+            ACL_CHECK(aclrtFree(cos_cache));
+        }
+    }
+
+    void* theta_scale_cache = nullptr;
+    int64_t theta_scale_length = 0;
+    // sin/cos cache, used only to accelerate first layer on each device
+    void* sin_cache = nullptr;
+    void* cos_cache = nullptr;
+    int64_t position_length = 0;
+    // Properties to check before reusing the sincos cache
+    bool cached = false;
+    float ext_factor = 0.0f;
+    float theta_scale = 0.0f;
+    float freq_scale = 0.0f;
+    float attn_factor = 0.0f;
+    bool is_neox = false;
+};
+
+struct ggml_cann_tensor_cache {
+    ~ggml_cann_tensor_cache() {
+        if(cache != nullptr) {
+            ACL_CHECK(aclrtFree(cache));
+        }
+    }
+
+    void* cache = nullptr;
+    int64_t size = 0;
+};
+
 /**
  * @brief Context for managing CANN backend operations.
  */
@@ -368,17 +467,18 @@ struct ggml_backend_cann_context {
     std::string name;                /**< Name of the device. */
     std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
-    void* init_ptr = nullptr;
-    void* sin_ptr = nullptr;
-    void* cos_ptr = nullptr;
-    int64_t max_prompt_length = 65536;
 #ifdef USE_ACL_GRAPH
     /// Cached CANN ACL graph used for executing the current ggml computation graph.
-    std::unique_ptr<ggml_cann_graph> cann_graph;
+    ggml_cann_graph_lru_cache graph_lru_cache;
+    bool acl_graph_mode = true;
 #endif
     cann_task_queue task_queue;
     bool async_mode;
-    bool support_set_rows;
+    // Rope Cache
+    ggml_cann_rope_cache rope_cache;
+    // Constant Pool
+    ggml_cann_tensor_cache rms_norm_one_tensor_cache;
+    ggml_cann_tensor_cache rms_norm_zero_tensor_cache;
 
     aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
@@ -394,14 +494,13 @@ struct ggml_backend_cann_context {
         async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
-
-        support_set_rows = parse_bool(get_env("LLAMA_SET_ROWS").value_or(""));
-        GGML_LOG_INFO("%s: LLAMA_SET_ROWS is %s\n", __func__, support_set_rows ? "ON" : "OFF");
-
-        if (!support_set_rows) {
-            GGML_LOG_INFO("%s: CANN Graph currently only supports execution when LLAMA_SET_ROWS is ON. "
-                    "Falling back to eager mode.\n", __func__);
-        }
+#ifdef USE_ACL_GRAPH
+        acl_graph_mode = parse_bool(get_env("GGML_CANN_ACL_GRAPH").value_or("on"));
+        GGML_LOG_INFO("%s: device %d execution mode is %s (%s)\n",
+              __func__, device,
+              acl_graph_mode ? "GRAPH" : "EAGER",
+              acl_graph_mode ? "acl graph enabled" : "acl graph disabled");
+#endif
     }
 
     /**
@@ -418,15 +517,6 @@ struct ggml_backend_cann_context {
                 ACL_CHECK(aclrtDestroyStream(streams[i]));
             }
         }
-        if(init_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(init_ptr));
-        }
-        if(sin_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(sin_ptr));
-        }
-        if(cos_ptr != nullptr) {
-            ACL_CHECK(aclrtFree(cos_ptr));
-        }
     }
 
     /**
@@ -436,7 +526,10 @@ struct ggml_backend_cann_context {
      */
     aclrtStream stream(int stream) {
         if (streams[stream] == nullptr) {
-            ggml_cann_set_device(device);
+            // If the device is not set here, destroying the stream later may cause a mismatch
+            // between the thread contexts where the stream was created and destroyed.
+            // However, I printed the device_id, thread_id, and stream, and they are all consistent.
+            ACL_CHECK(aclrtSetDevice(device));
             ACL_CHECK(aclrtCreateStream(&streams[stream]));
         }
         return streams[stream];

ggml/src/ggml-cann/ggml-cann.cpp

@@ -75,13 +75,12 @@
  * @param device The device ID to set.
  */
 void ggml_cann_set_device(const int32_t device) {
-    // TODO: uncomment these lines after empty context has fixed.
-    // int current_device;
-    // ACL_CHECK(aclrtGetDevice(&current_device));
+    int current_device = -1;
+    aclrtGetDevice(&current_device);
 
-    // if (device == current_device) {
-    //   return;
-    // }
+    if (device == current_device) {
+      return;
+    }
     ACL_CHECK(aclrtSetDevice(device));
 }
 
@@ -116,6 +115,24 @@ bool parse_bool(const std::string& value) {
     return valid_values.find(value) != valid_values.end();
 }
 
+/**
+ * @brief Parse a string as an integer, returning 0 if invalid.
+ *
+ * This function attempts to convert the input string `value` to an `int`.
+ * If the string is not a valid integer or is out of the `int` range,
+ * it returns 0.
+ *
+ * @param value The string to parse.
+ * @return The parsed integer, or 0 if conversion fails.
+ */
+int parse_integer(const std::string& value) {
+    try {
+        return std::stoi(value);
+    } catch (...) {
+        return 0;
+    }
+}
+
 /**
  * @brief Initialize the CANN device information.
  *
@@ -1116,30 +1133,65 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
-// ND to NZ Workspace Cache Management. Thread-safety: Not guaranteed
-namespace {
-    void* g_nz_workspace = nullptr;
-    size_t g_nz_workspace_allocated = 0;
+/**
+ * @brief Workspace for caching NZ buffers per device.
+ *
+ * This struct manages a device buffer used in NZ computations. It supports
+ * allocation, reallocation, and clearing of cached memory. The struct is
+ * designed to be used with a global array, one per device.
+ */
+struct ggml_cann_nz_workspace {
+    void*  ptr;       // Pointer to allocated device buffer
+    size_t allocated; // Size of currently allocated buffer in bytes
 
-    void release_nz_workspace() {
-        if (g_nz_workspace) {
-            aclrtFree(g_nz_workspace);
-            g_nz_workspace = nullptr;
-            g_nz_workspace_allocated = 0;
+    /**
+     * @brief Constructor. Initializes the workspace with no allocated memory.
+     */
+    ggml_cann_nz_workspace() : ptr(nullptr), allocated(0) {}
+
+    /**
+     * @brief Free cached memory and reset the workspace.
+     *
+     * If a buffer has been allocated, this function releases it using
+     * aclrtFree and resets internal state.
+     */
+    void clear() {
+        if (ptr) {
+            ACL_CHECK(aclrtFree(ptr));
+            ptr = nullptr;
+            allocated = 0;
         }
     }
 
-    void relloc_nz_workspace(size_t new_size) {
-        if (new_size > g_nz_workspace_allocated) {
-        if (g_nz_workspace) {
-            aclrtFree(g_nz_workspace);
-            g_nz_workspace = nullptr;
+    /**
+     * @brief Allocate or reallocate the workspace buffer.
+     *
+     * If the requested size is larger than the currently allocated size,
+     * the old buffer will be freed and a new buffer of the requested size
+     * will be allocated on the device.
+     *
+     * @param new_size Size in bytes to allocate for the workspace.
+     */
+    void realloc(size_t new_size) {
+        if (new_size > allocated) {
+            clear();
+            ACL_CHECK(aclrtMalloc(&ptr, new_size, ACL_MEM_MALLOC_HUGE_FIRST));
+            allocated = new_size;
         }
-        ACL_CHECK(aclrtMalloc(&g_nz_workspace, new_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        g_nz_workspace_allocated = new_size;
     }
-    }
-}
+
+    /**
+     * @brief Get the device buffer pointer.
+     *
+     * @return Pointer to the allocated buffer, or nullptr if not allocated.
+     */
+    void* get() const { return ptr; }
+};
+
+/**
+ * @brief Global array of NZ workspaces, one per device.
+ */
+static ggml_cann_nz_workspace g_nz_workspaces[GGML_CANN_MAX_DEVICES];
 
 /**
  * @brief Convert tensor weights to NZ format using Ascend CANN API.
@@ -1149,13 +1201,13 @@ namespace {
  * improve performance on certain hardware.
  *
  * @param tensor Pointer to the input ggml_tensor containing the weights.
- * @param data Pointer to the raw data buffer for the tensor weights.
  * @param offset Byte offset within the tensor data buffer where weights start.
+ * @param device device id.
  *
  * @note The workspace buffer used in this function is managed globally and reused
  *       across calls. This reduces overhead from repeated memory allocation and deallocation.
  */
-static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
+static void weight_format_to_nz(ggml_tensor *tensor, size_t offset, int device) {
     aclTensor* weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne,
                                     tensor->nb, 2, ACL_FORMAT_ND, offset);
     uint64_t workspaceSize = 0;
@@ -1165,7 +1217,9 @@ static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t of
     ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed,
                                                     &workspaceSize, &executor));
     // Avoid frequent malloc/free of the workspace.
-    relloc_nz_workspace(workspaceSize);
+    g_nz_workspaces[device].realloc(workspaceSize);
+
+    void* g_nz_workspace = g_nz_workspaces[device].get();
 
     ACL_CHECK(aclnnTransMatmulWeight(g_nz_workspace, workspaceSize, executor, nullptr));
     ACL_CHECK(aclDestroyTensor(weightTransposed));
@@ -1196,14 +1250,14 @@ static void ggml_backend_cann_buffer_set_tensor(
     // Why aclrtSynchronizeDevice?
 
     // Only check env once.
-    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
         if (weight_to_nz && is_matmul_weight((const ggml_tensor*)tensor)) {
             GGML_ASSERT(tensor->ne[2] == 1);
             GGML_ASSERT(tensor->ne[3] == 1);
-            weight_format_to_nz(tensor, data, offset);
+            weight_format_to_nz(tensor, offset, ctx->device);
         }
     } else {
         void *transform_buffer = malloc(size);
@@ -1279,6 +1333,10 @@ static bool ggml_backend_cann_buffer_cpy_tensor(
                                   ACL_MEMCPY_DEVICE_TO_DEVICE));
             return true;
         } else {
+#ifdef ASCEND_310P
+            // TODO: Support 310p P2P copy
+            return false;
+#endif
             // Different device but can access by peer.
             int32_t canAccessPeer = 0;
             ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device,
@@ -1439,7 +1497,7 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(
     int64_t ne0 = tensor->ne[0];
 
     // Only check env once.
-    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or(""));
+    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
 
     // last line must bigger than 32, because every single op deal at
     // least 32 bytes.
@@ -1670,6 +1728,7 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
             ggml_cann_get_rows(ctx, dst);
             break;
         case GGML_OP_SET_ROWS:
+            std::cout << "lcg GGML_OP_SET_ROWS"<< std::endl;
             ggml_cann_set_rows(ctx, dst);
             break;
         case GGML_OP_DUP:
@@ -2000,6 +2059,8 @@ static bool ggml_backend_cann_cpy_tensor_async(
     GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
                 ggml_backend_is_cann(backend_dst));
 
+    GGML_ASSERT(!is_matmul_weight((const ggml_tensor*)src));
+
     if (!ggml_backend_buffer_is_cann(src->buffer) ||
         !ggml_backend_buffer_is_cann(dst->buffer)) {
         return false;
@@ -2020,6 +2081,10 @@ static bool ggml_backend_cann_cpy_tensor_async(
         return true;
     }
     if (backend_src != backend_dst) {
+#ifdef ASCEND_310P
+        // TODO: Support 310p P2P copy
+        return false;
+#endif
         ggml_backend_cann_buffer_context* buf_ctx_src =
             (ggml_backend_cann_buffer_context*)buf_src->context;
         ggml_backend_cann_buffer_context* buf_ctx_dst =
@@ -2036,7 +2101,6 @@ static bool ggml_backend_cann_cpy_tensor_async(
         }
 
         // need open both directions for memcpyasync between devices.
-        ggml_cann_set_device(cann_ctx_dst->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
         ggml_cann_set_device(cann_ctx_src->device);
         ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
@@ -2046,9 +2110,17 @@ static bool ggml_backend_cann_cpy_tensor_async(
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
                                    ACL_MEMCPY_DEVICE_TO_DEVICE,
                                    cann_ctx_src->stream()));
+        // record event on src stream after the copy
+        // TODO: this event is not effective with acl graph mode, change to use aclrtSynchronizeStream
+        // if (!cann_ctx_src->copy_event) {
+        //     ACL_CHECK(aclrtCreateEventWithFlag(&cann_ctx_src->copy_event, ACL_EVENT_SYNC));
+        // }
+        // ACL_CHECK(aclrtRecordEvent(cann_ctx_src->copy_event, cann_ctx_src->stream()));
 
-        //TODO: workaround for Event didn`t work here.
-        aclrtSynchronizeStream(cann_ctx_src->stream());
+        // // wait on dst stream for the copy to complete
+        // ggml_cann_set_device(cann_ctx_dst->device);
+        // ACL_CHECK(aclrtStreamWaitEvent(cann_ctx_dst->stream(), cann_ctx_src->copy_event));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx_src->stream()));
     } else {
         // src and dst are on the same backend
         ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
@@ -2077,30 +2149,52 @@ static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
 
 #ifdef USE_ACL_GRAPH
 /**
- * @brief Populate the internal CANN graph node properties from the ggml computation graph.
+ * @brief Add a new CANN graph to the LRU cache by populating node properties from the ggml graph.
  *
- * This function copies all node attributes (operation type, dimensions, strides, input sources,
- * and operation parameters) into the cached CANN graph structure for later reuse or comparison.
+ * This function creates a new ggml_cann_graph object and fills its node properties
+ * (operation type, dimensions, strides, input sources, and operation parameters)
+ * based on the current ggml computation graph.
  *
- * @param cann_ctx  The CANN backend context.
- * @param cgraph    The ggml computational graph.
+ * Each node in the ggml graph is mapped to a property entry in the new CANN graph:
+ * - node address
+ * - operation type
+ * - shape (ne) and strides (nb)
+ * - source tensor addresses
+ * - operation parameters
+ *
+ * After initialization, the new graph is pushed into the LRU cache owned by the
+ * CANN backend context. The cache takes ownership of the graph and manages its
+ * lifetime (including deletion upon eviction).
+ *
+ * @param cann_ctx  The CANN backend context containing the graph cache.
+ * @param cgraph    The current ggml computation graph.
  */
-static void set_ggml_graph_node_properties(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    for (int node_idx = 0; node_idx < cgraph->n_nodes; node_idx++) {
-        ggml_tensor * node = cgraph->nodes[node_idx];
-        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_address = node->data;
-        cann_ctx->cann_graph->ggml_graph_properties[node_idx].node_op = node->op;
+static void add_lru_matched_graph_node_properties(
+        ggml_backend_cann_context * cann_ctx,
+        ggml_cgraph * cgraph) {
+    // Create a new ggml_cann_graph object on the heap (its lifetime is managed by the cache).
+    ggml_cann_graph * new_graph = new ggml_cann_graph();
+    new_graph->ggml_graph_properties.resize(cgraph->n_nodes);
 
-        for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].ne[dim] = node->ne[dim];
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].nb[dim] = node->nb[dim];
+    for (int node_idx = 0; node_idx < cgraph->n_nodes; ++node_idx) {
+        ggml_tensor * node = cgraph->nodes[node_idx];
+        auto & prop = new_graph->ggml_graph_properties[node_idx];
+
+        prop.node_address = node->data;
+        prop.node_op      = node->op;
+
+        std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
+        std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
+
+        for (int src = 0; src < GGML_MAX_SRC; ++src) {
+            prop.src_address[src] = node->src[src] ? node->src[src]->data : nullptr;
         }
-        for (int src = 0; src < GGML_MAX_SRC; src++) {
-            cann_ctx->cann_graph->ggml_graph_properties[node_idx].src_address[src] =
-                node->src[src] ? node->src[src]->data : nullptr;
-        }
-        memcpy(cann_ctx->cann_graph->ggml_graph_properties[node_idx].op_params, node->op_params, GGML_MAX_OP_PARAMS);
+
+        memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
     }
+
+    // Insert into the LRU cache (cache takes ownership and will delete it when evicted).
+    cann_ctx->graph_lru_cache.push(new_graph);
 }
 
 /**
@@ -2145,30 +2239,45 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
 }
 
 /**
- * @brief Determine if the CANN graph needs to be rebuilt due to graph changes.
+ * @brief Check whether there is a cached CANN graph that matches the current ggml graph.
  *
- * This checks whether the number or properties of ggml graph nodes have changed
- * compared to the last captured CANN graph. If so, the CANN graph must be re-captured.
+ * This function iterates through the cached CANN graphs stored in the LRU cache and
+ * compares them against the given ggml computation graph. A match requires that the
+ * number of nodes is the same and that each node’s properties (operation type,
+ * dimensions, strides, inputs, and operation parameters) are identical.
  *
- * @param cann_ctx  The CANN backend context.
+ * If a matching graph is found, it is promoted to the front of the LRU cache and the
+ * function returns true. Otherwise, the function returns false, indicating that a new
+ * CANN graph needs to be captured.
+ *
+ * @param cann_ctx  The CANN backend context containing the graph cache.
  * @param cgraph    The current ggml computation graph.
- * @return true if an update is required; false otherwise.
+ * @return true if a matching cached graph exists; false otherwise.
  */
-static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
-    // The number of nodes is different, so the graph needs to be reconstructed.
-    if (cann_ctx->cann_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
-        cann_ctx->cann_graph->ggml_graph_properties.resize(cgraph->n_nodes);
-        return true;
-    }
+static bool is_matched_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph) {
+    ggml_cann_graph_lru_cache &lru_cache = cann_ctx->graph_lru_cache;
+    for (auto &graph_ptr : lru_cache.cache_list) {
+        // Skip graphs with a different number of nodes.
+        if (graph_ptr->ggml_graph_properties.size() != static_cast<size_t>(cgraph->n_nodes)) {
+            continue;
+        }
 
-    // The number of nodes is the same; iterate over each node to check whether they match.
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        bool has_matching_properties = ggml_graph_node_has_matching_properties(
-            cgraph->nodes[i], &cann_ctx->cann_graph->ggml_graph_properties[i]);
-        if(!has_matching_properties) {
+        // Check if all nodes match.
+        bool all_match = true;
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            if (!ggml_graph_node_has_matching_properties(cgraph->nodes[i], &graph_ptr->ggml_graph_properties[i])) {
+                all_match = false;
+                break;
+            }
+        }
+
+        if (all_match) {
+            // update cache_list && renturn graph_ptr
+            lru_cache.move_to_front(graph_ptr);
             return true;
         }
     }
+
     return false;
 }
 #endif  // USE_ACL_GRAPH
@@ -2187,17 +2296,13 @@ static bool is_cann_graph_update_required(ggml_backend_cann_context * cann_ctx,
  * @param cann_graph_update_required Whether graph capture is needed due to graph changes.
  */
 static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx, ggml_cgraph * cgraph,
-    bool & use_cann_graph, bool & cann_graph_update_required) {
+    bool & use_cann_graph,  bool & cann_graph_update_required) {
 #ifdef USE_ACL_GRAPH
+    ggml_cann_graph* matched_graph = cann_ctx->graph_lru_cache.cache_list.front();
     if (use_cann_graph && cann_graph_update_required) {
-        if (cann_ctx->cann_graph->graph != nullptr) {
-            ACL_CHECK(aclmdlRIDestroy(cann_ctx->cann_graph->graph));
-            cann_ctx->cann_graph->graph = nullptr;
-        }
         ACL_CHECK(aclmdlRICaptureBegin(cann_ctx->stream(), ACL_MODEL_RI_CAPTURE_MODE_GLOBAL));
     }
 #endif // USE_ACL_GRAPH
-
     // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
     // With the use of CANN graphs, the execution will be performed by the graph launch.
     if (!use_cann_graph || cann_graph_update_required) {
@@ -2218,12 +2323,12 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
 
 #ifdef USE_ACL_GRAPH
     if (use_cann_graph && cann_graph_update_required) { // End CANN graph capture
-        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &cann_ctx->cann_graph->graph));
+        ACL_CHECK(aclmdlRICaptureEnd(cann_ctx->stream(), &matched_graph->graph));
     }
 
     if (use_cann_graph) {
         // Execute graph
-        ACL_CHECK(aclmdlRIExecuteAsync(cann_ctx->cann_graph->graph, cann_ctx->stream()));
+        ACL_CHECK(aclmdlRIExecuteAsync(matched_graph->graph, cann_ctx->stream()));
     }
 #endif // USE_ACL_GRAPH
 }
@@ -2246,30 +2351,46 @@ static enum ggml_status ggml_backend_cann_graph_compute(
     ggml_backend_cann_context* cann_ctx =
         (ggml_backend_cann_context*)backend->context;
     ggml_cann_set_device(cann_ctx->device);
-    release_nz_workspace();
+    g_nz_workspaces[cann_ctx->device].clear();
+
+    // calculate rope cache for fist layer in current device.
+    cann_ctx->rope_cache.cached = false;
+
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph = true;
     bool cann_graph_update_required = false;
 
-    // check environment LLAMA_SET_ROWS
-    if (!cann_ctx->support_set_rows) {
+    static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
+    if (!prefill_use_graph) {
+        // Do not use acl_graph for prefill.
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+            // TODO: Optimize here. Currently, we can only
+            // get seq_len by FA's input.
+            if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+                // Q -> src[0], shape: [B, S, N, D]
+                use_cann_graph = (node->src[0]->ne[1] == 1);
+                break;
+            }
+        }
+    }
+
+    if (!cann_ctx->acl_graph_mode) {
         use_cann_graph = false;
     }
 
     if (use_cann_graph) {
-        if (cann_ctx->cann_graph == nullptr) {
-            cann_ctx->cann_graph.reset(new ggml_cann_graph());
-            cann_graph_update_required = true;
+        // If no matching graph is found, the graph needs to be recaptured.
+        cann_graph_update_required = !is_matched_graph(cann_ctx, cgraph);
+        if (cann_graph_update_required) {
+            // If no matching graph is found, add a new ACL graph.
+            add_lru_matched_graph_node_properties(cann_ctx, cgraph);
         }
-
-        cann_graph_update_required = is_cann_graph_update_required(cann_ctx, cgraph);
-        set_ggml_graph_node_properties(cann_ctx, cgraph);
     }
 #else
     bool use_cann_graph = false;
     bool cann_graph_update_required = false;
 #endif  // USE_ACL_GRAPH
-
     evaluate_and_capture_cann_graph(
         cann_ctx,
         cgraph,
@@ -2336,7 +2457,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2354,7 +2475,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
 #ifdef ASCEND_310P
-                    // Q4 && Q8 per group is not suppor on 310p device
+                    // Q4 && Q8 per group is not support on 310p device
                     return false;
 #endif
                     // only support contiguous for quantized types.
@@ -2405,16 +2526,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         }
         case GGML_OP_ROPE: {
             // TODO: with ops-test v == 1
-            float ext_factor = 0.0f;
-            memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
             // TODO: n_dims <= ne0
             if (op->src[0]->ne[0] != op->op_params[1]) {
                 return false;
             }
-            // TODO: ext_factor != 0
-            if (ext_factor != 0) {
-                return false;
-            }
 
             const int mode = ((const int32_t *) op->op_params)[2];
             if (mode & GGML_ROPE_TYPE_MROPE) {
@@ -2423,10 +2538,11 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (mode & GGML_ROPE_TYPE_VISION) {
                 return false;
             }
-
+#ifdef ASCEND_310P
             if(!ggml_is_contiguous(op->src[0])){
                 return false;
             }
+#endif
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -2488,15 +2604,17 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_ARGMAX:
         case GGML_OP_COS:
         case GGML_OP_SIN:
-        case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_LOG:
         case GGML_OP_MEAN:
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
+            return (op->src[0]->ne[0] - 1) <= 255;
         case GGML_OP_SCALE:
             float bias;
-            memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
+            memcpy(&bias, (const float *)(op->op_params) + 1, sizeof(float));
             return bias == 0.0f; // TODO: support bias != 0.0f
         case GGML_OP_SOFT_MAX:
             // TODO: support attention sinks [TAG_ATTN_SINKS]
@@ -2505,6 +2623,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
             return true;
         case GGML_OP_FLASH_ATTN_EXT:{
+#ifdef ASCEND_310P
+            // FA not support on 310p device
+            return false;
+#endif
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
                 return false;
@@ -2523,15 +2645,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // different head sizes of K and V are not supported yet
                 return false;
             }
-            if (op->src[0]->ne[0] == 192) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek MLA
+            if (op->src[0]->ne[0] % 16 != 0) {
+                // TODO: padding to support
                 return false;
             }
             float logitSoftcap = 0.0f;
-            memcpy(&logitSoftcap,  (float*)op->op_params + 2, sizeof(float));
+            memcpy(&logitSoftcap, (const float *)(op->op_params) + 2, sizeof(float));
             if(logitSoftcap != 0.0f) {
                 return false;
             }
@@ -2638,6 +2757,7 @@ static const ggml_backend_i ggml_backend_cann_interface = {
     /* .graph_compute           = */ ggml_backend_cann_graph_compute,
     /* .event_record            = */ ggml_backend_cann_event_record,
     /* .event_wait              = */ ggml_backend_cann_event_wait,
+    /* .optimize_graph          = */ NULL,
 };
 
 /**

ggml/src/ggml-cpu/CMakeLists.txt

@@ -224,7 +224,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME)
                     string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
                     if (NOT ${feature_pos} EQUAL -1)
-                        message(STATUS "ARM feature ${feature} enabled")
+                        # Special handling for MATMUL_INT8 when machine doesn't support i8mm
+                        if ("${feature}" STREQUAL "MATMUL_INT8" AND GGML_MACHINE_SUPPORTS_noi8mm)
+                            message(STATUS "ARM feature ${feature} detected but unsetting due to machine not supporting i8mm")
+                            list(APPEND ARCH_FLAGS -U__ARM_FEATURE_MATMUL_INT8)
+                        else()
+                            message(STATUS "ARM feature ${feature} enabled")
+                        endif()
                     endif()
                 endforeach()
             endif()
@@ -433,15 +439,22 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ggml-cpu/arch/riscv/quants.c
             ggml-cpu/arch/riscv/repack.cpp
             )
-        if (GGML_RVV)
-            if (GGML_XTHEADVECTOR)
-                list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
-            elseif (GGML_RV_ZFH)
-                list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -mabi=lp64d)
-            else()
-                list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+        set(MARCH_STR "rv64gc")
+        if (GGML_RV_ZFH)
+            string(APPEND MARCH_STR "_zfh")
+        endif()
+        if (GGML_XTHEADVECTOR)
+            string(APPEND MARCH_STR "_xtheadvector")
+        elseif (GGML_RVV)
+            string(APPEND MARCH_STR "_v")
+            if (GGML_RV_ZVFH)
+                string(APPEND MARCH_STR "_zvfh")
             endif()
         endif()
+        if (GGML_RV_ZICBOP)
+            string(APPEND MARCH_STR "_zicbop")
+        endif()
+        list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
         list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
@@ -450,7 +463,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
-            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -472,11 +484,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
             list(APPEND ARCH_DEFINITIONS GGML_VXE)
         endif()
-
-        if (GGML_NNPA)
-            message(STATUS "NNPA enabled")
-            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
-        endif()
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
@@ -497,9 +504,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.11.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.13.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "3fe9e5ab964c375c53839296eb71eaa2")
+        set(KLEIDIAI_ARCHIVE_MD5  "d82a8de939d9814621a5ba23907bdac1")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -555,6 +562,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         list(APPEND GGML_KLEIDIAI_SOURCES
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
@@ -576,7 +584,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
-                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
+                ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
             set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 

ggml/src/ggml-cpu/arch-fallback.h

@@ -150,8 +150,6 @@
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
-#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K

ggml/src/ggml-cpu/arch/riscv/quants.c

@@ -1270,29 +1270,40 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
             const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
-            int tmp, tmp2, sumi;
+            float ftmp, ft2;
+            const uint8_t * restrict q40;
+            const uint8_t * restrict q41;
+            const uint8_t * restrict q42;
+            const uint8_t * restrict q43;
+            const int8_t  * restrict q80;
+            const int8_t  * restrict q81;
+            const int8_t  * restrict q82;
+            const int8_t  * restrict q83;
+            int s0, s1, s2, s3;
+
             __asm__ __volatile__(
-                "vsetivli zero, 12, e8, m1\n\t"
-                "vle8.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
-                "vsetivli zero, 4, e32, m1\n\t"
+                "li %[s1], 8\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vle32.v v1, (%[s6b])\n\t"
+                "vslide1down.vx v1, v1, zero\n\t"
+                "vmv.v.x v16, zero\n\t"
                 "vslidedown.vi v2, v1, 2\n\t"
                 "vmv1r.v v3, v2\n\t"
                 "vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
-                "vsetivli zero, 2, e32, m1\n\t"
+                "vsetivli zero, 2, e32, m1, ta, ma\n\t"
                 "vmv.v.i v4, 4\n\t"
                 "vand.vx v8, v1, %[kmask1]\n\t"
                 "vslide1up.vx v5, v4, zero\n\t" // {0, 4}
                 "vsrl.vi v6, v1, 6\n\t"
                 "vsrl.vv v7, v2, v5\n\t"
+                "vsse32.v v8, (%[utmp]), %[s1]\n\t"
                 "vand.vx v0, v6, %[kmask3]\n\t"
                 "vand.vx v2, v7, %[kmask2]\n\t"
                 "vsll.vi v6, v0, 4\n\t"
-                "li %[t2], 8\n\t"
-                "addi %[t1], %[utmp], 4\n\t"
+                "addi %[s0], %[utmp], 4\n\t"
                 "vor.vv v1, v6, v2\n\t"
-                "vsse32.v v8, (%[utmp]), %[t2]\n\t"
-                "vsse32.v v1, (%[t1]), %[t2]\n\t"
-                "vsetivli zero, 8, e16, m1\n\t"
+                "vsse32.v v1, (%[s0]), %[s1]\n\t"
+                "vsetivli zero, 8, e16, m1, ta, ma\n\t"
                 "vle32.v v2, (%[bsums])\n\t"
                 "vnsrl.wi v0, v2, 0\n\t"
                 "vnsrl.wi v1, v2, 16\n\t"
@@ -1300,13 +1311,131 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 "vle8.v v3, (%[mins])\n\t"
                 "vzext.vf2 v4, v3\n\t"
                 "vwmul.vv v6, v4, v2\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vredsum.vs v0, v6, v16\n\t"
+                "vredsum.vs v0, v7, v0\n\t"
+                "vfcvt.f.x.v v0, v0\n\t"
+                "vfmv.f.s %[ftmp], v0\n\t"
+                "vsetivli zero, 16, e8, m1, ta, ma\n\t"
+                "vle8.v v0, (%[xs])\n\t"
+                "fnmsub.s %[sumf], %[dmin], %[ftmp], %[sumf]\n\t"
+                "addi %[q40], %[xs], 64\n\t"
+                "addi %[q41], %[xs], 16\n\t"
+                "addi %[q42], %[xs], 32\n\t"
+                "addi %[q43], %[xs], 48\n\t"
+                "addi %[q80], %[ys], 64\n\t"
+                "vle8.v v1, (%[q41])\n\t"
+                "vle8.v v2, (%[q42])\n\t"
+                "addi %[q81], %[ys], 16\n\t"
+                "addi %[q41], %[q41], 64\n\t"
+                "addi %[q82], %[ys], 32\n\t"
+                "vle8.v v3, (%[q43])\n\t"
+                "vle8.v v8, (%[ys])\n\t"
+                "addi %[q42], %[q42], 64\n\t"
+                "addi %[q83], %[ys], 48\n\t"
+                "addi %[q43], %[q43], 64\n\t"
+                "vsrl.vi v4, v0, 4\n\t"
+                "vle8.v v9, (%[q81])\n\t"
+                "vle8.v v10, (%[q82])\n\t"
+                "vand.vi v0, v0, 0xF\n\t"
+                "addi %[q81], %[q81], 64\n\t"
+                "vsrl.vi v5, v1, 4\n\t"
+                "addi %[q82], %[q82], 64\n\t"
+                "vle8.v v11, (%[q83])\n\t"
+                "vle8.v v12, (%[q80])\n\t"
+                "vand.vi v1, v1, 0xF\n\t"
+                "addi %[q83], %[q83], 64\n\t"
+                "vsrl.vi v6, v2, 4\n\t"
+                "addi %[q80], %[q80], 64\n\t"
+                "vle8.v v13, (%[q81])\n\t"
+                "vle8.v v14, (%[q82])\n\t"
+                "vand.vi v2, v2, 0xF\n\t"
+                "addi %[q81], %[q81], 64\n\t"
+                "vsrl.vi v7, v3, 4\n\t"
+                "addi %[q82], %[q82], 64\n\t"
+                "vwmul.vv v16, v0, v8\n\t"
+                "vle8.v v15, (%[q83])\n\t"
+                "vle8.v v0, (%[q40])\n\t"
+                "vand.vi v3, v3, 0xF\n\t"
+                "addi %[q83], %[q83], 64\n\t"
+                "vwmul.vv v24, v2, v12\n\t"
+                "vwmul.vv v20, v4, v10\n\t"
+                "vwmul.vv v28, v6, v14\n\t"
+                "vwmacc.vv v16, v1, v9\n\t"
+                "vle8.v v1, (%[q41])\n\t"
+                "vle8.v v2, (%[q42])\n\t"
+                "vwmacc.vv v24, v3, v13\n\t"
+                "vwmacc.vv v20, v5, v11\n\t"
+                "vwmacc.vv v28, v7, v15\n\t"
+                "addi %[q40], %[q80], 64\n\t"
+                "addi %[q41], %[q81], 64\n\t"
+                "vle8.v v3, (%[q43])\n\t"
+                "vle8.v v8, (%[q80])\n\t"
+                "addi %[q42], %[q82], 64\n\t"
+                "addi %[q43], %[q83], 64\n\t"
+                "vsrl.vi v4, v0, 4\n\t"
+                "vle8.v v9, (%[q81])\n\t"
+                "vle8.v v10, (%[q82])\n\t"
+                "vand.vi v0, v0, 0xF\n\t"
+                "vsrl.vi v5, v1, 4\n\t"
+                "vsrl.vi v7, v3, 4\n\t"
+                "vand.vi v3, v3, 0xF\n\t"
+                "vle8.v v11, (%[q83])\n\t"
+                "vle8.v v12, (%[q40])\n\t"
+                "vand.vi v1, v1, 0xF\n\t"
+                "vsrl.vi v6, v2, 4\n\t"
+                "vand.vi v2, v2, 0xF\n\t"
+                "vwmul.vv v18, v0, v8\n\t"
+                "vle8.v v13, (%[q41])\n\t"
+                "vle8.v v14, (%[q42])\n\t"
+                "vwmul.vv v26, v2, v12\n\t"
+                "vwmul.vv v22, v4, v10\n\t"
+                "vwmul.vv v30, v6, v14\n\t"
+                "vwmacc.vv v18, v1, v9\n\t"
+                "vle8.v v15, (%[q43])\n\t"
+                "vwmacc.vv v26, v3, v13\n\t"
+                "vwmacc.vv v22, v5, v11\n\t"
+                "vwmacc.vv v30, v7, v15\n\t"
                 "vmv.v.x v0, zero\n\t"
-                "vsetivli zero, 8, e32, m2\n\t"
-                "vredsum.vs v0, v6, v0\n\t"
-                "vmv.x.s %[sumi], v0"
-                : [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi)
-                : [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
-                , [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1)
+                "vsetivli zero, 16, e16, m2, ta, ma\n\t"
+                "vwredsum.vs v4, v16, v0\n\t"
+                "lbu %[s0], 0(%[scale])\n\t"
+                "vwredsum.vs v5, v20, v0\n\t"
+                "lbu %[s1], 1(%[scale])\n\t"
+                "vwredsum.vs v6, v24, v0\n\t"
+                "lbu %[s2], 2(%[scale])\n\t"
+                "vwredsum.vs v7, v28, v0\n\t"
+                "lbu %[s3], 3(%[scale])\n\t"
+                "vwredsum.vs v8, v18, v0\n\t"
+                "lbu %[q40], 4(%[scale])\n\t"
+                "vwredsum.vs v9, v22, v0\n\t"
+                "lbu %[q41], 5(%[scale])\n\t"
+                "vwredsum.vs v10, v26, v0\n\t"
+                "lbu %[q42], 6(%[scale])\n\t"
+                "vwredsum.vs v11, v30, v0\n\t"
+                "lbu %[q43], 7(%[scale])\n\t"
+                "vsetivli zero, 4, e32, m1, ta, ma\n\t"
+                "vmul.vx v0, v4, %[s0]\n\t"
+                "vmul.vx v1, v8, %[q40]\n\t"
+                "vmacc.vx v0, %[s1], v5\n\t"
+                "vmacc.vx v1, %[q41], v9\n\t"
+                "vmacc.vx v0, %[s2], v6\n\t"
+                "vmacc.vx v1, %[q42], v10\n\t"
+                "vmacc.vx v0, %[s3], v7\n\t"
+                "vmacc.vx v1, %[q43], v11\n\t"
+                "vfcvt.f.x.v v0, v0\n\t"
+                "vfcvt.f.x.v v1, v1\n\t"
+                "vfmv.f.s %[ft2], v0\n\t"
+                "vfmv.f.s %[ftmp], v1\n\t"
+                "fadd.s %[ft2], %[ft2], %[ftmp]\n\t"
+                "fmadd.s %[sumf], %[d], %[ft2], %[sumf]"
+                : [ftmp] "=&f" (ftmp), [sumf] "+&f" (sumf), [ft2] "=&f" (ft2)
+                , [s0] "=&r" (s0), [s1] "=&r" (s1), [s2] "=&r" (s2), [s3] "=&r" (s3)
+                , [q40] "=&r" (q40), [q41] "=&r" (q41), [q42] "=&r" (q42), [q43] "=&r" (q43)
+                , [q80] "=&r" (q80), [q81] "=&r" (q81), [q82] "=&r" (q82), [q83] "=&r" (q83)
+                : [d] "f" (d), [ys] "r" (y[i].qs), [xs] "r" (x[i].qs), [scale] "r" (scales)
+                , [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp)
+                , [s6b] "r" (&x[i]), [kmask1] "r" (kmask1), [dmin] "f" (dmin)
                 , [kmask2] "r" (kmask2), [kmask3] "r" (kmask3)
                 : "memory"
                 , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
@@ -1314,59 +1443,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
                 , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
             );
-            sumf -= dmin * sumi;
-
-            const uint8_t * restrict q4 = x[i].qs;
-            const int8_t  * restrict q8 = y[i].qs;
-
-            sumi = 0;
-            const uint8_t * scale = scales;
-
-            for (int j = 0; j < QK_K/128; ++j) {
-                int vl128 = 128, vl64 = 64, vl32 = 32;
-                __asm__ __volatile__(
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
-                    "vle8.v v8, (%[q8])\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
-                    "vle8.v v0, (%[q4])\n\t"
-                    "vsrl.vi v4, v0, 4\n\t"
-                    "vand.vi v0, v0, 0xF\n\t"
-                    "vsetvli zero, %[vl32], e8, m2\n\t"
-                    "vwmul.vv v28, v6, v14\n\t"
-                    "vwmul.vv v20, v4, v10\n\t"
-                    "vwmul.vv v24, v2, v12\n\t"
-                    "vwmul.vv v16, v0, v8\n\t"
-                    "vsetivli zero, 4, e32, m1\n\t"
-                    "vle8.v v2, (%[scale])\n\t"
-                    "vmv.v.x v0, zero\n\t"
-                    "vzext.vf4 v1, v2\n\t"
-                    "vsetvli zero, %[vl32], e16, m4\n\t"
-                    "vwredsum.vs v6, v24, v0\n\t"
-                    "vwredsum.vs v7, v28, v0\n\t"
-                    "vwredsum.vs v4, v16, v0\n\t"
-                    "vwredsum.vs v5, v20, v0\n\t"
-                    "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v6, v7, 1\n\t"
-                    "vslideup.vi v4, v5, 1\n\t"
-                    "vslideup.vi v4, v6, 2\n\t"
-                    "vmul.vv v8, v4, v1\n\t"
-                    "vredsum.vs v0, v8, v0\n\t"
-                    "vmv.x.s %[tmp], v0\n\t"
-                    "add %[sumi], %[sumi], %[tmp]"
-                    : [tmp] "=&r" (tmp), [sumi] "+&r" (sumi)
-                    : [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32)
-                    , [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale)
-                    : "memory"
-                    , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
-                    , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
-                    , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
-                    , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
-                );
-
-                q4 += 64;    q8 += 128;    scale += 4;
-            }
-
-            sumf += d * sumi;
         }
         break;
     default:
@@ -1693,6 +1769,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 128:
         for (int i = 0; i < nb; ++i) {
 
+            __builtin_prefetch(&x[i + 1].d, 0, 1);
+
             const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * restrict q6 = x[i].ql;
@@ -1701,23 +1779,59 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             const int8_t * restrict scale = x[i].scales;
 
-            int sum_t = 0;
-            int t0;
+            int q6h;
+            float ftmp;
 
             for (int j = 0; j < QK_K/128; ++j) {
                 __asm__ __volatile__(
+                    "addi %[q6h], %[q6], 32\n\t"
+                    "ld t0, 0(%[scale])\n\t"
+                    "addi %[scale], %[scale], 8\n\t"
+                    "slli t6, t0, 1 * 8\n\t"
+                    "lb zero, 0(%[q6])\n\t"
+                    "slli t5, t0, 2 * 8\n\t"
+                    "slli t4, t0, 3 * 8\n\t"
+                    "lb zero, 0(%[q6h])\n\t"
+                    "slli t3, t0, 4 * 8\n\t"
+                    "slli t2, t0, 5 * 8\n\t"
+                    "lb zero, 0(%[qh])\n\t"
+                    "lb zero, 31(%[q6h])\n\t"
+                    "slli t1, t0, 6 * 8\n\t"
+                    "srai a7, t0, 56\n\t"
                     "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "vle8.v v8, (%[q6])\n\t"
+                    "srai t6, t6, 56\n\t"
+                    "srai t5, t5, 56\n\t"
+                    "srai t4, t4, 56\n\t"
+                    "srai t3, t3, 56\n\t"
+                    "vle8.v v10, (%[q6h])\n\t"
+                    "addi %[q6], %[q6], 64\n\t"
+                    "slli t0, t0, 7 * 8\n\t"
+                    "srai t2, t2, 56\n\t"
+                    "srai t1, t1, 56\n\t"
+                    "srai t0, t0, 56\n\t"
                     "vle8.v v4, (%[qh])\n\t"
+                    "vsrl.vi v12, v8, 4\n\t"
+                    "vsrl.vi v14, v10, 4\n\t"
+                    "lb zero, 0(%[q8])\n\t"
+                    "vand.vi v8, v8, 0xF\n\t"
+                    "vand.vi v10, v10, 0xF\n\t"
+                    "lb zero, 32(%[q8])\n\t"
                     "vsll.vi v0, v4, 4\n\t"
                     "vsll.vi v2, v4, 2\n\t"
+                    "lb zero, 64(%[q8])\n\t"
                     "vsrl.vi v6, v4, 2\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
-                    "vle8.v v8, (%[q6])\n\t"
-                    "vsrl.vi v12, v8, 4\n\t"
-                    "vand.vi v8, v8, 0xF\n\t"
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vand.vx v0, v0, %[mask]\n\t"
+                    "lb zero, 96(%[q8])\n\t"
+                    "vand.vx v2, v2, %[mask]\n\t"
+                    "vand.vx v4, v4, %[mask]\n\t"
+                    "vand.vx v6, v6, %[mask]\n\t"
                     "vor.vv v8, v8, v0\n\t"
+                    "lb zero, 127(%[q8])\n\t"
+                    "vor.vv v10, v10, v2\n\t"
+                    "vor.vv v12, v12, v4\n\t"
+                    "vor.vv v14, v14, v6\n\t"
+                    "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vle8.v v0, (%[q8])\n\t"
                     "vsub.vx v8, v8, %[vl32]\n\t"
                     "vsetvli zero, %[vl64], e8, m4\n\t"
@@ -1734,34 +1848,34 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     "vwredsum.vs v13, v28, v0\n\t"
                     "vwredsum.vs v14, v30, v0\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v2, (%[scale])\n\t"
-                    "vsext.vf4 v4, v2\n\t"
-                    "vmul.vv v2, v4, v10\n\t"
-                    "vredsum.vs v0, v2, v0\n\t"
-                    "vmv.x.s %[t0], v0\n\t"
-                    "add %[sumi], %[sumi], %[t0]"
-                    : [sumi] "+&r" (sum_t), [t0] "=&r" (t0)
-                    : [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale)
+                    "vmul.vx v0, v10, t0\n\t"
+                    "vmul.vx v1, v9, t1\n\t"
+                    "vmacc.vx v0, t2, v8\n\t"
+                    "vmacc.vx v1, t3, v7\n\t"
+                    "vmacc.vx v0, t4, v11\n\t"
+                    "vmacc.vx v1, t5, v12\n\t"
+                    "vmacc.vx v0, t6, v13\n\t"
+                    "vmacc.vx v1, a7, v14\n\t"
+                    "vadd.vv v0, v0, v1\n\t"
+                    "vfcvt.f.x.v v0, v0\n\t"
+                    "vfmv.f.s %[ftmp], v0\n\t"
+                    "fmadd.s %[sumf], %[d], %[ftmp], %[sumf]"
+                    : [q6] "+&r" (q6), [q6h] "=&r" (q6h)
+                    , [scale] "+&r" (scale)
+                    , [sumf] "+&f" (sumf), [ftmp] "=&f" (ftmp)
+                    : [qh] "r" (qh), [q8] "r" (q8)
                     , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
-                    , [mask] "r" (0x30)
+                    , [mask] "r" (0x30), [d] "f" (d)
                     : "memory"
                     , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
                     , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
                     , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
                     , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31"
+                    , "t0", "t1", "t2", "t3", "t4", "t5", "t6", "a7"
+                    , "a6", "a5", "a4", "a3"
                 );
-                q6 += 64;   qh += 32;   q8 += 128;   scale += 8;
+                qh += 32;   q8 += 128;
             }
-
-            sumf += d * sum_t;
-
         }
         break;
     default:

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -23,6 +23,27 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__VXE__) || defined(__VXE2__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const __attribute__((aligned(16))) uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b ) << 4
+static const __attribute__((aligned(16))) uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+
+// permute mask for byteswapping
+static const uint8x16_t v_kperm = (const uint8x16_t){
+     7,  6,  5,  4,  3,  2, 1, 0,
+    15, 14, 13, 12, 11, 10, 9, 8
+};
+#endif
+
 void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
@@ -32,9 +53,9 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__VXE__) || defined(__VXE2__)
     for (int i = 0; i < nb; i++) {
-        __vector float srcv [8];
-        __vector float asrcv[8];
-        __vector float amaxv[8];
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
 
         for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
         for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
@@ -53,8 +74,8 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
-            const __vector float v = vec_mul(srcv[j], vec_splats(id));
-            const __vector int32_t vi = vec_signed(v);
+            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
+            const int32x4_t vi = vec_signed(v);
 
             y[i].qs[4*j + 0] = vec_extract(vi, 0);
             y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -77,9 +98,9 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__VXE__) || defined(__VXE2__)
     for (int i = 0; i < nb; i++) {
-        __vector float srcv [8];
-        __vector float asrcv[8];
-        __vector float amaxv[8];
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
 
         for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j);
         for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
@@ -97,11 +118,11 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
-        __vector int32_t acc = vec_splats(0);
+        int32x4_t acc = vec_splats(0);
 
         for (int j = 0; j < 8; j++) {
-            const __vector float v = vec_mul(srcv[j], vec_splats(id));
-            const __vector int32_t vi = vec_signed(v);
+            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
+            const int32x4_t vi = vec_signed(v);
 
             y[i].qs[4*j + 0] = vec_extract(vi, 0);
             y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -141,37 +162,36 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
 #if defined(__VXE__) || defined(__VXE2__)
-    __vector float acc = vec_splats(0.0f);
+    float32x4_t acc = vec_splats(0.0f);
 
-    const __vector uint8_t v_m = vec_splats((const uint8_t)0x0F);
-    const __vector int8_t  v_s = vec_splats( (const int8_t)0x08);
+    const uint8x16_t v_m = vec_splats((const uint8_t)0x0F);
+    const int8x16_t  v_s = vec_splats( (const int8_t)0x08);
 
     for (; ib < nb; ++ib) {
-        const __vector uint8_t v_x = vec_xl(0, x[ib].qs);
-        const __vector int8_t v_xl = (const __vector int8_t)(v_x & v_m);
-        const __vector int8_t v_xh = (const __vector int8_t)(v_x >> 4);
+        const uint8x16_t v_x = vec_xl(0, x[ib].qs);
+        const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
+        const int8x16_t v_xh = (const int8x16_t)(v_x >> 4);
 
-        const __vector int8_t v_xls = vec_sub(v_xl, v_s);
-        const __vector int8_t v_xhs = vec_sub(v_xh, v_s);
+        const int8x16_t v_xls = vec_sub(v_xl, v_s);
+        const int8x16_t v_xhs = vec_sub(v_xh, v_s);
 
-        const __vector int8_t v_yl = vec_xl(0      , y[ib].qs);
-        const __vector int8_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
+        const int8x16_t v_yl = vec_xl(0      , y[ib].qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs);
 
-        const __vector int16_t v_xylso = vec_mulo(v_xls, v_yl);
-        const __vector int16_t v_xylse = vec_mule(v_xls, v_yl);
-        const __vector int16_t v_xyhso = vec_mulo(v_xhs, v_yh);
-        const __vector int16_t v_xyhse = vec_mule(v_xhs, v_yh);
+        const int16x8_t v_xylso = vec_mulo(v_xls, v_yl);
+        const int16x8_t v_xylse = vec_mule(v_xls, v_yl);
+        const int16x8_t v_xyhso = vec_mulo(v_xhs, v_yh);
+        const int16x8_t v_xyhse = vec_mule(v_xhs, v_yh);
 
-        __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
+        int16x8_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
 
-        const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
-        const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_xy = vec_float(vec_unpackh(v_xy_));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3];
-
+    sumf = vec_hsum_f32x4(acc);
     *s = sumf;
 #else
     UNUSED(nb);
@@ -228,8 +248,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs;
-
+    sumf = vec_hsum_f32x4(acc) + summs;
     *s = sumf;
 #else
     UNUSED(nb);
@@ -241,6 +260,301 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    uint32_t qh0, qh1;
+    uint64_t tmp0[4], tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, (const uint8_t *)x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, (const uint8_t *)x1->qs);
+
+        int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_sub(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_sub(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_sub(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_sub(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+        const int8x16_t v_y1l = vec_xl(0,       (const int8_t *)y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_0/2, (const int8_t *)y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum_f32x4(v_sum0) + vec_hsum_f32x4(v_sum1);
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_0 * GGML_RESTRICT y0 = &y[ib];
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, (const uint8_t *)x0->qs);
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_sub(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_sub(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0,       (const int8_t *)y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, (const int8_t *)y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));
+
+        sumf += vec_hsum_f32x4(v_acc);
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    float32x4_t v_sum0 = vec_splats(0.0f);
+    float32x4_t v_sum1 = vec_splats(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    const uint8x16_t v_m = vec_splats((uint8_t)0x0F);
+
+    #pragma GCC unroll 4
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
+
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        int8x16_t v_qh0l = vec_xl(0, (const int8_t *)(tmp0 + 0));
+        int8x16_t v_qh0h = vec_xl(0, (const int8_t *)(tmp0 + 2));
+        int8x16_t v_qh1l = vec_xl(0, (const int8_t *)(tmp1 + 0));
+        int8x16_t v_qh1h = vec_xl(0, (const int8_t *)(tmp1 + 2));
+
+        // required for fixing the byteorder
+        v_qh0l = vec_perm(v_qh0l, v_qh0l, v_kperm);
+        v_qh0h = vec_perm(v_qh0h, v_qh0h, v_kperm);
+        v_qh1l = vec_perm(v_qh1l, v_qh1l, v_kperm);
+        v_qh1h = vec_perm(v_qh1h, v_qh1h, v_kperm);
+
+        const uint8x16_t v_x0 = vec_xl(0, x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, x1->qs);
+
+        const int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        const int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        const int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        const int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        const int8x16_t v_x0lf = vec_or(v_x0l, v_qh0l);
+        const int8x16_t v_x0hf = vec_or(v_x0h, v_qh0h);
+        const int8x16_t v_x1lf = vec_or(v_x1l, v_qh1l);
+        const int8x16_t v_x1hf = vec_or(v_x1h, v_qh1h);
+
+        const int8x16_t v_y0l = vec_xl(0      , y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_1/2, y0->qs);
+        const int8x16_t v_y1l = vec_xl(0      , y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_1/2, y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0lf, v_y0l), v_x0hf, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1lf, v_y1l), v_x1hf, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_sum0 = vec_madd(v_xy0f, v_d0, v_sum0);
+        v_sum1 = vec_madd(v_xy1f, v_d1, v_sum1);
+    }
+
+    sumf += vec_hsum_f32x4(v_sum0) + vec_hsum_f32x4(v_sum1) + summs0 + summs1;
+
+    #pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        float summs = GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        uint64_t tmp[4];
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        int8x16_t v_qhl = vec_xl(0, (const int8_t *)(tmp + 0));
+        int8x16_t v_qhh = vec_xl(0, (const int8_t *)(tmp + 2));
+
+        // required for fixing the byteorder
+        v_qhl = vec_perm(v_qhl, v_qhl, v_kperm);
+        v_qhh = vec_perm(v_qhh, v_qhh, v_kperm);
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+        const int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        const int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        const int8x16_t v_xlf = vec_or(v_xl, v_qhl);
+        const int8x16_t v_xhf = vec_or(v_xh, v_qhh);
+
+        const int8x16_t v_yl = vec_xl(0      , y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_1/2, y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xlf, v_yl), v_xhf, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
+
+        sumf += vec_hsum_f32x4(v_acc) + summs;
+    }
+
+    *s = sumf;
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -259,7 +573,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
 #if defined(__VXE__) || defined(__VXE2__)
-    __vector float acc = vec_splats(0.0f);
+    float32x4_t acc = vec_splats(0.0f);
 
 #pragma GCC unroll 8
     for (; ib < nb; ++ib) {
@@ -278,7 +592,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         acc = vec_madd(v_xy, v_d, acc);
     }
 
-    sumf = acc[0] + acc[1] + acc[2] + acc[3];
+    sumf = vec_hsum_f32x4(acc);
 
     *s = sumf;
 #else
@@ -402,10 +716,10 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
             isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
 
-            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
-            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
-            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
-            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+            isum += vec_hsum_i32x4(isum0) * scale[0];
+            isum += vec_hsum_i32x4(isum1) * scale[1];
+            isum += vec_hsum_i32x4(isum2) * scale[2];
+            isum += vec_hsum_i32x4(isum3) * scale[3];
 
             scale += 4;
 
@@ -503,7 +817,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             v_xl[1] = (int8x16_t)vec_and(v_x[1], v_lm);
 
             const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
-            sumi1 += (p1[0] + p1[1] + p1[2] + p1[3]) * scales[2*j+0];
+            sumi1 += vec_hsum_i32x4(p1) * scales[2*j+0];
 
             v_y[0] = vec_xl(0 , y0);
             v_y[1] = vec_xl(16, y0);
@@ -513,7 +827,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             v_xl[1] = (int8x16_t)vec_sr(v_x[1], 4);
 
             const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]);
-            sumi2 += (p2[0] + p2[1] + p2[2] + p2[3]) * scales[2*j+1];
+            sumi2 += vec_hsum_i32x4(p2) * scales[2*j+1];
         }
 
         sumf += d * (sumi1 + sumi2);
@@ -595,7 +909,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh);
         const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh);
         const int32x4_t v_mins = vec_add(v_minsho, v_minshe);
-        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+        const int32_t mins = vec_hsum_i32x4(v_mins);
 
         const uint8_t * scales = (const uint8_t *)utmp;
         const uint8_t * GGML_RESTRICT x0l = x[i].qs;
@@ -632,8 +946,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int32x4_t sumi0 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[0], v_y[0]), q5b[1], v_y[1]);
             int32x4_t sumi1 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[2], v_y[2]), q5b[3], v_y[3]);
 
-            sumi += (sumi0[0] + sumi0[1] + sumi0[2] + sumi0[3]) * *scales++;
-            sumi += (sumi1[0] + sumi1[1] + sumi1[2] + sumi1[3]) * *scales++;
+            sumi += vec_hsum_i32x4(sumi0) * *scales++;
+            sumi += vec_hsum_i32x4(sumi1) * *scales++;
         }
 
         sumf += d * sumi - dmin * mins;
@@ -704,7 +1018,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh);
         const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe;
 
-        const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3];
+        const int32_t mins = vec_hsum_i32x4(v_mins);
 
         int32_t isum = 0;
         for (int j = 0; j < QK_K/128; ++j) {
@@ -744,10 +1058,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int32x4_t summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
             int32x4_t summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
 
-            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
-                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
-                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
-                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+            isum += vec_hsum_i32x4(summs0) * scale[0] +
+                    vec_hsum_i32x4(summs1) * scale[1] +
+                    vec_hsum_i32x4(summs2) * scale[2] +
+                    vec_hsum_i32x4(summs3) * scale[3];
 
             scale += 4;
 
@@ -778,10 +1092,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]);
             summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]);
 
-            isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] +
-                    (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] +
-                    (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] +
-                    (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3];
+            isum += vec_hsum_i32x4(summs0) * scale[0] +
+                    vec_hsum_i32x4(summs1) * scale[1] +
+                    vec_hsum_i32x4(summs2) * scale[2] +
+                    vec_hsum_i32x4(summs3) * scale[3];
 
             scale += 4;
         }
@@ -969,7 +1283,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
         const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
 
-        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * vec_hsum_i32x4(v_xy);
     }
 
     *s = sumf;
@@ -1038,8 +1352,8 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
             h >>= 4;
 
-            sumi1 += (vsumi0[0] + vsumi0[1] + vsumi0[2] + vsumi0[3]) * ls1;
-            sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
+            sumi1 += vec_hsum_i32x4(vsumi0) * ls1;
+            sumi2 += vec_hsum_i32x4(vsumi1) * ls2;
         }
 
         sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);