cont : fix comments

ggml-ci

.clang-format

@@ -22,8 +22,8 @@ AllowShortIfStatementsOnASingleLine: Never
 AllowShortLambdasOnASingleLine: Inline
 AllowShortLoopsOnASingleLine: false
 AlwaysBreakBeforeMultilineStrings: true
-BinPackArguments: false
-BinPackParameters: false # OnePerLine
+BinPackArguments: true
+BinPackParameters: true # OnePerLine
 BitFieldColonSpacing: Both
 BreakBeforeBraces: Custom # Attach
 BraceWrapping:
@@ -70,17 +70,14 @@ ExperimentalAutoDetectBinPacking: false
 FixNamespaceComments: true
 IncludeBlocks:   Regroup
 IncludeCategories:
-  - Regex:           '".*"'
+  - Regex:           '^<.*\.h>'
     Priority:        1
     SortPriority:    0
-  - Regex:           '^<.*\.h>'
+  - Regex:           '^<.*'
     Priority:        2
     SortPriority:    0
-  - Regex:           '^<.*'
-    Priority:        3
-    SortPriority:    0
   - Regex:           '.*'
-    Priority:        4
+    Priority:        3
     SortPriority:    0
 IncludeIsMainRegex: '([-_](test|unittest))?$'
 IncludeIsMainSourceRegex: ''

.devops/cann.Dockerfile

@@ -1,130 +0,0 @@
-# ==============================================================================
-# ARGUMENTS
-# ==============================================================================
-
-# Define the CANN base image for easier version updates later
-ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
-
-# ==============================================================================
-# BUILD STAGE
-# Compile all binary files and libraries
-# ==============================================================================
-FROM ${CANN_BASE_IMAGE} AS build
-
-# Define the Ascend chip model for compilation. Default is Ascend910B3
-ARG ASCEND_SOC_TYPE=Ascend910B3
-
-# -- Install build dependencies --
-RUN yum install -y gcc g++ cmake make git libcurl-devel python3 python3-pip && \
-    yum clean all && \
-    rm -rf /var/cache/yum
-
-# -- Set the working directory --
-WORKDIR /app
-
-# -- Copy project files --
-COPY . .
-
-# -- Set CANN environment variables (required for compilation) --
-# Using ENV instead of `source` allows environment variables to persist across the entire image layer
-ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
-ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
-ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
-ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
-ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
-# ... You can add other environment variables from the original file as needed ...
-# For brevity, only core variables are listed here. You can paste the original ENV list here.
-
-# -- Build llama.cpp --
-# Use the passed ASCEND_SOC_TYPE argument and add general build options
-RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
-    && \
-    cmake -B build \
-        -DGGML_CANN=ON \
-        -DCMAKE_BUILD_TYPE=Release \
-        -DSOC_TYPE=${ASCEND_SOC_TYPE} \
-        . && \
-    cmake --build build --config Release -j$(nproc)
-
-# -- Organize build artifacts for copying in later stages --
-# Create a lib directory to store all .so files
-RUN mkdir -p /app/lib && \
-    find build -name "*.so" -exec cp {} /app/lib \;
-
-# Create a full directory to store all executables and Python scripts
-RUN mkdir -p /app/full && \
-    cp build/bin/* /app/full/ && \
-    cp *.py /app/full/ && \
-    cp -r gguf-py /app/full/ && \
-    cp -r requirements /app/full/ && \
-    cp requirements.txt /app/full/
-    # If you have a tools.sh script, make sure it is copied here
-    # cp .devops/tools.sh /app/full/tools.sh
-
-# ==============================================================================
-# BASE STAGE
-# Create a minimal base image with CANN runtime and common libraries
-# ==============================================================================
-FROM ${CANN_BASE_IMAGE} AS base
-
-# -- Install runtime dependencies --
-RUN yum install -y libgomp curl && \
-    yum clean all && \
-    rm -rf /var/cache/yum
-
-# -- Set CANN environment variables (required for runtime) --
-ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
-ENV LD_LIBRARY_PATH=/app:${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
-ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
-ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
-# ... You can add other environment variables from the original file as needed ...
-
-WORKDIR /app
-
-# Copy compiled .so files from the build stage
-COPY --from=build /app/lib/ /app
-
-# ==============================================================================
-# FINAL STAGES (TARGETS)
-# ==============================================================================
-
-### Target: full
-# Complete image with all tools, Python bindings, and dependencies
-# ==============================================================================
-FROM base AS full
-
-COPY --from=build /app/full /app
-
-# Install Python dependencies
-RUN yum install -y git python3 python3-pip && \
-    pip3 install --no-cache-dir --upgrade pip setuptools wheel && \
-    pip3 install --no-cache-dir -r requirements.txt && \
-    yum clean all && \
-    rm -rf /var/cache/yum
-
-# You need to provide a tools.sh script as the entrypoint
-ENTRYPOINT ["/app/tools.sh"]
-# If there is no tools.sh, you can set the default to start the server
-# ENTRYPOINT ["/app/llama-server"]
-
-### Target: light
-# Lightweight image containing only llama-cli
-# ==============================================================================
-FROM base AS light
-
-COPY --from=build /app/full/llama-cli /app
-
-ENTRYPOINT [ "/app/llama-cli" ]
-
-### Target: server
-# Dedicated server image containing only llama-server
-# ==============================================================================
-FROM base AS server
-
-ENV LLAMA_ARG_HOST=0.0.0.0
-
-COPY --from=build /app/full/llama-server /app
-
-HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]
-
-ENTRYPOINT [ "/app/llama-server" ]

.devops/cloud-v-pipeline

@@ -0,0 +1,22 @@
+node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
+    stage('Cleanup'){
+        cleanWs()               // Cleaning previous CI build in workspace
+    }
+    stage('checkout repo'){
+        retry(5){               // Retry if the cloning fails due to some reason
+            checkout scm        // Clone the repo on Runner
+        }
+    }
+    stage('Compiling llama.cpp'){
+        sh'''#!/bin/bash
+            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
+        '''
+    }
+    stage('Running llama.cpp'){
+        sh'''#!/bin/bash
+            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
+            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
+            cat llama_log.txt                   # Printing results
+        '''
+    }
+}

.devops/cpu.Dockerfile

@@ -4,6 +4,8 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
 
+ARG GGML_CPU_ARM_ARCH=armv8-a
+
 RUN apt-get update && \
     apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
@@ -11,8 +13,10 @@ WORKDIR /app
 
 COPY . .
 
-RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
+RUN if [ "$TARGETARCH" = "amd64" ]; then \
         cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+    elif [ "$TARGETARCH" = "arm64" ]; then \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -61,7 +61,7 @@ RUN apt-get update \
     python3 \
     python3-pip \
     && pip install --upgrade pip setuptools wheel \
-    && pip install --break-system-packages -r requirements.txt \
+    && pip install -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/intel.Dockerfile

@@ -49,23 +49,19 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update && \
-    apt-get install -y \
-        git \
-        python3 \
-        python3-pip \
-        python3-venv && \
-    python3 -m venv /opt/venv && \
-    . /opt/venv/bin/activate && \
-    pip install --upgrade pip setuptools wheel && \
-    pip install -r requirements.txt && \
-    apt autoremove -y && \
-    apt clean -y && \
-    rm -rf /tmp/* /var/tmp/* && \
-    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
-    find /var/cache -type f -delete
+RUN apt-get update \
+    && apt-get install -y \
+    git \
+    python3 \
+    python3-pip \
+    && pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt \
+    && apt autoremove -y \
+    && apt clean -y \
+    && rm -rf /tmp/* /var/tmp/* \
+    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
+    && find /var/cache -type f -delete
 
-ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 

.devops/musa.Dockerfile

@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc4.2.0
+ARG MUSA_VERSION=rc4.0.1
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-devel-ubuntu${UBUNTU_VERSION}
 
-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-mudnn-runtime-ubuntu${UBUNTU_VERSION}
 
 FROM ${BASE_MUSA_DEV_CONTAINER} AS build
 

.devops/nix/package.nix

@@ -47,7 +47,6 @@ let
   inherit (lib)
     cmakeBool
     cmakeFeature
-    optionalAttrs
     optionals
     strings
     ;
@@ -198,7 +197,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     ];
 
   # Environment variables needed for ROCm
-  env = optionalAttrs useRocm {
+  env = optionals useRocm {
     ROCM_PATH = "${rocmPackages.clr}";
     HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
   };

.devops/rocm.Dockerfile

@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04
 
 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=6.4
-ARG AMDGPU_VERSION=6.4
+ARG ROCM_VERSION=6.3
+ARG AMDGPU_VERSION=6.3
 
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete

.devops/tools.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 # Read the first argument into a variable

.devops/vulkan.Dockerfile

@@ -2,30 +2,14 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-# Ref: https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html
-
 # Install build tools
-RUN apt update && apt install -y git build-essential cmake wget xz-utils
+RUN apt update && apt install -y git build-essential cmake wget
 
-# 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
+# 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
 
 # Build it
 WORKDIR /app

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
     validations:
       required: true

.github/copilot-instructions.md

@@ -1,262 +0,0 @@
-# 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/labeler.yml

@@ -1,4 +1,10 @@
 # https://github.com/actions/labeler
+Kompute:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-kompute.h
+            - ggml/src/ggml-kompute/**
+            - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
@@ -22,11 +28,6 @@ Vulkan:
         - any-glob-to-any-file:
             - ggml/include/ggml-vulkan.h
             - ggml/src/ggml-vulkan/**
-IBM zDNN:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-zdnn.h
-            - ggml/src/ggml-zdnn/**
 documentation:
     - changed-files:
         - any-glob-to-any-file:
@@ -92,8 +93,3 @@ Ascend NPU:
             - ggml/include/ggml-cann.h
             - ggml/src/ggml-cann/**
             - docs/backend/CANN.md
-OpenCL:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-opencl.h
-            - ggml/src/ggml-opencl/**

.github/workflows/build-cmake-pkg.yml

@@ -1,51 +0,0 @@
-name: Build relocatable cmake package
-on:
-  workflow_dispatch:
-  workflow_call:
-
-jobs:
-  linux:
-    runs-on: ubuntu-24.04
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y build-essential tcl
-
-      - name: Build
-        run: |
-          PREFIX="$(pwd)"/inst
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
-                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release
-          cmake --install build --prefix "$PREFIX" --config Release
-
-          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
-          set build(version) "0.0.$build(number)"
-
-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
-
-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
-                  exit 1
-              }
-          }
-          puts "success."
-          EOF
-
-          cd examples/simple-cmake-pkg
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build

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

@@ -48,98 +48,98 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  # ubuntu-24-riscv64-vulkan-cross:
-  #   runs-on: ubuntu-24.04
+  ubuntu-24-riscv64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
-  #   steps:
-  #     - uses: actions/checkout@v4
-  #     - name: Setup Riscv
-  #       run: |
-  #         sudo dpkg --add-architecture riscv64
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo dpkg --add-architecture riscv64
 
-  #         # Add arch-specific repositories for non-amd64 architectures
-  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-  #         EOF
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
 
-  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 glslc \
-  #                 gcc-14-riscv64-linux-gnu \
-  #                 g++-14-riscv64-linux-gnu \
-  #                 libvulkan-dev:riscv64
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu \
+                  libvulkan-dev:riscv64
 
-  #     - name: Build
-  #       run: |
-  #         cmake -B build -DLLAMA_CURL=OFF \
-  #                        -DCMAKE_BUILD_TYPE=Release \
-  #                        -DGGML_VULKAN=ON \
-  #                        -DGGML_OPENMP=OFF \
-  #                        -DLLAMA_BUILD_EXAMPLES=ON \
-  #                        -DLLAMA_BUILD_TOOLS=ON \
-  #                        -DLLAMA_BUILD_TESTS=OFF \
-  #                        -DCMAKE_SYSTEM_NAME=Linux \
-  #                        -DCMAKE_SYSTEM_PROCESSOR=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
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=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 --build build --config Release -j $(nproc)
+          cmake --build build --config Release -j $(nproc)
 
-  # ubuntu-24-arm64-vulkan-cross:
-  #   runs-on: ubuntu-24.04
+  ubuntu-24-arm64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
-  #   steps:
-  #     - uses: actions/checkout@v4
-  #     - name: Setup Arm64
-  #       run: |
-  #         sudo dpkg --add-architecture arm64
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup Arm64
+        run: |
+          sudo dpkg --add-architecture arm64
 
-  #         # Add arch-specific repositories for non-amd64 architectures
-  #         cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
-  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-  #         deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-  #         EOF
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
 
-  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 glslc \
-  #                 crossbuild-essential-arm64 \
-  #                 libvulkan-dev:arm64
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  crossbuild-essential-arm64 \
+                  libvulkan-dev:arm64
 
-  #     - name: Build
-  #       run: |
-  #         cmake -B build -DLLAMA_CURL=OFF \
-  #                        -DCMAKE_BUILD_TYPE=Release \
-  #                        -DGGML_VULKAN=ON \
-  #                        -DGGML_OPENMP=OFF \
-  #                        -DLLAMA_BUILD_EXAMPLES=ON \
-  #                        -DLLAMA_BUILD_TOOLS=ON \
-  #                        -DLLAMA_BUILD_TESTS=OFF \
-  #                        -DCMAKE_SYSTEM_NAME=Linux \
-  #                        -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
-  #                        -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
-  #                        -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
-  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=aarch64 \
+                         -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \
+                         -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-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)
+          cmake --build build --config Release -j $(nproc)
 
   ubuntu-24-ppc64el-cpu-cross:
     runs-on: ubuntu-24.04
@@ -185,52 +185,52 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  # ubuntu-24-ppc64el-vulkan-cross:
-  #   runs-on: ubuntu-24.04
+  ubuntu-24-ppc64el-vulkan-cross:
+    runs-on: ubuntu-24.04
 
-  #   steps:
-  #     - uses: actions/checkout@v4
-  #     - name: Setup PowerPC64le
-  #       run: |
-  #         sudo dpkg --add-architecture ppc64el
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
 
-  #         # Add arch-specific repositories for non-amd64 architectures
-  #         cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-  #         EOF
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
 
-  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 glslc \
-  #                 gcc-14-powerpc64le-linux-gnu \
-  #                 g++-14-powerpc64le-linux-gnu \
-  #                 libvulkan-dev:ppc64el
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu \
+                  libvulkan-dev:ppc64el
 
-  #     - name: Build
-  #       run: |
-  #         cmake -B build -DLLAMA_CURL=OFF \
-  #                        -DCMAKE_BUILD_TYPE=Release \
-  #                        -DGGML_VULKAN=ON \
-  #                        -DGGML_OPENMP=OFF \
-  #                        -DLLAMA_BUILD_EXAMPLES=ON \
-  #                        -DLLAMA_BUILD_TOOLS=ON \
-  #                        -DLLAMA_BUILD_TESTS=OFF \
-  #                        -DCMAKE_SYSTEM_NAME=Linux \
-  #                        -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-  #                        -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-  #                        -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-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)
+          cmake --build build --config Release -j $(nproc)
 
   debian-13-loongarch64-cpu-cross:
     runs-on: ubuntu-24.04

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

@@ -1,60 +0,0 @@
-name: Build on RISCV Linux Machine by Cloud-V
-on:
-  pull_request:
-  workflow_dispatch:
-  workflow_call:
-
-jobs:
-  debian-13-riscv64-native: # Bianbu 2.2
-    runs-on: self-hosted
-
-    steps:
-      - name: Install prerequisites
-        run: |
-          sudo apt-get update || true
-          sudo apt-get install -y libatomic1
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo apt-get update || true
-          sudo apt-get install -y --no-install-recommends \
-                  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_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

@@ -5,43 +5,10 @@ on:
   push:
     branches:
       - master
-    paths: [
-      '.github/workflows/build.yml',
-      '.github/workflows/build-linux-cross.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.swift',
-      '**/*.m',
-      '**/*.metal',
-      '**/*.comp'
-    ]
-
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build.yml',
-      '.github/workflows/build-linux-cross.yml',
-      '.github/workflows/build-cmake-pkg.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.cu',
-      '**/*.cuh',
-      '**/*.swift',
-      '**/*.m',
-      '**/*.metal',
-      '**/*.comp'
-    ]
+    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -64,7 +31,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -84,8 +51,7 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=OFF \
-            -DGGML_METAL_SHADER_DEBUG=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -104,7 +70,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -135,53 +101,6 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-  macOS-latest-cmake-arm64-webgpu:
-    runs-on: macos-14
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: macOS-latest-cmake-arm64-webgpu
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v1.0.0"
-          DAWN_OWNER="reeselevine"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-macos-latest-Release.tar.gz"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Build
-        id: cmake_build
-        run: |
-          export CMAKE_PREFIX_PATH=dawn
-          cmake -B build -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
   ubuntu-cpu-cmake:
     strategy:
       matrix:
@@ -199,7 +118,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -251,7 +170,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-sanitizer-${{ matrix.sanitizer }}
           evict-old-files: 1d
@@ -330,7 +249,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-rpc
           evict-old-files: 1d
@@ -363,7 +282,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -389,61 +308,11 @@ jobs:
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4200
-
-  ubuntu-22-cmake-webgpu:
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: ubuntu-22-cmake-webgpu
-          evict-old-files: 1d
-
-      - name: Vulkan SDK Dependencies
-        id: vulkan-depends
-        run: |
-          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
-          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v1.0.0"
-          DAWN_OWNER="reeselevine"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-ubuntu-latest-Release.tar.gz"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
-
-      - name: Build
-        id: cmake_build
-        run: |
-          export Dawn_DIR=dawn/lib64/cmake/Dawn
-          cmake -B build -DGGML_WEBGPU=ON
-          cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
-    container: rocm/dev-ubuntu-22.04:6.1.2
+    container: rocm/dev-ubuntu-22.04:6.0.2
 
     steps:
       - name: Clone
@@ -457,7 +326,7 @@ jobs:
           sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-hip
           evict-old-files: 1d
@@ -471,9 +340,19 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
+      - name: Build with legacy HIP support
+        id: cmake_build_legacy_hip
+        run: |
+          cmake -B build2 -S . \
+            -DCMAKE_C_COMPILER=hipcc \
+            -DCMAKE_CXX_COMPILER=hipcc \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
+            -DGGML_HIP=ON
+          cmake --build build2 --config Release -j $(nproc)
+
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
-    container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
+    container: mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 
     steps:
       - name: Clone
@@ -487,7 +366,7 @@ jobs:
           apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-musa
           evict-old-files: 1d
@@ -532,7 +411,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl
           evict-old-files: 1d
@@ -580,7 +459,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl-fp16
           evict-old-files: 1d
@@ -599,9 +478,6 @@ jobs:
   build-linux-cross:
     uses: ./.github/workflows/build-linux-cross.yml
 
-  build-cmake-pkg:
-    uses: ./.github/workflows/build-cmake-pkg.yml
-
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -611,7 +487,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-ios
           evict-old-files: 1d
@@ -648,7 +524,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-tvos
           evict-old-files: 1d
@@ -720,7 +596,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-swift
           evict-old-files: 1d
@@ -752,7 +628,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     strategy:
       fail-fast: false
@@ -766,7 +642,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-msys2
           variant: ccache
@@ -802,31 +678,28 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
         include:
           - build: 'cpu-x64 (static)'
-            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
-            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
-            arch: 'x64'
-            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
-            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
-            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
+         # - build: 'kompute-x64'
+         #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
       - name: Clone
@@ -834,12 +707,18 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.build }}
           variant: ccache
           evict-old-files: 1d
 
+      - name: Clone Kompute submodule
+        id: clone_kompute
+        if: ${{ matrix.build == 'kompute-x64' }}
+        run: |
+          git submodule update --init ggml/src/ggml-kompute/kompute
+
       - name: Download OpenBLAS
         id: get_openblas
         if: ${{ matrix.build == 'openblas-x64' }}
@@ -855,9 +734,9 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -890,8 +769,6 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -901,7 +778,6 @@ jobs:
           cmake -S . -B build ${{ matrix.defines }} `
             -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include"
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
@@ -912,7 +788,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.arch == 'x64' }}
+        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -948,7 +824,7 @@ jobs:
               apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
 
         - name: ccache
-          uses: ggml-org/ccache-action@v1.2.16
+          uses: hendrikmuhs/ccache-action@v1.2.16
           with:
             key: ubuntu-latest-cmake-cuda
             evict-old-files: 1d
@@ -977,7 +853,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -1017,7 +893,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     defaults:
       run:
@@ -1033,7 +909,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -1051,7 +927,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     steps:
       - name: Clone
@@ -1070,8 +946,7 @@ jobs:
           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"
           write-host "Installing AMD HIP SDK"
-          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
@@ -1080,7 +955,7 @@ jobs:
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
       - name: Install ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ${{ github.job }}
           evict-old-files: 1d
@@ -1114,11 +989,6 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
-      - name: Setup Xcode
-        uses: maxim-lobanov/setup-xcode@v1
-        with:
-          xcode-version: latest-stable
-
       - name: Build
         id: cmake_build
         run: |
@@ -1152,7 +1022,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: android-build
           evict-old-files: 1d

.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: "refactor,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

@@ -1,57 +0,0 @@
-name: "Copilot Setup Steps"
-
-# Automatically run the setup steps when they are changed to allow for easy validation, and
-# allow manual testing through the repository's "Actions" tab
-on:
-  workflow_dispatch:
-  push:
-    paths:
-      - .github/workflows/copilot-setup-steps.yml
-  pull_request:
-    paths:
-      - .github/workflows/copilot-setup-steps.yml
-
-jobs:
-  # The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
-  copilot-setup-steps:
-    runs-on: ubuntu-latest
-
-    # Set the permissions to the lowest permissions possible needed for your steps.
-    # Copilot will be given its own token for its operations.
-    permissions:
-      # If you want to clone the repository as part of your setup steps, for example to install dependencies, you'll need the `contents: read` permission. If you don't clone the repository in your setup steps, Copilot will do this for you automatically after the steps complete.
-      contents: read
-
-    # You can define any steps you want, and they will run before the agent starts.
-    # If you do not check out your code, Copilot will do this for you.
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: copilot-setup-steps
-          evict-old-files: 1d
-
-      - name: Dependencies
-        id: depends
-        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
-        with:
-          python-version: '3.11'
-
-      - name: Install Python dependencies
-        run: |
-          python3 -m venv .venv
-          .venv/bin/activate
-          pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
-          pip install flake8 pyright pre-commit

.github/workflows/pre-tokenizer-hashes.yml

@@ -1,45 +0,0 @@
-name: Check Pre-Tokenizer Hashes
-
-on:
-    push:
-        paths:
-            - 'convert_hf_to_gguf.py'
-            - 'convert_hf_to_gguf_update.py'
-    pull_request:
-        paths:
-            - 'convert_hf_to_gguf.py'
-            - 'convert_hf_to_gguf_update.py'
-
-jobs:
-    pre-tokenizer-hashes:
-        runs-on: ubuntu-latest
-
-        steps:
-        - name: Checkout repository
-          uses: actions/checkout@v4
-
-        - name: Set up Python
-          uses: actions/setup-python@v5
-          with:
-              python-version: '3.11'
-
-        - name: Install Python dependencies
-          run: |
-              python3 -m venv .venv
-              .venv/bin/pip install -r requirements/requirements-convert_hf_to_gguf_update.txt
-
-        - name: Update pre-tokenizer hashes
-          run: |
-              cp convert_hf_to_gguf.py /tmp
-              .venv/bin/python convert_hf_to_gguf_update.py --check-missing
-
-        - name: Check if committed pre-tokenizer hashes matches generated version
-          run: |
-              if ! diff -q convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py; then
-                  echo "Model pre-tokenizer hashes (in convert_hf_to_gguf.py) do not match generated hashes (from convert_hf_to_gguf_update.py)."
-                  echo "To fix: run ./convert_hf_to_gguf_update.py and commit the updated convert_hf_to_gguf.py along with your changes"
-                  echo "Differences found:"
-                  diff convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py || true
-                  exit 1
-              fi
-              echo "Model pre-tokenizer hashes are up to date."

.github/workflows/release.yml

@@ -32,7 +32,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -49,8 +49,7 @@ jobs:
         run: |
           sysctl -a
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='@loader_path' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -85,7 +84,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -104,8 +103,7 @@ jobs:
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='@loader_path' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
@@ -147,7 +145,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -162,8 +160,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DGGML_BACKEND_DL=ON \
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -198,7 +194,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -215,8 +211,6 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
-            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DGGML_BACKEND_DL=ON \
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -241,7 +235,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -256,7 +250,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-cpu-${{ matrix.arch }}
           variant: ccache
@@ -277,7 +271,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DGGML_NATIVE=OFF ^
@@ -294,7 +288,7 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -304,11 +298,11 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows:
-    runs-on: windows-2025
+    runs-on: windows-latest
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.309.0
 
     strategy:
       matrix:
@@ -328,7 +322,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
           variant: ccache
@@ -338,7 +332,7 @@ jobs:
         id: get_vulkan
         if: ${{ matrix.backend == 'vulkan' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -398,7 +392,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -454,7 +448,7 @@ jobs:
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     defaults:
       run:
@@ -471,7 +465,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -526,7 +520,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-2022
+    runs-on: windows-latest
 
     strategy:
       matrix:
@@ -545,7 +539,7 @@ jobs:
           git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
+        uses: hendrikmuhs/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-hip-${{ matrix.name }}-x64
           evict-old-files: 1d
@@ -557,8 +551,7 @@ jobs:
           write-host "Downloading AMD HIP SDK Installer"
           Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
-          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
-          $proc.WaitForExit(600000)
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP SDK installation"
 
       - name: Verify ROCm
@@ -601,7 +594,7 @@ jobs:
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode-build:
-    runs-on: macos-15
+    runs-on: macos-latest
 
     steps:
       - name: Checkout code
@@ -609,10 +602,6 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Setup Xcode
-        run: |
-          sudo xcode-select -s /Applications/Xcode_16.4.app
-
       - name: Build
         id: cmake_build
         run: |

.github/workflows/update-ops-docs.yml

@@ -1,40 +0,0 @@
-name: Update Operations Documentation
-
-on:
-    push:
-        paths:
-            - 'docs/ops/**'
-            - 'scripts/create_ops_docs.py'
-    pull_request:
-        paths:
-            - 'docs/ops/**'
-            - 'scripts/create_ops_docs.py'
-
-jobs:
-    update-ops-docs:
-        runs-on: ubuntu-latest
-
-        steps:
-        - name: Checkout repository
-          uses: actions/checkout@v4
-
-        - name: Set up Python
-          uses: actions/setup-python@v5
-          with:
-              python-version: '3.x'
-
-        - name: Generate operations documentation to temporary file
-          run: |
-              mkdir -p /tmp/ops_check
-              ./scripts/create_ops_docs.py /tmp/ops_check/ops.md
-
-        - name: Check if docs/ops.md matches generated version
-          run: |
-              if ! diff -q docs/ops.md /tmp/ops_check/ops.md; then
-                  echo "Operations documentation (docs/ops.md) is not up to date with the backend CSV files."
-                  echo "To fix: run ./scripts/create_ops_docs.py and commit the updated docs/ops.md along with your changes"
-                  echo "Differences found:"
-                  diff docs/ops.md /tmp/ops_check/ops.md || true
-                  exit 1
-              fi
-              echo "Operations documentation is up to date."

.gitignore

@@ -82,7 +82,6 @@ models/*
 models-mnt
 !models/.editorconfig
 !models/ggml-vocab-*.gguf*
-!models/templates
 
 # Zig
 zig-out/
@@ -147,4 +146,3 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
-.ccache/

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "kompute"]
+	path = ggml/src/ggml-kompute/kompute
+	url = https://github.com/nomic-ai/kompute.git

CMakeLists.txt

@@ -12,8 +12,6 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()
 
-message("CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}")
-
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 
@@ -91,14 +89,6 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
-if (NOT DEFINED LLAMA_BUILD_NUMBER)
-    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-endif()
-if (NOT DEFINED LLAMA_BUILD_COMMIT)
-    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-endif()
-set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
-
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -122,6 +112,7 @@ endfunction()
 
 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
+llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
@@ -164,8 +155,6 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
-    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
-    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
@@ -215,6 +204,10 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
+set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")

CMakePresets.json

@@ -55,17 +55,6 @@
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
         }
     },
-    {
-        "name": "x64-linux-gcc", "hidden": true,
-        "cacheVariables": {
-            "CMAKE_C_COMPILER": "gcc",
-            "CMAKE_CXX_COMPILER": "g++"
-        }
-    },
-    { "name": "x64-linux-gcc-debug", "inherits": [ "base", "x64-linux-gcc", "debug" ] },
-    { "name": "x64-linux-gcc-release", "inherits": [ "base", "x64-linux-gcc", "release" ] },
-    { "name": "x64-linux-gcc-reldbg", "inherits": [ "base", "x64-linux-gcc", "reldbg" ] },
-    { "name": "x64-linux-gcc+static-release", "inherits": [ "base", "x64-linux-gcc", "release", "static" ] },
 
     { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },

CODEOWNERS

@@ -5,8 +5,7 @@
 /tools/server/ @ngxson
 /ggml/src/ggml-cuda/fattn* @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.* @JohannesGaessler
+/ggml/src/ggml-cuda/mmv.* @JohannesGaessler
 /ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
 /ggml/src/ggml-opt.cpp @JohannesGaessler
 /ggml/src/gguf.cpp @JohannesGaessler
-/ggml/src/ggml-vulkan/ @0cc4m
-/ggml/src/ggml-zdnn/ @taronaeo

README.md

@@ -6,9 +6,9 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
-LLM inference in C/C++
+Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
 ## Recent API changes
 
@@ -17,12 +17,11 @@ LLM inference in C/C++
 
 ## Hot topics
 
-- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
-- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
-- Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
-- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
-- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
+- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
+- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
+- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -136,7 +135,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [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)
 
 #### Multimodal
 
@@ -151,7 +149,6 @@ 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>
 
@@ -243,7 +240,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Infrastructure</summary>
 
-- [Paddler](https://github.com/intentee/paddler) - Open-source LLMOps platform for hosting and scaling AI in your own infrastructure
+- [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
@@ -273,7 +270,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
 ## Obtaining and quantizing models
@@ -439,7 +435,7 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
 ## [`llama-perplexity`](tools/perplexity)
 
-#### A tool for measuring the [perplexity](tools/perplexity/README.md) [^1] (and other quality metrics) of a model over a given text.
+#### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.
 
 - <details open>
     <summary>Measure the perplexity over a text file</summary>
@@ -462,7 +458,8 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
-[^1]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
+[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md)
+[^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
 
 ## [`llama-bench`](tools/llama-bench)
 

build-xcframework.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #
 # Options
 IOS_MIN_OS_VERSION=16.4

ci/README.md

@@ -54,7 +54,7 @@ docker run --privileged -it \
     -v $HOME/llama.cpp/ci-cache:/ci-cache \
     -v $HOME/llama.cpp/ci-results:/ci-results \
     -v $PWD:/ws -w /ws \
-    mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
+    mthreads/musa:rc4.0.1-mudnn-devel-ubuntu22.04
 ```
 
 Inside the container, execute the following commands:

ci/run.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -16,9 +16,6 @@
 # # with VULKAN support
 # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with WebGPU support
-# GG_BUILD_WEBGPU=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 # # with MUSA support
 # GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
@@ -42,7 +39,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
@@ -84,10 +81,6 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
 fi
 
-if [ ! -z ${GG_BUILD_WEBGPU} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
-fi
-
 if [ ! -z ${GG_BUILD_MUSA} ]; then
     # Use qy1 by default (MTT S80)
     MUSA_ARCH=${MUSA_ARCH:-21}
@@ -106,7 +99,7 @@ function gg_wget {
     cd $out
 
     # should not re-download if file is the same
-    wget -nv -c -N $url
+    wget -nv -N $url
 
     cd $cwd
 }
@@ -786,7 +779,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/CMakeLists.txt

@@ -23,21 +23,31 @@ if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        # For build-info.cpp below
-        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
 endif()
 
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
-configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+    WORKING_DIRECTORY "${PROJECT_SOURCE_DIR}"
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
 set(TARGET build_info)
-add_library(${TARGET} OBJECT ${OUTPUT_FILE})
+add_library(${TARGET} OBJECT build-info.cpp)
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
@@ -86,7 +96,8 @@ if (LLAMA_CURL)
     endif()
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
+    find_library(CURL_LIBRARY curl REQUIRED)
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
 endif ()
 
 if (LLAMA_LLGUIDANCE)
@@ -111,13 +122,13 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v1.0.1:
-        GIT_TAG d795912fedc7d393de740177ea9ea761e7905774
+        # v0.7.20 (+ fix to build on GCC 15):
+        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE
         CONFIGURE_COMMAND ""
-        BUILD_COMMAND cargo build --release --package llguidance
+        BUILD_COMMAND cargo build --release
         INSTALL_COMMAND ""
         BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
         UPDATE_COMMAND ""

common/arg.cpp

@@ -24,7 +24,6 @@
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
-#include <list>
 #include <regex>
 #include <set>
 #include <string>
@@ -749,39 +748,6 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 // utils
 //
 
-// Helper function to parse tensor buffer override strings
-static void parse_tensor_buffer_overrides(const std::string & value, std::vector<llama_model_tensor_buft_override> & overrides) {
-    std::map<std::string, ggml_backend_buffer_type_t> buft_list;
-    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-        auto * dev = ggml_backend_dev_get(i);
-        auto * buft = ggml_backend_dev_buffer_type(dev);
-        if (buft) {
-            buft_list[ggml_backend_buft_name(buft)] = buft;
-        }
-    }
-
-    for (const auto & override : string_split<std::string>(value, ',')) {
-        std::string::size_type pos = override.find('=');
-        if (pos == std::string::npos) {
-            throw std::invalid_argument("invalid value");
-        }
-        std::string tensor_name = override.substr(0, pos);
-        std::string buffer_type = override.substr(pos + 1);
-
-        if (buft_list.find(buffer_type) == buft_list.end()) {
-            printf("Available buffer types:\n");
-            for (const auto & it : buft_list) {
-                printf("  %s\n", ggml_backend_buft_name(it.second));
-            }
-            throw std::invalid_argument("unknown buffer type");
-        }
-        // 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(tensor_name);
-        overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
-    }
-}
-
 struct handle_model_result {
     bool found_mmproj = false;
     common_params_model mmproj;
@@ -1011,10 +977,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
             string_process_escapes(seq_breaker);
         }
-        for (auto & pair : params.speculative.replacements) {
-            string_process_escapes(pair.first);
-            string_process_escapes(pair.second);
-        }
     }
 
     if (!params.kv_overrides.empty()) {
@@ -1026,8 +988,8 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
-    if (!params.speculative.tensor_buft_overrides.empty()) {
-        params.speculative.tensor_buft_overrides.push_back({nullptr, nullptr});
+    if (params.reranking && params.embedding) {
+        throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
     }
 
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
@@ -1238,7 +1200,6 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
             common_params_print_completion(ctx_arg);
             exit(0);
         }
-        params.lr.init();
     } catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
@@ -1507,22 +1468,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
-    add_opt(common_arg(
-        {"--swa-checkpoints"}, "N",
-        string_format("max number of SWA checkpoints per slot to create (default: %d)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_swa_checkpoints),
-        [](common_params & params, int value) {
-            params.n_swa_checkpoints = value;
-        }
-    ).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--kv-unified", "-kvu"},
-        string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
-            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_unified ? "true" : "false"),
-        [](common_params & params) {
-            params.kv_unified = true;
-        }
-    ).set_env("LLAMA_ARG_KV_SPLIT"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -1530,13 +1475,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ctx_shift = false;
         }
     ).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 ? "enabled" : "disabled"),
-        [](common_params & params) {
-            params.ctx_shift = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY}).set_env("LLAMA_ARG_CONTEXT_SHIFT"));
     add_opt(common_arg(
         {"--chunks"}, "N",
         string_format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
@@ -1670,7 +1608,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.antiprompt.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"-sp", "--special"},
         string_format("special tokens output enabled (default: %s)", params.special ? "true" : "false"),
@@ -1755,7 +1693,7 @@ 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, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -1830,7 +1768,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
         }
-    ).set_sparam());
+    ).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
@@ -2149,13 +2087,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.no_kv_offload = true;
         }
     ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
-    add_opt(common_arg(
-        {"-nr", "--no-repack"},
-        "disable weight repacking",
-        [](common_params & params) {
-            params.no_extra_bufts = true;
-        }
-    ).set_env("LLAMA_ARG_NO_REPACK"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         string_format(
@@ -2254,11 +2185,9 @@ 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 (DEPRECATED)"),
+        string_format("KV cache defragmentation threshold (default: %.1f, < 0 - disabled)", (double)params.defrag_thold),
         [](common_params & params, const std::string & value) {
-            GGML_UNUSED(params);
-            GGML_UNUSED(value);
-            LOG_WRN("DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify\n");
+            params.defrag_thold = std::stof(value);
         }
     ).set_env("LLAMA_ARG_DEFRAG_THOLD"));
     add_opt(common_arg(
@@ -2404,58 +2333,38 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
         "override tensor buffer type", [](common_params & params, const std::string & value) {
-            parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
+            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+            if (buft_list.empty()) {
+                // enumerate all the devices and add their buffer types to the list
+                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    auto * buft = ggml_backend_dev_buffer_type(dev);
+                    if (buft) {
+                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                    }
+                }
+            }
+
+            for (const auto & override : string_split<std::string>(value, ',')) {
+                std::string::size_type pos = override.find('=');
+                if (pos == std::string::npos) {
+                    throw std::invalid_argument("invalid value");
+                }
+                std::string tensor_name = override.substr(0, pos);
+                std::string buffer_type = override.substr(pos + 1);
+
+                if (buft_list.find(buffer_type) == buft_list.end()) {
+                    printf("Available buffer types:\n");
+                    for (const auto & it : buft_list) {
+                        printf("  %s\n", ggml_backend_buft_name(it.second));
+                    }
+                    throw std::invalid_argument("unknown buffer type");
+                }
+                // FIXME: this leaks memory
+                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+            }
         }
     ));
-    add_opt(common_arg(
-        {"--override-tensor-draft", "-otd"}, "<tensor name pattern>=<buffer type>,...",
-        "override tensor buffer type for draft model", [](common_params & params, const std::string & value) {
-            parse_tensor_buffer_overrides(value, params.speculative.tensor_buft_overrides);
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--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()});
-        }
-    ).set_env("LLAMA_ARG_CPU_MOE"));
-    add_opt(common_arg(
-        {"--n-cpu-moe", "-ncmoe"}, "N",
-        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU",
-        [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("invalid value");
-            }
-            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));
-                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
-            }
-        }
-    ).set_env("LLAMA_ARG_N_CPU_MOE"));
-    add_opt(common_arg(
-        {"--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()});
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
-    add_opt(common_arg(
-        {"--n-cpu-moe-draft", "-ncmoed"}, "N",
-        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model",
-        [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("invalid value");
-            }
-            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));
-                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",
@@ -2706,7 +2615,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -2714,15 +2623,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_out_freq = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
-    add_opt(common_arg(
-        {"--output-format"}, "{gguf,dat}",
-        string_format("output format for imatrix file (default: %s)", params.imat_dat > 0 ? "dat" : "gguf"),
-        [](common_params & params, const std::string & value) {
-            /**/ if (value == "gguf") { params.imat_dat = -1; }
-            else if (value == "dat")  { params.imat_dat = 1;  }
-            else { throw std::invalid_argument("invalid output format"); }
-        }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--save-frequency"}, "N",
         string_format("save an imatrix copy every N iterations (default: %d)", params.n_save_freq),
@@ -2751,13 +2651,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
-    add_opt(common_arg(
-        {"--show-statistics"},
-        string_format("show imatrix statistics and then exit (default: %s)", params.show_statistics ? "true" : "false"),
-        [](common_params & params) {
-            params.show_statistics = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
         string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),
@@ -2817,13 +2710,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.embd_sep = value;
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
-    add_opt(common_arg(
-        {"--cls-separator"}, "STRING",
-        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
-        [](common_params & params, const std::string & value) {
-            params.cls_sep = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
         string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -2845,13 +2731,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.public_path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
-    add_opt(common_arg(
-        {"--api-prefix"}, "PREFIX",
-        string_format("prefix path the server serves from, without the trailing slash (default: %s)", params.api_prefix.c_str()),
-        [](common_params & params, const std::string & value) {
-            params.api_prefix = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
     add_opt(common_arg(
         {"--no-webui"},
         string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
@@ -2868,10 +2747,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
         [](common_params & params) {
-            params.embedding = true;
-            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
+            params.reranking = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -2912,16 +2790,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ssl_file_cert = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
-    add_opt(common_arg(
-        {"--chat-template-kwargs"}, "STRING",
-        string_format("sets additional params for the json template parser"),
-        [](common_params & params, const std::string &  value) {
-            auto parsed = json::parse(value);
-            for (const auto & item : parsed.items()) {
-                params.default_template_kwargs[item.key()] = item.value().dump();
-            }
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -2998,9 +2866,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
         "- none: leaves thoughts unparsed in `message.content`\n"
         "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
-        "(default: auto)",
+        "(default: deepseek)",
         [](common_params & params, const std::string & value) {
-            params.reasoning_format = common_reasoning_format_from_name(value);
+            /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
+            else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
     add_opt(common_arg(
@@ -3181,7 +3052,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-tbd", "--threads-batch-draft"}, "N",
         "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
@@ -3191,7 +3062,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
     add_opt(common_arg(
         {"-Cd", "--cpu-mask-draft"}, "M",
         "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
@@ -3342,39 +3213,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
-    add_opt(common_arg(
-        {"--spec-replace"}, "TARGET", "DRAFT",
-        "translate the string in TARGET into DRAFT if the draft model and main model are not compatible",
-        [](common_params & params, const std::string & tgt, const std::string & dft) {
-            params.speculative.replacements.push_back({ tgt, dft });
-        }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for K for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_k)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_k = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
-    add_opt(common_arg(
-        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
-        string_format(
-            "KV cache data type for V for the draft model\n"
-            "allowed values: %s\n"
-            "(default: %s)",
-            get_all_kv_cache_types().c_str(),
-            ggml_type_name(params.speculative.cache_type_v)
-        ),
-        [](common_params & params, const std::string & value) {
-            params.speculative.cache_type_v = kv_cache_type_from_str(value);
-        }
-    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));
 
     add_opt(common_arg(
         {"-mv", "--model-vocoder"}, "FNAME",
@@ -3538,97 +3376,5 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
-    add_opt(common_arg(
-        { "--diffusion-steps" }, "N",
-        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
-        [](common_params & params, int value) { params.diffusion.steps = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-visual" },
-        string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
-                      params.diffusion.visual_mode ? "true" : "false"),
-        [](common_params & params) { params.diffusion.visual_mode = true; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-
-    add_opt(common_arg(
-        { "--diffusion-eps" }, "F",
-        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
-        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-algorithm" }, "N",
-        string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)",
-                      params.diffusion.algorithm),
-        [](common_params & params, int value) { params.diffusion.algorithm = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-alg-temp" }, "F",
-        string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
-        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-
-    add_opt(common_arg(
-        { "--diffusion-block-length" }, "N",
-        string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
-        [](common_params & params, int value) { params.diffusion.block_length = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-cfg-scale" }, "F",
-        string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
-        [](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-add-gumbel-noise" }, "F",
-        string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
-        [](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-
-
-    add_opt(
-        common_arg({ "-lr", "--learning-rate" }, "ALPHA",
-                   string_format(
-                       "adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
-                       (double) params.lr.lr0),
-                   [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(
-        common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
-                   string_format(
-                       "(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
-                       (double) params.lr.lr_min),
-                   [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(
-        common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
-                   string_format(
-                       "(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
-                       (double) params.lr.decay_epochs),
-                   [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg(
-                { "-wd", "--weight-decay" }, "WD",
-                string_format(
-                    "adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
-                    (double) params.lr.wd),
-                [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
-                       string_format("fraction of data to use as validation set for training (default: %.2g).",
-                                     (double) params.val_split),
-                       [](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-epochs", "--epochs" }, "N",
-                       string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
-                       [](common_params & params, int epochs) { params.lr.epochs = epochs; })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
-                       [](common_params & params, const std::string & name) {
-                           params.optimizer = common_opt_get_optimizer(name.c_str());
-                           if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
-                               throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
-                           }
-                       })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-
     return ctx_arg;
 }

common/build-info.cpp.in

@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/chat-parser.cpp

@@ -49,21 +49,12 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
-
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
     std::string name = tool_call.contains("name") ? tool_call.at("name") : "";
     std::string id = tool_call.contains("id") ? tool_call.at("id") : "";
-    std::string arguments = "";
-    if (tool_call.contains("arguments")) {
-        if (tool_call.at("arguments").is_object()) {
-            arguments = tool_call.at("arguments").dump();
-        } else {
-            arguments = tool_call.at("arguments");
-        }
-    }
-
+    std::string arguments = tool_call.contains("arguments") ? tool_call.at("arguments") : "";
     return add_tool_call(name, id, arguments);
 }
 
@@ -387,7 +378,3 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
-
-void common_chat_msg_parser::clear_tools() {
-    result_.tool_calls.clear();
-}

common/chat-parser.h

@@ -115,6 +115,4 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
-
-    void clear_tools();
 };

common/chat.cpp

@@ -17,8 +17,6 @@
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
-
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -126,8 +124,6 @@ std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const comm
 typedef minja::chat_template common_chat_template;
 
 struct common_chat_templates {
-    bool add_bos;
-    bool add_eos;
     bool has_explicit_template; // Model had builtin template or template overridde was specified.
     std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
     std::unique_ptr<common_chat_template> template_tool_use;
@@ -144,10 +140,6 @@ struct templates_params {
     bool add_generation_prompt = true;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
-    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) {
@@ -297,7 +289,6 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
-            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -451,8 +442,6 @@ std::string common_chat_format_single(
 
     common_chat_templates_inputs inputs;
     inputs.use_jinja = use_jinja;
-    inputs.add_bos = tmpls->add_bos;
-    inputs.add_eos = tmpls->add_eos;
 
     std::string fmt_past_msg;
     if (!past_msg.empty()) {
@@ -474,12 +463,9 @@ std::string common_chat_format_single(
     return ss.str();
 }
 
-std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja, const std::map<std::string, std::string> & chat_template_kwargs) {
+std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) {
     common_chat_templates_inputs inputs;
     inputs.use_jinja = use_jinja;
-    inputs.add_bos = tmpls->add_bos;
-    inputs.add_eos = tmpls->add_eos;
-    inputs.chat_template_kwargs = chat_template_kwargs;
     auto add_simple_msg = [&](auto role, auto content) {
         common_chat_msg msg;
         msg.role = role;
@@ -555,21 +541,8 @@ common_chat_templates_ptr common_chat_templates_init(
             default_template_src = CHATML_TEMPLATE_SRC;
         }
     }
-
-    // TODO @ngxson : this is a temporary hack to prevent chat template from throwing an error
-    // Ref: https://github.com/ggml-org/llama.cpp/pull/15230#issuecomment-3173959633
-    if (default_template_src.find("<|channel|>") != std::string::npos
-            // search for the error message and patch it
-            && default_template_src.find("in message.content or") != std::string::npos) {
-        string_replace_all(default_template_src,
-            "{%- if \"<|channel|>analysis<|message|>\" in message.content or \"<|channel|>final<|message|>\" in message.content %}",
-            "{%- if false %}");
-    }
-
     std::string token_bos = bos_token_override;
     std::string token_eos = eos_token_override;
-    bool add_bos = false;
-    bool add_eos = false;
     if (model) {
         const auto * vocab = llama_model_get_vocab(model);
         const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
@@ -584,13 +557,9 @@ common_chat_templates_ptr common_chat_templates_init(
         };
         token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
         token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
-        add_bos = llama_vocab_get_add_bos(vocab);
-        add_eos = llama_vocab_get_add_eos(vocab);
     }
     common_chat_templates_ptr tmpls(new common_chat_templates());
     tmpls->has_explicit_template = has_explicit_template;
-    tmpls->add_bos = add_bos;
-    tmpls->add_eos = add_eos;
     try {
         tmpls->template_default = std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos);
     } catch (const std::exception & e) {
@@ -620,8 +589,6 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
         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";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -630,7 +597,6 @@ const char * common_chat_format_name(common_chat_format format) {
 const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
-        case COMMON_REASONING_FORMAT_AUTO:     return "auto";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
         case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
@@ -638,19 +604,6 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     }
 }
 
-common_reasoning_format common_reasoning_format_from_name(const std::string & format) {
-    if (format == "none") {
-        return COMMON_REASONING_FORMAT_NONE;
-    } else if (format == "auto") {
-        return COMMON_REASONING_FORMAT_AUTO;
-    } else if (format == "deepseek") {
-        return COMMON_REASONING_FORMAT_DEEPSEEK;
-    } else if (format == "deepseek-legacy") {
-        return COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY;
-    }
-    throw std::runtime_error("Unknown reasoning format: " + format);
-}
-
 static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
     std::string arguments;
     if (builder.is_partial()) {
@@ -767,23 +720,16 @@ static void foreach_function(const json & tools, const std::function<void(const
 
 static std::string apply(
     const common_chat_template & tmpl,
-    const struct templates_params & inputs,
-    const std::optional<json> & messages_override = std::nullopt,
-    const std::optional<json> & tools_override = std::nullopt,
-    const std::optional<json> & additional_context = std::nullopt)
+    const nlohmann::ordered_json & messages,
+    const nlohmann::ordered_json & tools,
+    bool add_generation_prompt,
+    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
 {
     minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
-    if (tools_override) {
-        tmpl_inputs.tools = *tools_override;
-    } else {
-        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
-    }
-    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
-    tmpl_inputs.extra_context = inputs.extra_context;
-    if (additional_context) {
-        tmpl_inputs.extra_context.merge_patch(*additional_context);
-    }
+    tmpl_inputs.messages = messages;
+    tmpl_inputs.tools = tools;
+    tmpl_inputs.add_generation_prompt = add_generation_prompt;
+    tmpl_inputs.extra_context = extra_context;
     // TODO: add flag to control date/time, if only for testing purposes.
     // tmpl_inputs.now = std::chrono::system_clock::now();
 
@@ -792,10 +738,10 @@ static std::string apply(
     // instead of using `chat_template_options.use_bos_token = false`, since these tokens
     // may be needed inside the template / between messages too.
     auto result = tmpl.apply(tmpl_inputs, tmpl_opts);
-    if (inputs.add_bos && string_starts_with(result, tmpl.bos_token())) {
+    if (string_starts_with(result, tmpl.bos_token())) {
         result = result.substr(tmpl.bos_token().size());
     }
-    if (inputs.add_eos && string_ends_with(result, tmpl.eos_token())) {
+    if (string_ends_with(result, tmpl.eos_token())) {
         result = result.substr(0, result.size() - tmpl.eos_token().size());
     }
     return result;
@@ -882,7 +828,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
         inputs.messages,
         "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
 
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
+    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_GENERIC;
     return data;
 }
@@ -958,7 +904,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.preserved_tokens = {
         "[TOOL_CALLS]",
     };
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
@@ -988,7 +934,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             adjusted_messages.push_back(msg);
         }
     }
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
+    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
         if (!inputs.enable_thinking) {
@@ -1176,7 +1122,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
         {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1241,7 +1187,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs);
+    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
 
     // Hacks to fix the official (broken) prompt.
     // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1333,209 +1279,10 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
         tool_calls_end);
 }
 
-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;
-
-    // These special tokens are required to parse properly, so we include them
-    // even if parse_tool_calls is false.
-    data.preserved_tokens = {
-        "<|channel|>",
-        "<|constrain|>",
-        "<|message|>",
-        "<|start|>",
-        "<|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) {
-            // tool calls can appear in commentary or analysis channels
-            auto channel = builder.add_rule("channel", "\"<|channel|>\" ( \"commentary\" | \"analysis\" )");
-
-            std::vector<std::string> tool_rules_recipient_in_role;
-            std::vector<std::string> tool_rules_recipient_in_channel;
-            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_recipient_in_role.push_back(
-                    builder.add_rule(name + "-call",
-                        "\"" + name + "\"" + channel + " \" <|constrain|>json\"? \"<|message|>\" " +
-                        builder.add_schema(name + "-args", parameters)
-                    )
-                );
-
-                tool_rules_recipient_in_channel.push_back(
-                    builder.add_rule(name + "-call",
-                        "\"" + name + "\"" + " \" <|constrain|>json\"? \"<|message|>\" " +
-                        builder.add_schema(name + "-args", parameters)
-                    )
-                );
-            });
-
-            auto recipient_in_role = builder.add_rule("recipient_in_role",
-                "\"<|start|>assistant\"? \" to=functions.\" ( " +
-                string_join(tool_rules_recipient_in_role, " | ") + " )"
-            );
-
-            auto recipient_in_channel = builder.add_rule("recipient_in_channel",
-                channel + " \" to=functions.\" ( " +
-                string_join(tool_rules_recipient_in_channel, " | ") + " )"
-            );
-
-            builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
-
-            // Trigger on tool calls that appear in the commentary channel
-            data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-                "<\\|channel\\|>(commentary|analysis) to"
-            });
-
-            // Trigger tool calls that appear in the role section, either at the
-            // start or in the middle.
-            data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
-                "^ to"
-            });
-
-            data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-                "<\\|start\\|>assistant to"
-            });
-        });
-    }
-
-    return data;
-}
-static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
-    static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
-    static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
-
-    static const common_regex start_regex("<\\|start\\|>assistant");
-    static const common_regex analysis_regex("<\\|channel\\|>analysis");
-    static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
-    static const common_regex preamble_regex("<\\|channel\\|>commentary");
-    static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
-    static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
-
-    auto consume_end = [&](bool include_end = false) {
-        if (auto res = builder.try_find_literal("<|end|>")) {
-            return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
-        }
-        return builder.consume_rest();
-    };
-
-    auto handle_tool_call = [&](const std::string & name) {
-        if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
-            if (builder.syntax().parse_tool_calls) {
-                if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
-                    throw common_chat_msg_partial_exception("incomplete tool call");
-                }
-            } else if (args->is_partial) {
-                throw common_chat_msg_partial_exception("incomplete tool call");
-            }
-        }
-    };
-
-    auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
-        auto match = regex.search(input, 0, true);
-        if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
-            return match;
-        }
-        return std::nullopt;
-    };
-
-    do {
-        auto header_start_pos = builder.pos();
-        auto content_start = builder.try_find_literal("<|message|>");
-        if (!content_start) {
-            throw common_chat_msg_partial_exception("incomplete header");
-        }
-
-        auto header = content_start->prelude;
-
-        if (auto match = regex_match(tool_call1_regex, header)) {
-            auto group = match->groups[1];
-            auto name = header.substr(group.begin, group.end - group.begin);
-            handle_tool_call(name);
-            continue;
-        }
-
-        if (auto match = regex_match(tool_call2_regex, header)) {
-            auto group = match->groups[2];
-            auto name = header.substr(group.begin, group.end - group.begin);
-            handle_tool_call(name);
-            continue;
-        }
-
-        if (regex_match(analysis_regex, header)) {
-            builder.move_to(header_start_pos);
-            if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
-                builder.add_content(consume_end(true));
-            } else {
-                builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
-            }
-            continue;
-        }
-
-        if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
-            builder.add_content(consume_end());
-            continue;
-        }
-
-        // Possibly a malformed message, attempt to recover by rolling
-        // back to pick up the next <|start|>
-        LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
-        builder.move_to(header_start_pos);
-    } while (builder.try_find_regex(start_regex, std::string::npos, false));
-
-    auto remaining = builder.consume_rest();
-    if (!remaining.empty()) {
-        LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
-    }
-}
-
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
+    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
@@ -1591,7 +1338,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
     // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1718,7 +1465,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
 
-    data.prompt = apply(tmpl, inputs);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     // TODO: if (has_raw_python)
     return data;
 }
@@ -1751,15 +1498,14 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
-    json extra_context = json {
+    json additional_context = {
         {"enable_thinking", inputs.enable_thinking},
     };
-    extra_context.update(inputs.extra_context);
 
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!extra_context["enable_thinking"]) {
+        if (!inputs.enable_thinking) {
             data.prompt += "</think>";
         } else {
             data.thinking_forced_open = true;
@@ -1889,7 +1635,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
         "|<function name=\"([^\"]+)\">"  // match 5 (function name again)
     );
 
-    while (auto res = builder.try_find_regex(open_regex)) {
+    if (auto res = builder.try_find_regex(open_regex)) {
         const auto & block_start = res->groups[1];
         std::string block_end = block_start.empty() ? "" : "```";
 
@@ -1911,6 +1657,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_literal(block_end);
                     builder.consume_spaces();
                 }
+                builder.add_content(builder.consume_rest());
             } else {
                 throw common_chat_msg_partial_exception("failed to parse tool call");
             }
@@ -1935,124 +1682,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_spaces();
                 }
             }
-        }
-    }
-
-    builder.add_content(builder.consume_rest());
-}
-
-static common_chat_params common_chat_params_init_granite(const common_chat_template & tmpl, const struct templates_params & inputs) {
-    common_chat_params data;
-
-    // Pass thinking context for Granite template
-    json additional_context = {
-        {"thinking", inputs.enable_thinking},
-    };
-
-    data.prompt = apply(tmpl, inputs, /* messages_override= */ std::nullopt, /* tools_override= */ std::nullopt, additional_context);
-    data.format = COMMON_CHAT_FORMAT_GRANITE;
-
-    if (string_ends_with(data.prompt, "<think>\n") || string_ends_with(data.prompt, "<think>")) {
-        if (!inputs.enable_thinking) {
-            data.prompt += "</think>";
-        } else {
-            data.thinking_forced_open = true;
-        }
-    }
-
-    if (!inputs.tools.is_null()) {
-        // Granite uses <|tool_call|> followed by JSON list
-        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(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", builder.add_schema(name +
-"-args", {
-                    {"type", "object"},
-                    {"properties", {
-                        {"name", {{"const", name}}},
-                        {"arguments", parameters},
-                    }},
-                    {"required", json::array({"name", "arguments"})},
-                })));
-            });
-
-            auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | "));
-            auto tool_list = builder.add_rule("tool_list", "\"[\" space " + tool_call + " (\",\" space " + tool_call + ")* space \"]\"");
-
-            if (data.thinking_forced_open) {
-                builder.add_rule("root", "\"</think>\" space \"<response>\" space [^<]* \"</response>\" space \"<|tool_call|>\" space " + tool_list);
-            } else {
-                builder.add_rule("root", "\"<|tool_call|>\" space " + tool_list);
-            }
-
-            data.grammar_triggers.push_back({
-                COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
-                "<|tool_call|>"
-            });
-
-            data.preserved_tokens = {
-                "<think>",
-                "</think>",
-                "<response>",
-                "</response>",
-                "<|tool_call|>",
-            };
-        });
-    } else {
-        // Handle thinking tags for non-tool responses
-        if (data.thinking_forced_open && inputs.enable_thinking) {
-            data.grammar_lazy = false;
-            data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-                builder.add_rule("root", "\"</think>\" space \"<response>\" space .* \"</response>\" space");
-            });
-            data.preserved_tokens = {
-                "<think>",
-                "</think>",
-                "<response>",
-                "</response>",
-            };
-        }
-    }
-
-    return data;
-}
-
-static void common_chat_parse_granite(common_chat_msg_parser & builder) {
-    // Parse thinking tags
-    builder.try_parse_reasoning("<think>", "</think>");
-
-    // Parse response tags using regex
-    static const common_regex response_regex("<response>([\\s\\S]*?)</response>");
-    if (auto res = builder.try_find_regex(response_regex)) {
-        // Extract the content between the tags (capture group 1)
-        auto content = builder.str(res->groups[1]);
-        builder.add_content(content);
-        builder.move_to(res->groups[0].end);
-    }
-
-    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("<|tool_call|>"));
-    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.add_tool_calls(tool_calls_data.json)) {
-                builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
-            }
-        } else {
-            builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
+            builder.add_content(builder.consume_rest());
         }
     } else {
         builder.add_content(builder.consume_rest());
@@ -2061,7 +1691,7 @@ static void common_chat_parse_granite(common_chat_msg_parser & builder) {
 
 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);
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     data.grammar_lazy = false;
     if (!inputs.json_schema.is_null()) {
@@ -2092,14 +1722,6 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
-    params.add_bos = tmpls->add_bos;
-    params.add_eos = tmpls->add_eos;
-
-    params.extra_context = json::object();
-    for (auto el : inputs.chat_template_kwargs) {
-        params.extra_context[el.first] = json::parse(el.second);
-    }
-
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }
@@ -2130,21 +1752,11 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_command_r7b(tmpl, params);
     }
 
-    // Granite (IBM) - detects thinking / tools support
-    if (src.find("elif thinking") != std::string::npos && src.find("<|tool_call|>") != std::string::npos) {
-        return common_chat_params_init_granite(tmpl, params);
-    }
-
     // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
     if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_hermes_2_pro(tmpl, params);
     }
 
-    // GPT-OSS
-    if (src.find("<|channel|>") != std::string::npos) {
-        return common_chat_params_init_gpt_oss(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())) {
@@ -2195,7 +1807,6 @@ static common_chat_params common_chat_templates_apply_legacy(
     int alloc_size = 0;
     std::vector<llama_chat_message> chat;
     std::vector<std::string> contents;
-
     for (const auto & msg : inputs.messages) {
         auto content = msg.content;
         for (const auto & part : msg.content_parts) {
@@ -2227,7 +1838,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     if (res < 0) {
         // if the custom "tmpl" is not supported, we throw an error
         // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-        throw std::runtime_error("this custom template is not supported, try using --jinja");
+        throw std::runtime_error("this custom template is not supported");
     }
 
     // if it turns out that our buffer is too small, we resize it
@@ -2297,12 +1908,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_COMMAND_R7B:
             common_chat_parse_command_r7b(builder);
             break;
-        case COMMON_CHAT_FORMAT_GRANITE:
-            common_chat_parse_granite(builder);
-            break;
-        case COMMON_CHAT_FORMAT_GPT_OSS:
-            common_chat_parse_gpt_oss(builder);
-            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }
@@ -2316,14 +1921,10 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            builder.clear_tools();
-            builder.move_to(0);
-            common_chat_parse_content_only(builder);
+            throw std::runtime_error(ex.what());
         }
     }
     auto msg = builder.result();
-    if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
-    }
+    LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
     return msg;
 }

common/chat.h

@@ -7,7 +7,6 @@
 #include <chrono>
 #include <string>
 #include <vector>
-#include <map>
 
 struct common_chat_templates;
 
@@ -109,8 +108,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
-    COMMON_CHAT_FORMAT_GRANITE,
-    COMMON_CHAT_FORMAT_GPT_OSS,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };
@@ -128,9 +125,6 @@ struct common_chat_templates_inputs {
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
-    std::map<std::string, std::string> chat_template_kwargs;
-    bool add_bos = false;
-    bool add_eos = false;
 };
 
 struct common_chat_params {
@@ -187,12 +181,10 @@ std::string common_chat_format_single(
 // Returns an example of formatted chat
 std::string common_chat_format_example(
     const struct common_chat_templates * tmpls,
-    bool use_jinja,
-    const std::map<std::string, std::string> & chat_template_kwargs);
+    bool use_jinja);
 
 const char*               common_chat_format_name(common_chat_format format);
 const char*               common_reasoning_format_name(common_reasoning_format format);
-common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
 common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);

common/cmake/build-info-gen-cpp.cmake

@@ -0,0 +1,24 @@
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
+
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMMIT ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_COMPILER ${CMAKE_MATCH_1})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
+    set(OLD_TARGET ${CMAKE_MATCH_1})
+    if (
+        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
+        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
+        NOT OLD_TARGET   STREQUAL BUILD_TARGET
+    )
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+    endif()
+else()
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+endif()

common/common.cpp

@@ -41,7 +41,6 @@
 #endif
 #include <locale>
 #include <windows.h>
-#include <string.h>
 #include <fcntl.h>
 #include <io.h>
 #else
@@ -449,15 +448,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
     return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
 }
-
-bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
-    bool has_suffix = string_ends_with(str, suffix);
-    if (has_suffix) {
-        str = str.substr(0, str.size() - suffix.size());
-    }
-    return has_suffix;
-}
-
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
     if (!str.empty() && !stop.empty()) {
         const char text_last_char = str.back();
@@ -558,6 +548,13 @@ 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);
     }
@@ -582,6 +579,13 @@ 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])
@@ -702,17 +706,11 @@ bool fs_validate_filename(const std::string & filename) {
         // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
-
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
 #endif
 
         filename_utf32 = converter.from_bytes(filename);
@@ -769,9 +767,6 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
-#include <iostream>
-
-
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -789,16 +784,9 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // process path from front to back, procedurally creating directories
     while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) {
         const std::wstring subpath = wpath.substr(0, pos_slash);
+        const wchar_t * test = subpath.c_str();
 
-        pos_slash += 1;
-
-        // skip the drive letter, in some systems it can return an access denied error
-        if (subpath.length() == 2 && subpath[1] == ':') {
-            continue;
-        }
-
-        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
-
+        const bool success = CreateDirectoryW(test, NULL);
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -812,6 +800,8 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
+
+        pos_slash += 1;
     }
 
     return true;
@@ -907,6 +897,34 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -948,35 +966,6 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
-    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1001,19 +990,13 @@ struct common_init_result common_init_from_params(common_params & params) {
         params.sampling.ignore_eos = false;
     }
 
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
     if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+            if (llama_vocab_is_eog(vocab, i)) {
+                LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+                params.sampling.logit_bias.push_back({i, -INFINITY});
+            }
+        }
     }
 
     if (params.sampling.penalty_last_n == -1) {
@@ -1109,7 +1092,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
-    mparams.use_extra_bufts = !params.no_extra_bufts;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
@@ -1152,6 +1134,7 @@ 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.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
@@ -1159,7 +1142,11 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.no_perf           = params.no_perf;
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
-    cparams.kv_unified        = params.kv_unified;
+
+    if (params.reranking) {
+        cparams.embeddings    = true;
+        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
+    }
 
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
@@ -1293,9 +1280,6 @@ std::vector<llama_token> common_tokenize(
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
     n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
-    if (n_tokens == std::numeric_limits<int32_t>::min()) {
-        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
-    }
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
@@ -1551,56 +1535,3 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
 
     return result;
 }
-
-ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
-    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
-    const lr_opt &            d      = *(lr_opt *) userdata;
-    result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
-    result.sgd.wd = result.adamw.wd = d.wd;
-    return result;
-}
-
-// TODO make all command line args case-insensitive
-static inline bool eq_case_insensitive(char const* a, char const* b) {
-    return !
-#if defined(_MSC_VER)
-        _stricmp
-#else
-        strcasecmp
-#endif // defined(_MSC_VER)
-        (a, b);
-}
-
-enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
-    if (eq_case_insensitive("adamw", n)) {
-        return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
-    }
-    if (eq_case_insensitive("sgd", n)) {
-        return GGML_OPT_OPTIMIZER_TYPE_SGD;
-    }
-    return GGML_OPT_OPTIMIZER_TYPE_COUNT;
-}
-
-// TODO simplify to use just log and exp
-static float const k_log_2 = std::log(2.f);
-
-void lr_opt::init() {
-    if (lr_min > 0 && lr_min < lr0) {
-        float nhalf = std::log(lr0 / lr_min) / k_log_2;
-        float e     = epochs;
-        if (decay_epochs > 0 && decay_epochs < e) {
-            e = decay_epochs;
-        } else {
-            decay_epochs = e;
-        }
-        scale_epoch = nhalf / e;
-    }
-}
-
-float lr_opt::get_lr(float epoch) const {
-    float r = lr_min <= 0 ? lr0 :
-        epoch >= decay_epochs ? lr_min :
-        lr0 * std::pow(0.5f, epoch * scale_epoch);
-    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
-    return r;
-}

common/common.h

@@ -2,17 +2,13 @@
 
 #pragma once
 
+#include "llama-cpp.h"
+
 #include <set>
-#include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
-#include <map>
 #include <sstream>
-#include <cmath>
-
-#include "ggml-opt.h"
-#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -84,8 +80,6 @@ enum llama_example {
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
-    LLAMA_EXAMPLE_DIFFUSION,
-    LLAMA_EXAMPLE_FINETUNE,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -182,8 +176,7 @@ struct common_params_sampling {
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
-    std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
-    std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
+    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
 
     // print the parameters into a string
     std::string print() const;
@@ -205,11 +198,6 @@ struct common_params_speculative {
     int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
-    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
-
-    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
-    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
@@ -225,50 +213,12 @@ struct common_params_vocoder {
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
 
-struct common_params_diffusion {
-    int32_t steps         = 128;
-    bool    visual_mode   = false;
-
-    float   eps           = 0;        // epsilon for timesteps
-    int32_t block_length  = 0;        // block length for generation
-
-    int32_t algorithm     = 4;        // default algorithm: low-confidence
-    float   alg_temp      = 0.0f;     // algorithm temperature
-
-    float   cfg_scale     = 0;        // classifier-free guidance scale
-    bool    add_gumbel_noise = false; // add gumbel noise to the logits if temp > 0.0
-};
-
-// reasoning API response format (not to be confused as chat template's reasoning format)
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
     COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
-    // do not extend this enum unless you absolutely have to
-    // in most cases, use COMMON_REASONING_FORMAT_AUTO
-    // see: https://github.com/ggml-org/llama.cpp/pull/15408
 };
 
-
-struct lr_opt {
-    float    lr0          = 1e-5; // learning rate at first epoch
-    float    lr_min       = -1;
-    float    decay_epochs = -1;   // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
-    float    scale_epoch  = 0;
-    float    wd           = 0;
-    unsigned epochs       = 2;
-
-    unsigned epoch; // set by optimizer outer (epochs) loop
-    // learning rate decay - constant LR per epoch only for now
-    float get_lr(float e) const;
-    float get_lr() const { return get_lr(epoch); }
-    // must call after arg parse, before get_lr
-    void init();
-};
-
-struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
-
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -288,6 +238,7 @@ struct common_params {
     float   yarn_beta_fast        = 32.0f; // 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
@@ -313,7 +264,6 @@ struct common_params {
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
-    struct common_params_diffusion   diffusion;
 
     struct common_params_model model;
 
@@ -374,9 +324,8 @@ struct common_params {
     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 infinite text generation
+    bool ctx_shift         = true;  // context shift on inifinite 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
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads
@@ -387,7 +336,6 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
-    bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -402,33 +350,26 @@ struct common_params {
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
-    // finetune
-    struct lr_opt lr;
-    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
-    float val_split = 0.05f; // fraction of the data used for the validation set
-
     // embedding
     bool embedding         = false; // get only sentence embedding
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    std::string cls_sep    = "\t";  // separator of classification sequences
+    bool reranking         = false; // enable reranking support on server
 
     // server params
-    int32_t port              = 8080;         // server listens on this network port
-    int32_t timeout_read      = 600;          // http read timeout in seconds
-    int32_t timeout_write     = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
-    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
+    int32_t port           = 8080;         // server listens on this network port
+    int32_t timeout_read   = 600;          // http read timeout in seconds
+    int32_t timeout_write  = timeout_read; // http write timeout in seconds
+    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
-    std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
@@ -437,8 +378,6 @@ struct common_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
-    std::map<std::string, std::string> default_template_kwargs;
-
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
     bool endpoint_slots   = false;
@@ -473,12 +412,10 @@ struct common_params {
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
-    int8_t  imat_dat    =  0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)
 
-    bool process_output  = false; // collect data for the output tensor
-    bool compute_ppl     = true;  // whether to compute perplexity
-    bool show_statistics = false; // show imatrix statistics per tensor
-    bool parse_special   = false; // whether to parse special tokens during imatrix tokenization
+    bool process_output = false; // collect data for the output tensor
+    bool compute_ppl    = true;  // whether to compute perplexity
+    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;
@@ -578,7 +515,6 @@ static bool string_starts_with(const std::string & str,
 
 // While we wait for C++20's std::string::ends_with...
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
-bool string_remove_suffix(std::string & str, const std::string_view & suffix);
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
@@ -735,6 +671,3 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 //
 
 ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
-
-// "adamw" or "sgd" (case insensitive)
-enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);

common/json-schema-to-grammar.cpp

@@ -41,6 +41,49 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
+/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
+class string_view {
+    const std::string & _str;
+    const size_t _start;
+    const size_t _end;
+public:
+    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
+
+    size_t size() const {
+        return _end - _start;
+    }
+
+    size_t length() const {
+        return size();
+    }
+
+    operator std::string() const {
+        return str();
+    }
+
+    std::string str() const {
+        return _str.substr(_start, _end - _start);
+    }
+
+    string_view substr(size_t pos, size_t len = std::string::npos) const {
+        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
+    }
+
+    char operator[](size_t pos) const {
+        auto index = _start + pos;
+        if (index >= _end) {
+            throw std::out_of_range("string_view index out of range");
+        }
+        return _str[_start + pos];
+    }
+
+    bool operator==(const string_view & other) const {
+        std::string this_str = *this;
+        std::string other_str = other;
+        return this_str == other_str;
+    }
+};
+
 static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
     auto has_min = min_value != std::numeric_limits<int>::min();
     auto has_max = max_value != std::numeric_limits<int>::max();
@@ -69,14 +112,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         }
         out << "}";
     };
-    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
-        [&](const std::string_view & from, const std::string_view & to) {
+    std::function<void(const string_view &, const string_view &)> uniform_range =
+        [&](const string_view & from, const string_view & to) {
             size_t i = 0;
             while (i < from.length() && i < to.length() && from[i] == to[i]) {
                 i++;
             }
             if (i > 0) {
-                out << "\"" << from.substr(0, i) << "\"";
+                out << "\"" << from.substr(0, i).str() << "\"";
             }
             if (i < from.length() && i < to.length()) {
                 if (i > 0) {

common/speculative.cpp

@@ -1,39 +1,30 @@
 #include "speculative.h"
 
-#include "ggml.h"
-#include "llama.h"
 #include "log.h"
 #include "common.h"
 #include "sampling.h"
 
 #include <cstring>
 #include <algorithm>
-#include <map>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
 struct common_speculative {
-    struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
-    struct llama_context * ctx_dft;
+    struct llama_context * ctx;
     struct common_sampler * smpl;
 
     llama_batch batch;
-    llama_tokens prompt_dft;
-    bool vocab_dft_compatible = true; // whether retokenization is needed
-    std::map<std::string, std::string> tgt_dft_replacements = {};
+    llama_tokens prompt;
 };
 
 struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
         struct llama_context * ctx_dft) {
     auto * result = new common_speculative {
-        /* .ctx_tgt    = */ ctx_tgt,
-        /* .ctx_dft    = */ ctx_dft,
-        /* .smpl       = */ nullptr,
-        /* .batch      = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
-        /* .prompt_dft = */ {},
-        /* .vocab_dft_compatible = */ false,
+        /* .ctx    = */ ctx_dft,
+        /* .smpl   = */ nullptr,
+        /* .batch  = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .prompt = */ {},
     };
 
     // TODO: optimize or pass from outside?
@@ -68,9 +59,6 @@ struct common_speculative * common_speculative_init(
     }
 #endif
 
-    result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft);
-    LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible);
-
     return result;
 }
 
@@ -87,8 +75,8 @@ void common_speculative_free(struct common_speculative * spec) {
 }
 
 bool common_speculative_are_compatible(
-    const struct llama_context * ctx_tgt,
-    const struct llama_context * ctx_dft) {
+        const struct llama_context * ctx_tgt,
+        const struct llama_context * ctx_dft) {
     const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
     const struct llama_model * model_dft = llama_get_model(ctx_dft);
 
@@ -102,32 +90,31 @@ bool common_speculative_are_compatible(
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
-        LOG_DBG("%s: draft model vocab type must match target model to use speculation but ", __func__);
-        LOG_DBG("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
+        LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
+                     "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
         return false;
     }
 
-    if (
-        llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
+    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
         llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
         llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
-    ) {
-        LOG_DBG("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
+        LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
         return false;
     }
 
     {
         const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
         const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
-        const int vocab_diff  = n_vocab_tgt > n_vocab_dft
-            ? n_vocab_tgt - n_vocab_dft
-            : n_vocab_dft - n_vocab_tgt;
+
+        const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
 
         if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
-            LOG_DBG("%s: draft model vocab must closely match target model to use speculation but ", __func__);
-            LOG_DBG("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
-                    n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
+                         "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    __func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
             return false;
         }
 
@@ -135,8 +122,8 @@ bool common_speculative_are_compatible(
             const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
             const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
-                LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__);
-                LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i,
+                LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
+                             "token %d content differs - target '%s', draft '%s'\n", __func__, i,
                         common_token_to_piece(ctx_tgt, i).c_str(),
                         common_token_to_piece(ctx_dft, i).c_str());
                 return false;
@@ -147,93 +134,32 @@ bool common_speculative_are_compatible(
     return true;
 }
 
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest) {
-    spec->tgt_dft_replacements[source] = dest;
-}
-
-static std::string replace_to_dft(
-        struct common_speculative * spec,
-        const std::string& input) {
-    std::string result = input;
-    for (const auto & pair : spec->tgt_dft_replacements) {
-        size_t pos = result.find(pair.first);
-        while (pos != std::string::npos) {
-            result.replace(pos, pair.first.length(), pair.second);
-            pos = result.find(pair.first, pos + pair.second.length());
-        }
-    }
-    return result;
-}
-
-static std::string replace_to_tgt(
-        struct common_speculative * spec,
-        const std::string& input) {
-    std::string result = input;
-    for (const auto& pair : spec->tgt_dft_replacements) {
-        size_t pos = result.find(pair.second);
-        while (pos != std::string::npos) {
-            result.replace(pos, pair.second.length(), pair.first);
-            pos = result.find(pair.second, pos + pair.first.length());
-        }
-    }
-    return result;
-}
-
-
 llama_tokens common_speculative_gen_draft(
         struct common_speculative * spec,
         struct common_speculative_params params,
-        const llama_tokens & prompt_tgt_main_model, // specified in target model vocab
+        const llama_tokens & prompt_tgt,
         llama_token id_last) {
     auto & batch  = spec->batch;
-    auto & ctx_tgt = spec->ctx_tgt;
-    auto & ctx_dft = spec->ctx_dft;
+    auto & ctx    = spec->ctx;
     auto & smpl   = spec->smpl;
-    auto & prompt_dft = spec->prompt_dft;
+    auto & prompt = spec->prompt;
 
-    auto * mem_dft = llama_get_memory(ctx_dft);
+    auto * mem = llama_get_memory(ctx);
 
     int reuse_i = 0;
     int reuse_n = 0;
 
-    const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft;
-
-    llama_tokens prompt_tgt_draft_model;
-    if (!spec->vocab_dft_compatible) {
-        std::string text;
-        text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true);
-        text = replace_to_dft(spec, text);
-        LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
-        prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true);
-
-        // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
-        const auto * model_tgt = llama_get_model(ctx_tgt);
-        const auto * vocab_tgt = llama_model_get_vocab(model_tgt);
-
-        int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
-        GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
-        text.resize(-n_chars);
-        llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
-        text = replace_to_dft(spec, text);
-
-        LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
-        id_last = common_tokenize(ctx_dft, text, false, true)[0];
-    }
-    // prompt_tgt's tokens will always be compatible with ctx_dft
-    const llama_tokens &prompt_tgt =
-        spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model;
+    const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
 
     const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
 
     // reuse as much as possible from the old draft context
     // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
-    for (int i = 0; i < (int) prompt_dft.size(); ++i) {
+    for (int i = 0; i < (int) prompt.size(); ++i) {
         int cur = 0;
         while (i_start + cur < (int) prompt_tgt.size() &&
-               i       + cur < (int) prompt_dft.size() &&
-               prompt_tgt[i_start + cur] == prompt_dft[i + cur]) {
+               i       + cur < (int) prompt.size() &&
+               prompt_tgt[i_start + cur] == prompt[i + cur]) {
             cur++;
         }
 
@@ -243,20 +169,21 @@ llama_tokens common_speculative_gen_draft(
         }
     }
 
-    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
 
     llama_tokens result;
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_memory_clear(mem_dft, false);
-        prompt_dft.clear();
+        llama_memory_clear(mem, false);
+
+        prompt.clear();
     } else {
         // this happens when a previous draft has been discarded (for example, due to being too small), but the
         // target model agreed with it. in this case, we simply pass back the previous results to save compute
-        if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
-            for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
-                result.push_back(prompt_dft[i]);
+        if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
+            for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
+                result.push_back(prompt[i]);
 
                 if (params.n_draft <= (int) result.size()) {
                     break;
@@ -267,15 +194,16 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
-            llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem, 0, 0, reuse_i);
+            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
 
-            prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
+            prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
-        if (reuse_n < (int) prompt_dft.size()) {
-            llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
-            prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+        if (reuse_n < (int) prompt.size()) {
+            llama_memory_seq_rm (mem, 0, reuse_n, -1);
+
+            prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }
     }
 
@@ -286,28 +214,28 @@ llama_tokens common_speculative_gen_draft(
         //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
         common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
 
-        prompt_dft.push_back(prompt_tgt[i]);
+        prompt.push_back(prompt_tgt[i]);
     }
 
     // we should rarely end-up here during normal decoding
     if (batch.n_tokens > 0) {
         //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
 
-        llama_decode(ctx_dft, batch);
+        llama_decode(ctx, batch);
     }
 
-    const llama_pos n_past = prompt_dft.size();
+    const llama_pos n_past = prompt.size();
 
     LOG_DBG("%s: n_past = %d\n", __func__, n_past);
 
     common_batch_clear(batch);
     common_batch_add  (batch, id_last, n_past, { 0 }, true);
 
-    prompt_dft.push_back(id_last);
+    prompt.push_back(id_last);
 
-    LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
+    //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
 
-    llama_decode(ctx_dft, batch);
+    llama_decode(ctx, batch);
 
     common_sampler_reset(smpl);
 
@@ -315,13 +243,13 @@ llama_tokens common_speculative_gen_draft(
     for (int i = 0; i < params.n_draft; ++i) {
         common_batch_clear(batch);
 
-        common_sampler_sample(smpl, ctx_dft, 0, true);
+        common_sampler_sample(smpl, ctx, 0, true);
 
         const auto * cur_p = common_sampler_get_candidates(smpl);
 
         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",
-                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
         }
 
         // add drafted token for each sequence
@@ -343,19 +271,10 @@ llama_tokens common_speculative_gen_draft(
         common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
 
         // evaluate the drafted tokens on the draft model
-        llama_decode(ctx_dft, batch);
+        llama_decode(ctx, batch);
 
-        prompt_dft.push_back(id);
+        prompt.push_back(id);
     }
 
-    if (!spec->vocab_dft_compatible) {
-        std::string detokenized = common_detokenize(ctx_dft, result, true);
-        detokenized = replace_to_tgt(spec, detokenized);
-        LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
-        result = common_tokenize(ctx_tgt, detokenized, false, true);
-        if (result.size() > (size_t)params.n_draft) {
-            result.resize(params.n_draft);
-        }
-    }
     return result;
 }

common/speculative.h

@@ -12,10 +12,7 @@ struct common_speculative_params {
     float p_min = 0.75f; // min probability required to accept a token in the draft
 };
 
-struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
-        struct llama_context * ctx_dft
-);
+struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);
 
 void common_speculative_free(struct common_speculative * spec);
 
@@ -23,10 +20,6 @@ bool common_speculative_are_compatible(
         const struct llama_context * ctx_tgt,
         const struct llama_context * ctx_dft);
 
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest);
-
 // sample up to n_draft tokens and add them to the batch using the draft model
 llama_tokens common_speculative_gen_draft(
                struct common_speculative * spec,

convert_hf_to_gguf_update.py

@@ -7,6 +7,7 @@ import pathlib
 import re
 
 import requests
+import sys
 import json
 import shutil
 import argparse
@@ -59,10 +60,6 @@ parser.add_argument(
     "--full", action="store_true",
     help="download full list of models - make sure you have access to all of them",
 )
-parser.add_argument(
-    "--check-missing", action="store_true",
-    help="only check for missing pre-tokenizer hashes",
-)
 parser.add_argument(
     "hf_token",
     help="optional HF token",
@@ -72,11 +69,8 @@ args = parser.parse_args()
 hf_token = args.hf_token if args.hf_token is not None else hf_token
 
 if hf_token is None:
-    logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
-
-if args.check_missing and args.full:
-    logger.warning("Downloading full list of models requested, ignoring --check-missing!")
-    args.check_missing = False
+    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
+    sys.exit(1)
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
@@ -134,11 +128,6 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
-    {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
-    {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
-    {"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", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -147,22 +136,12 @@ pre_computed_hashes = [
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
-    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
     {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
-    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
-    {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
-    # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
-    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
-    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
-    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
-    {"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"},
 ]
 
 
 def download_file_with_auth(url, token, save_path):
-    headers = {"Authorization": f"Bearer {token}"} if token else None
+    headers = {"Authorization": f"Bearer {token}"}
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
@@ -232,19 +211,18 @@ if not args.full:
     all_models = models.copy()
     models = [model for model in all_models if model["name"] not in existing_models]
 
-if not args.check_missing:
-    logging.info(f"Downloading {len(models)} models...")
-    for model in models:
-        try:
-            download_model(model)
-        except Exception as e:
-            logger.error(f"Failed to download model {model['name']}. Error: {e}")
+logging.info(f"Downloading {len(models)} models...")
+for model in models:
+    try:
+        download_model(model)
+    except Exception as e:
+        logger.error(f"Failed to download model {model['name']}. Error: {e}")
 
 
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in [*pre_computed_hashes, *all_models]:
+for model in [*all_models, *pre_computed_hashes]:
     name = model["name"]
     tokt = model["tokt"]
     chkhsh = model.get("chkhsh")
@@ -252,6 +230,11 @@ for model in [*pre_computed_hashes, *all_models]:
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
 
+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
     # create the tokenizer
     if chkhsh is not None:
         # if the model has a pre-computed hash, use it
@@ -261,19 +244,15 @@ for model in [*pre_computed_hashes, *all_models]:
         chkhsh = existing_models[name]
     else:
         # otherwise, compute the hash of the tokenizer
-
-        # Fail if the tokenizer folder with config does not exist or there are other download issues previously
-        if not os.path.isfile(f"models/tokenizers/{name}/tokenizer_config.json"):
-            raise OSError(f"Config for tokenizer {name} not found. The model may not exist or is not accessible with the provided token.")
-
         try:
             logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
             if name == "t5":
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
             else:
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-        except Exception as e:
-            raise OSError(f"Error loading tokenizer for model {name}.") from e
+        except OSError as e:
+            logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+            continue  # Skip to the next model if the tokenizer can't be loaded
 
         chktok = tokenizer.encode(CHK_TXT)
         chkhsh = sha256(str(chktok).encode()).hexdigest()

convert_lora_to_gguf.py

@@ -340,7 +340,7 @@ if __name__ == '__main__':
             sys.exit(1)
     else:
         logger.info(f"Loading base model: {dir_base_model.name}")
-        hparams = ModelBase.load_hparams(dir_base_model, False)
+        hparams = ModelBase.load_hparams(dir_base_model)
 
     with torch.inference_mode():
         try:

docs/backend/CANN.md

@@ -310,7 +310,5 @@ Specifies the memory pool management strategy:
 
 Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
 
-### GGML_CANN_WEIGHT_NZ
-
-Converting the matmul weight format from ND to NZ can significantly improve performance on the 310I DUO NPU.
-
+## TODO
+- Support more models and data types.

docs/backend/SYCL.md

@@ -757,7 +757,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
+| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

docs/build-s390x.md

@@ -1,295 +0,0 @@
-> [!IMPORTANT]
-> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
-
-# Build llama.cpp locally (for s390x)
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
-
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
-
-**To get the code:**
-
-```bash
-git clone https://github.com/ggml-org/llama.cpp
-cd llama.cpp
-```
-
-## CPU Build with BLAS
-
-Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
-
-```bash
-cmake -S . -B build             \
-    -DCMAKE_BUILD_TYPE=Release  \
-    -DGGML_BLAS=ON              \
-    -DGGML_BLAS_VENDOR=OpenBLAS
-
-cmake --build build --config Release -j $(nproc)
-```
-
-**Notes**:
-
--   For faster repeated compilation, install [ccache](https://ccache.dev/)
--   By default, VXE/VXE2 is enabled. To disable it (not recommended):
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DGGML_VXE=OFF
-
-    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
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Debug    \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS
-    cmake --build build --config Debug -j $(nproc)
-    ```
-
--   For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
-
-    ```bash
-    cmake -S . -B build             \
-        -DCMAKE_BUILD_TYPE=Release  \
-        -DGGML_BLAS=ON              \
-        -DGGML_BLAS_VENDOR=OpenBLAS \
-        -DBUILD_SHARED_LIBS=OFF
-
-    cmake --build build --config Release -j $(nproc)
-    ```
-
-## IBM zDNN Accelerator
-
-This provides acceleration using the IBM zAIU co-processor located in the Telum I and Telum II processors. Make sure to have the [IBM zDNN library](https://github.com/IBM/zDNN) installed.
-
-#### Compile from source from IBM
-
-You may find the official build instructions here: [Building and Installing zDNN](https://github.com/IBM/zDNN?tab=readme-ov-file#building-and-installing-zdnn)
-
-### Compilation
-
-```bash
-cmake -S . -B build             \
-    -DCMAKE_BUILD_TYPE=Release  \
-    -DGGML_ZDNN=ON
-cmake --build build --config Release -j$(nproc)
-```
-
-## Getting GGUF Models
-
-All models need to be converted to Big-Endian. You can achieve this in three cases:
-
-1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
-
-    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
-
-    You can find popular models pre-converted and verified at [s390x Verified Models](https://huggingface.co/collections/taronaeo/s390x-verified-models-672765393af438d0ccb72a08) or [s390x Runnable Models](https://huggingface.co/collections/taronaeo/s390x-runnable-models-686e951824198df12416017e).
-
-    These models have already been converted from `safetensors` to `GGUF` Big-Endian and their respective tokenizers verified to run correctly on IBM z15 and later system.
-
-2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
-
-    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
-
-    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
-
-    Ensure that you have installed the required packages in advance
-
-    ```bash
-    pip3 install -r requirements.txt
-    ```
-
-    Convert the `safetensors` model to `GGUF`
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile model-name-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        model-directory/
-    ```
-
-    For example,
-
-    ```bash
-    python3 convert_hf_to_gguf.py \
-        --outfile granite-3.3-2b-instruct-be.f16.gguf \
-        --outtype f16 \
-        --bigendian \
-        granite-3.3-2b-instruct/
-    ```
-
-3. **Convert existing GGUF Little-Endian model to Big-Endian**
-
-    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
-
-    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B GGUF](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
-
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
-    ```
-
-    For example,
-
-    ```bash
-    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
-    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
-    ```
-
-    **Notes:**
-
-    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
-
-## IBM Accelerators
-
-### 1. SIMD Acceleration
-
-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)
-
-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
-
-_Only available with IBM z17 / LinuxONE 5 or later system. No support currently available._
-
-## Performance Tuning
-
-### 1. Virtualization Setup
-
-It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
-
-Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
-
-### 2. IFL (Core) Count
-
-It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
-
-Note: IFL count does not equate to vCPU count.
-
-### 3. SMT vs NOSMT (Simultaneous Multithreading)
-
-It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
-
-### 4. BLAS vs NOBLAS
-
-IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
-
-## Frequently Asked Questions (FAQ)
-
-1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
-
-    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
-
-    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
-
-2. I'm getting extremely poor performance when running inference on a model
-
-    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
-
-3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
-
-    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
-
-4. Failing to install the `sentencepiece` package using GCC 15+
-
-    Answer: The `sentencepiece` team are aware of this as seen in [this issue](https://github.com/google/sentencepiece/issues/1108).
-
-    As a temporary workaround, please run the installation command with the following environment variables.
-
-    ```bash
-    export CXXFLAGS="-include cstdint"
-    ```
-
-    For example,
-
-    ```bash
-    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**
-
-    Please file an issue in llama.cpp and ensure that the title contains "s390x".
-
-2. **Other Questions**
-
-    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
-
-## Appendix A: Hardware Support Matrix
-
-|          | Support | Minimum Compiler Version |
-| -------- | ------- | ------------------------ |
-| IBM z15  | ✅      |                          |
-| IBM z16  | ✅      |                          |
-| IBM z17  | ✅      | GCC 15.1.0               |
-| IBM zDNN | ✅      |                          |
-
--   ✅ - supported and verified to run as intended
--   🚫 - unsupported, we are unlikely able to provide support
-
-## Appendix B: SIMD Support Matrix
-
-|            | VX/VXE/VXE2 | NNPA | 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 Aug 22, 2025.

docs/build.md

@@ -1,6 +1,6 @@
 # Build llama.cpp locally
 
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
 
 The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
 
@@ -68,9 +68,6 @@ cmake --build build --config Release
       cmake --build build-x64-windows-llvm-release
       ```
 - Curl usage is enabled by default and can be turned off with `-DLLAMA_CURL=OFF`. Otherwise you need to install development libraries for libcurl.
-  - **Debian / Ubuntu:** `sudo apt-get install libcurl4-openssl-dev`  # (or `libcurl4-gnutls-dev` if you prefer GnuTLS)
-  - **Fedora / RHEL / Rocky / Alma:** `sudo dnf install libcurl-devel`
-  - **Arch / Manjaro:** `sudo pacman -S curl`  # includes libcurl headers
 
 ## BLAS Build
 
@@ -197,12 +194,13 @@ 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. 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.                                                                                                                                                                                           |
+| 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.                                                                                                  |
 
 ## MUSA
 
@@ -307,8 +305,9 @@ On Linux it is possible to use unified memory architecture (UMA) to share main m
 
 ## Vulkan
 
-### For Windows Users:
-**w64devkit**
+**Windows**
+
+### w64devkit
 
 Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
@@ -335,7 +334,7 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-**Git Bash MINGW64**
+### Git Bash MINGW64
 
 Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
 
@@ -358,8 +357,7 @@ Now you can load the model in conversation mode using `Vulkan`
 build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
 ```
 
-**MSYS2**
-
+### MSYS2
 Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
 ```sh
 pacman -S git \
@@ -375,9 +373,9 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-### For Docker users:
+**With docker**:
 
-You don't need to install the Vulkan SDK. It will be installed inside the container.
+You don't need to install Vulkan SDK. It will be installed inside the container.
 
 ```sh
 # Build the image
@@ -387,29 +385,32 @@ docker build -t llama-cpp-vulkan --target light -f .devops/vulkan.Dockerfile .
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
-### For Linux users:
+**Without docker**:
 
-First, follow the official LunarG instructions for the installation and setup of the Vulkan SDK in the [Getting Started with the Linux Tarball Vulkan SDK](https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html) guide.
+Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
 
-> [!IMPORTANT]
-> After completing the first step, ensure that you have used the `source` command on the `setup_env.sh` file inside of the Vulkan SDK in your current terminal session. Otherwise, the build won't work. Additionally, if you close out of your terminal, you must perform this step again if you intend to perform a build. However, there are ways to make this persistent. Refer to the Vulkan SDK guide linked in the first step for more information about any of this.
+For example, on Ubuntu 22.04 (jammy), use the command below:
 
-Second, after verifying that you have followed all of the SDK installation/setup steps, use this command to make sure before proceeding:
 ```bash
+wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
+wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+apt update -y
+apt-get install -y vulkan-sdk
+# To verify the installation, use the command below:
 vulkaninfo
 ```
 
-Then, assuming you have `cd` into your llama.cpp folder and there are no errors with running `vulkaninfo`, you can proceed to build llama.cpp using the CMake commands below:
+Alternatively your package manager might be able to provide the appropriate libraries.
+For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
+For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
+
+Then, build llama.cpp using the cmake command below:
+
 ```bash
 cmake -B build -DGGML_VULKAN=1
 cmake --build build --config Release
-```
-
-Finally, after finishing your build, you should be able to do something like this:
-```bash
-# Test the output binary
-# "-ngl 99" should offload all of the layers to GPU for most (if not all) models.
-./build/bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -ngl 99
+# Test the output binary (with "-ngl 33" to offload all layers to GPU)
+./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
 
 # You should see in the output, ggml_vulkan detected your GPU. For example:
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
@@ -556,27 +557,6 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
-## WebGPU [In Progress]
-
-The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
-
-In the llama.cpp directory, build with CMake:
-
-```
-cmake -B build -DGGML_WEBGPU=ON
-cmake --build build --config Release
-```
-
-### Browser Support
-
-WebGPU allows cross-platform access to the GPU from supported browsers. We utilize [Emscripten](https://emscripten.org/) to compile ggml's WebGPU backend to WebAssembly. Emscripten does not officially support WebGPU bindings yet, but Dawn currently maintains its own WebGPU bindings called emdawnwebgpu.
-
-Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/src/emdawnwebgpu/) to download or build the emdawnwebgpu package (Note that it might be safer to build the emdawbwebgpu package locally, so that it stays in sync with the version of Dawn you have installed above). When building using CMake, the path to the emdawnwebgpu port file needs to be set with the flag `EMDAWNWEBGPU_DIR`.
-
-## IBM Z & LinuxONE
-
-To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
-
 ## Notes about GPU-accelerated backends
 
 The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.

docs/development/HOWTO-add-model.md

@@ -23,19 +23,11 @@ The convert script reads the model configuration, tokenizer, tensor names+data a
 
 The required steps to implement for an HF model are:
 
-1. Define the model `ModelBase.register` annotation in a new `TextModel` or `MmprojModel` subclass, example:
+1. Define the model `Model.register` annotation in a new `Model` subclass, example:
 
 ```python
-@ModelBase.register("MyModelForCausalLM")
-class MyModel(TextModel):
-    model_arch = gguf.MODEL_ARCH.MYMODEL
-```
-
-or
-
-```python
-@ModelBase.register("MyModelForConditionalGeneration")
-class MyModel(MmprojModel):
+@Model.register("MyModelForCausalLM")
+class MyModel(Model):
     model_arch = gguf.MODEL_ARCH.MYMODEL
 ```
 
@@ -83,31 +75,28 @@ block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
 `transformer.blocks.{bid}.norm_1` will be mapped to `blk.{bid}.attn_norm` in GGUF.
 
 Depending on the model configuration, tokenizer, code and tensors layout, you will have to override:
-- `TextModel#set_gguf_parameters`
-- `MmprojModel#set_gguf_parameters`
-- `ModelBase#set_vocab`
-- `ModelBase#modify_tensors`
+- `Model#set_gguf_parameters`
+- `Model#set_vocab`
+- `Model#write_tensors`
 
 NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the convention and several tools like `quantize` expect this to proceed the weights.
 
 ### 2. Define the model architecture in `llama.cpp`
 
-The model params and tensors layout must be defined in `llama.cpp` source files:
-1. Define a new `llm_arch` enum value in `src/llama-arch.h`.
-2. In `src/llama-arch.cpp`:
-    - Add the architecture name to the `LLM_ARCH_NAMES` map.
-    - Add the tensor mappings to the `LLM_TENSOR_NAMES` map.
-3. Add any non-standard metadata loading in the `llama_model_loader` constructor in `src/llama-model-loader.cpp`.
-4. If the model has a RoPE operation, add a case for the architecture in `llama_model_rope_type` function in `src/llama-model.cpp`.
+The model params and tensors layout must be defined in `llama.cpp`:
+1. Define a new `llm_arch`
+2. Define the tensors layout in `LLM_TENSOR_NAMES`
+3. Add any non-standard metadata in `llm_load_hparams`
+4. Create the tensors for inference in `llm_load_tensors`
+5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
 
 NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorch` dimensions.
 
 ### 3. Build the GGML graph implementation
 
-This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `src/llama-model.cpp`.
-Create a new struct that inherits from `llm_graph_context` and implement the graph-building logic in its constructor.
-Have a look at existing implementations like `llm_build_llama`, `llm_build_dbrx` or `llm_build_bert`.
-Then, in the `llama_model::build_graph` method, add a case for your architecture to instantiate your new graph-building struct.
+This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
+
+Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
 
 Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
 

docs/docker.md

@@ -25,9 +25,6 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
@@ -110,7 +107,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
 
 The defaults are:
 
-- `MUSA_VERSION` set to `rc4.2.0`
+- `MUSA_VERSION` set to `rc4.0.1`
 
 The resulting images, are essentially the same as the non-MUSA images:
 

docs/function-calling.md

@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2
   - Functionary v3.1 / v3.2
   - Hermes 2/3, Qwen 2.5
-  - Qwen 2.5 Coder
+  - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034)
   - Mistral Nemo
   - Firefunction v2
   - Command R7B

docs/multimodal.md

@@ -97,9 +97,6 @@ NOTE: some models may require large context window, for example: `-c 8192`
 # Qwen2-Audio and SeaLLM-Audio
 # note: no pre-quantized GGUF this model, as they have very poor result
 # ref: https://github.com/ggml-org/llama.cpp/pull/13760
-
-# Mistral's Voxtral
-(tool_name) -hf ggml-org/Voxtral-Mini-3B-2507-GGUF
 ```
 
 **Mixed modalities**:
@@ -110,7 +107,3 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
-
-## Finding more models:
-
-GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf

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 -j 32
+make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
 ```
 ### run on Orin
 ### case 1

docs/multimodal/minicpmo2.6.md

@@ -13,7 +13,7 @@ If there are differences in usage, please refer to the official build [documenta
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggml-org/llama.cpp
+git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
 ```
 
@@ -29,8 +29,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
 
 ```bash
-python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --minicpmv_version 4
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 # quantize int4 version

docs/multimodal/minicpmo4.0.md

@@ -1,47 +0,0 @@
-## MiniCPM-o 4
-
-### Prepare models and code
-
-Download [MiniCPM-o-4](https://huggingface.co/openbmb/MiniCPM-o-4) PyTorch model from huggingface to "MiniCPM-o-4" folder.
-
-
-### Build llama.cpp
-Readme modification time: 20250206
-
-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-o 4
-
-Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-4-gguf) by us)
-
-```bash
-python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-4
-python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-4 --minicpmv-projector ../MiniCPM-o-4/minicpmv.projector --output-dir ../MiniCPM-o-4/ --minicpmv_version 6
-python ./convert_hf_to_gguf.py ../MiniCPM-o-4/model
-
-# quantize int4 version
-./build/bin/llama-quantize ../MiniCPM-o-4/model/ggml-model-f16.gguf ../MiniCPM-o-4/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-o-4/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-4/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-o-4/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-4/mmproj-model-f16.gguf
-```

docs/multimodal/minicpmv2.5.md

@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
 ```bash
-python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --minicpmv_version 2
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version

docs/multimodal/minicpmv2.6.md

@@ -12,7 +12,7 @@ If there are differences in usage, please refer to the official build [documenta
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggml-org/llama.cpp
+git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
 ```
 
@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
 
 ```bash
-python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-2_6
-python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --minicpmv_version 3
+python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 # quantize int4 version

docs/multimodal/minicpmv4.0.md

@@ -1,47 +0,0 @@
-## MiniCPM-V 4
-
-### Prepare models and code
-
-Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model from huggingface to "MiniCPM-V-4" folder.
-
-
-### Build llama.cpp
-Readme modification time: 20250206
-
-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-gguf) by us)
-
-```bash
-python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4
-python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4 --minicpmv-projector ../MiniCPM-V-4/minicpmv.projector --output-dir ../MiniCPM-V-4/ --minicpmv_version 5
-python ./convert_hf_to_gguf.py ../MiniCPM-V-4/model
-
-# quantize int4 version
-./build/bin/llama-quantize ../MiniCPM-V-4/model/ggml-model-f16.gguf ../MiniCPM-V-4/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/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4/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/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4/mmproj-model-f16.gguf
-```

docs/ops.md

@@ -1,103 +0,0 @@
-# GGML Operations
-
-List of GGML operations and backend support status.
-
-## How to add a backend to this table:
-
-1. Run `test-backend-ops support --output csv` with your backend name and redirect output to a csv file in `docs/ops/` (e.g., `docs/ops/CUDA.csv`)
-2. Regenerate `/docs/ops.md` via `./scripts/create_ops_docs.py`
-
-Legend:
-- ✅ Fully supported by this backend
-- 🟡 Partially supported by this backend
-- ❌ Not supported by this backend
-
-| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
-|-----------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
-|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
-|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |

examples/CMakeLists.txt

@@ -33,8 +33,6 @@ else()
     add_subdirectory(speculative-simple)
     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/Miku.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"

examples/batched.swift/README.md

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

examples/chat-13B.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -e
 

examples/chat-persistent.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -euo pipefail
 

examples/chat-vicuna.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 set -e
 

examples/chat.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 
 #
 # Temporary script - will be removed in the future

examples/diffusion/CMakeLists.txt

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

examples/diffusion/README.md

@@ -1,13 +0,0 @@
-# Diffusion Text Generation
-
-This directory contains implementations for Diffusion LLMs (DLLMs)
-
-More Info:
-- https://github.com/ggml-org/llama.cpp/pull/14644
-- https://github.com/ggml-org/llama.cpp/pull/14771
-
-
-Example of using Dream architechture: `llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual`
-
-Example of using LLaDA architechture: `llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual`
-

examples/diffusion/diffusion-cli.cpp

@@ -1,683 +0,0 @@
-#include "arg.h"
-#include "chat.h"
-#include "common.h"
-#include "llama.h"
-#include "log.h"
-
-#include <limits.h>
-
-#include <algorithm>
-#include <cmath>
-#include <cstring>
-#include <limits>
-#include <random>
-#include <string>
-#include <vector>
-
-enum diffusion_algorithm { ORIGIN = 0, ENTROPY_BASED = 1, MARGIN_BASED = 2, RANDOM = 3, CONFIDENCE_BASED = 4 };
-
-// Unified transfer scheduling methods
-enum transfer_schedule {
-    TIMESTEP_BASED = 0,  // Dream-style: (1.0 - s/t) * remaining
-    BLOCK_BASED    = 1,  // LLaDA-style: process in blocks with get_num_transfer_tokens
-};
-
-typedef bool (*diffusion_step_callback_t)(int32_t             step,
-                                          int32_t             total_steps,
-                                          const llama_token * tokens,
-                                          int32_t             n_tokens,
-                                          void *              user_data);
-
-struct diffusion_params {
-    int32_t                   steps                   = 0;
-    float                     temperature             = 0;
-    llama_token               mask_token_id           = LLAMA_TOKEN_NULL;
-    diffusion_step_callback_t step_callback           = nullptr;
-    void *                    step_callback_user_data = nullptr;
-    int32_t                   seed                    = 0;
-    bool                      visual_mode             = false;
-    bool                      shift_logits            = false;  // Shift logits by -1 after decode
-
-    float   top_p = 0.;
-    int32_t top_k = 0.;
-
-    diffusion_algorithm algorithm = CONFIDENCE_BASED;
-    transfer_schedule   schedule  = TIMESTEP_BASED;
-
-    float   cfg_scale        = 0.;     // Config scale for classifier-free guidance
-    float   eps              = 0.;     // Timestep scheduling
-    int32_t block_length     = 0;      // Block size (for block scheduling)
-    float   alg_temp         = 0;      // algorithm temperature (0.0 = deterministic)
-    bool    add_gumbel_noise = false;  // Add gumbel noise to the logits if temp > 0.0
-
-    int32_t max_length = 0;            // Maximum sequence length
-};
-
-struct callback_data {
-    diffusion_params *  diff_params;
-    const llama_vocab * vocab;
-    int32_t             n_input;
-};
-
-static float calculate_confidence(const llama_token_data_array & cur_p,
-                                  diffusion_algorithm            algorithm,
-                                  std::mt19937 &                 rng) {
-    switch (algorithm) {
-        case CONFIDENCE_BASED:
-            return cur_p.data[cur_p.selected].p;  // Selected token probability
-
-        case ENTROPY_BASED:
-            {
-                float       entropy = 0.0f;
-                const float epsilon = 1e-10f;
-                for (size_t i = 0; i < cur_p.size; i++) {
-                    float prob = cur_p.data[i].p;
-                    entropy += prob * logf(prob + epsilon);
-                }
-                return -entropy;  // Higher entropy = lower confidence
-            }
-
-        case MARGIN_BASED:
-            return (cur_p.size > 1) ? cur_p.data[0].p - cur_p.data[1].p : cur_p.data[0].p;
-
-        case RANDOM:
-            {
-                std::uniform_real_distribution<float> uniform(0.0f, 1.0f);
-                return uniform(rng);  // Random confidence
-            }
-
-        case ORIGIN:
-            return cur_p.data[cur_p.selected].p;
-
-        default:
-            return 0.0f;
-    }
-}
-
-// Unified transfer count calculation function
-static int32_t calculate_transfer_count(int32_t                      step,
-                                        int32_t                      total_steps,
-                                        int32_t                      remaining_masked,
-                                        transfer_schedule            schedule,
-                                        float                        eps,
-                                        const std::vector<int32_t> & num_transfer_tokens = {}) {
-    switch (schedule) {
-        case TIMESTEP_BASED:
-            {
-                float t          = 1.0f - (float) step / total_steps * (1.0f - eps);
-                float s          = 1.0f - (float) (step + 1) / total_steps * (1.0f - eps);
-                float p_transfer = (step < total_steps - 1) ? (1.0f - s / t) : 1.0f;
-                return (int32_t) (remaining_masked * p_transfer);
-            }
-
-        case BLOCK_BASED:
-            if (!num_transfer_tokens.empty() && step < (int32_t) num_transfer_tokens.size()) {
-                return num_transfer_tokens[step];
-            }
-            return remaining_masked / (total_steps - step);  // Fallback
-
-        default:
-            return remaining_masked / (total_steps - step);
-    }
-}
-
-static bool diffusion_step_callback(int32_t             step,
-                                    int32_t             total_steps,
-                                    const llama_token * tokens,
-                                    int32_t             n_tokens,
-                                    void *              user_data) {
-    (void) user_data;
-
-    callback_data * data = static_cast<callback_data *>(user_data);
-
-    auto print_progress_bar = [](int32_t step, int32_t total_steps) {
-        int progress_percent = (step * 100) / total_steps;
-        int progress_bars    = (step * 50) / total_steps;
-        LOG_INF("\rdiffusion step: %d/%d [%s%s] %d%%",
-                step,
-                total_steps,
-                std::string(progress_bars, '=').c_str(),
-                std::string(50 - progress_bars, ' ').c_str(),
-                progress_percent);
-    };
-
-    if (data->diff_params->visual_mode) {
-        // Visual mode: clear
-        LOG_INF("\033[2J\033[H");  // Clear screen and move cursor to top-left
-
-        print_progress_bar(step, total_steps);
-
-        LOG_INF("\n");
-
-        std::string current_text = " ";
-
-        for (int32_t i = data->n_input; i < n_tokens; i++) {
-            std::string token_str;
-            if (tokens[i] != llama_vocab_mask(data->vocab)) {
-                char piece[256];
-                int  n_chars = llama_token_to_piece(data->vocab, tokens[i], piece, sizeof(piece), 0, false);
-                if (n_chars > 0) {
-                    piece[n_chars] = '\0';
-                    token_str      = piece;
-                }
-            } else {
-                token_str = " ";
-            }
-
-            current_text += token_str;
-        }
-
-        LOG_INF("%s\n", current_text.c_str());
-    } else {
-        print_progress_bar(step, total_steps);
-    }
-
-    return true;
-}
-
-static void add_gumbel_noise(float * logits, int32_t n_vocab, float temperature, std::mt19937 & rng) {
-    if (temperature == 0.0f) {
-        return;
-    }
-
-    std::uniform_real_distribution<double> uniform(0.0, 1.0);
-    for (int32_t i = 0; i < n_vocab; i++) {
-        double noise        = uniform(rng);
-        // Prevent log(0)
-        noise               = std::max(noise, 1e-20);
-        double gumbel_noise = std::pow(-std::log(noise), temperature);
-        logits[i]           = std::exp(logits[i]) / gumbel_noise;
-    }
-}
-
-static std::vector<int32_t> get_num_transfer_tokens(int32_t mask_count, int32_t steps) {
-    std::vector<int32_t> num_transfer_tokens(steps);
-
-    int32_t base      = mask_count / steps;
-    int32_t remainder = mask_count % steps;
-
-    for (int32_t i = 0; i < steps; i++) {
-        num_transfer_tokens[i] = base + (i < remainder ? 1 : 0);
-    }
-
-    return num_transfer_tokens;
-}
-
-static void diffusion_generate(llama_context *          ctx,
-                               const llama_token *      input_tokens,
-                               llama_token *            output_tokens,
-                               int32_t                  n_input,
-                               const diffusion_params & params,
-                               int32_t &                n_generated) {
-    n_generated = 0;
-    if (!ctx || !input_tokens || !output_tokens || n_input <= 0 || params.max_length <= n_input) {
-        return;
-    }
-
-    const llama_model * model = llama_get_model(ctx);
-
-    // Initialize with input and pad with mask tokens
-    std::copy(input_tokens, input_tokens + n_input, output_tokens);
-    std::fill(output_tokens + n_input, output_tokens + params.max_length, params.mask_token_id);
-
-    std::mt19937 rng(params.seed);
-
-    llama_set_causal_attn(ctx, false);
-
-    int32_t n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(model));
-
-    std::vector<llama_token_data> candidates(n_vocab);
-    std::vector<llama_token_data> conf_candidates;
-    conf_candidates.reserve(params.max_length);
-    std::vector<int32_t> mask_positions;
-    mask_positions.reserve(params.max_length);
-
-    // Setup sampler chain
-    struct llama_sampler * sampler = llama_sampler_chain_init(llama_sampler_chain_default_params());
-    if (params.top_k > 0) {
-        llama_sampler_chain_add(sampler, llama_sampler_init_top_k(params.top_k));
-    }
-    if (params.top_p < 1.0f) {
-        llama_sampler_chain_add(sampler, llama_sampler_init_top_p(params.top_p, 1));
-    }
-    if (params.temperature > 0.0f) {
-        llama_sampler_chain_add(sampler, llama_sampler_init_temp(params.temperature));
-    }
-    llama_sampler_chain_add(sampler, llama_sampler_init_dist(params.seed));
-
-    struct llama_sampler * dist_sampler = llama_sampler_init_dist(params.seed);
-
-    llama_batch batch = llama_batch_init(params.max_length, 0, 1);
-    batch.n_tokens    = params.max_length;
-
-    // Pre-allocate buffers for CFG if needed
-    int32_t                  logits_size = n_vocab * params.max_length;
-    std::vector<float>       cond_logits_buffer;
-    std::vector<llama_token> un_x_buffer;
-    if (params.cfg_scale > 0.0f) {
-        cond_logits_buffer.resize(logits_size);
-        un_x_buffer.resize(params.max_length);
-    }
-
-    // For block-based processing
-    std::vector<int32_t> num_transfer_tokens;
-    int32_t              num_blocks      = 1;
-    int32_t              steps_per_block = params.steps;
-
-    if (params.schedule == BLOCK_BASED) {
-        GGML_ASSERT(params.max_length % params.block_length == 0);
-        num_blocks = params.max_length / params.block_length;
-        GGML_ASSERT(params.steps % num_blocks == 0);
-        steps_per_block = params.steps / num_blocks;
-    }
-
-    std::vector<float> confidence(params.max_length);
-
-    int64_t total_sampling_time = 0;
-    int64_t total_time          = 0;
-    int64_t time_start          = ggml_time_us();
-
-    for (int block_num = 0; block_num < num_blocks; block_num++) {
-        int32_t block_start = (params.schedule == BLOCK_BASED) ? n_input + block_num * params.block_length : 0;
-        int32_t block_end   = (params.schedule == BLOCK_BASED) ?
-                                  std::min(n_input + (block_num + 1) * params.block_length, params.max_length) :
-                                  params.max_length;
-
-        // Count masked tokens in current block for block-based processing
-        if (params.schedule == BLOCK_BASED) {
-            int32_t block_mask_count = 0;
-            for (int i = block_start; i < block_end; i++) {
-                if (output_tokens[i] == params.mask_token_id) {
-                    block_mask_count++;
-                }
-            }
-            num_transfer_tokens = get_num_transfer_tokens(block_mask_count, steps_per_block);
-        }
-
-        for (int32_t step = 0; step < steps_per_block; step++) {
-            int32_t global_step = block_num * steps_per_block + step;
-
-            if (params.step_callback) {
-                if (!params.step_callback(
-                        global_step, params.steps, output_tokens, params.max_length, params.step_callback_user_data)) {
-                    break;
-                }
-            }
-
-            // Setup batch
-            for (int32_t i = 0; i < params.max_length; i++) {
-                batch.token[i]     = output_tokens[i];
-                batch.pos[i]       = i;
-                batch.n_seq_id[i]  = 1;
-                batch.seq_id[i][0] = 0;
-                batch.logits[i]    = 1;
-            }
-
-            float * logits = nullptr;
-
-            if (params.cfg_scale > 0.0f) {
-                int ret = llama_decode(ctx, batch);
-                if (ret != 0) {
-                    LOG_ERR("Failed to generate conditional");
-                    break;
-                }
-                float * cond_logits_ptr = llama_get_logits(ctx);
-                std::memcpy(cond_logits_buffer.data(), cond_logits_ptr, logits_size * sizeof(float));
-
-                // Unconditional generation (mask input)
-                std::copy(output_tokens, output_tokens + params.max_length, un_x_buffer.begin());
-                for (int32_t i = 0; i < n_input; i++) {
-                    un_x_buffer[i] = params.mask_token_id;
-                }
-
-                for (int32_t i = 0; i < params.max_length; i++) {
-                    batch.token[i] = un_x_buffer[i];
-                }
-                ret = llama_decode(ctx, batch);
-                if (ret != 0) {
-                    LOG_ERR("Failed to generate unconditional");
-                    break;
-                }
-                float * uncond_logits = llama_get_logits(ctx);
-
-                // Apply CFG
-                for (int32_t i = 0; i < logits_size; i++) {
-                    cond_logits_buffer[i] =
-                        uncond_logits[i] + (params.cfg_scale + 1.0f) * (cond_logits_buffer[i] - uncond_logits[i]);
-                }
-                logits = cond_logits_buffer.data();
-            } else {
-                int ret = llama_decode(ctx, batch);
-                if (ret != 0) {
-                    LOG_ERR("%s: failed to decode at step %d, ret = %d\n", __func__, global_step, ret);
-                    break;
-                }
-                logits = llama_get_logits(ctx);
-            }
-
-            if (!logits) {
-                LOG_ERR("%s: failed to get logits at step %d\n", __func__, global_step);
-                break;
-            }
-
-            auto get_logits_for_pos = [&](int32_t pos) -> const float * {
-                if (params.shift_logits) {
-                    return pos == 0 ? logits : logits + (pos - 1) * n_vocab;
-                }
-                return logits + (pos) *n_vocab;
-            };
-
-            int64_t time_start_sampling = ggml_time_us();
-
-            mask_positions.clear();
-            for (int32_t i = 0; i < params.max_length; i++) {
-                if (output_tokens[i] == params.mask_token_id) {
-                    // For block-based, only consider current block
-                    if (params.schedule != BLOCK_BASED || (i >= block_start && i < block_end)) {
-                        mask_positions.push_back(i);
-                    }
-                }
-            }
-
-            if (mask_positions.empty()) {
-                break;
-            }
-
-            if (params.add_gumbel_noise && params.temperature > 0.0f) {
-                add_gumbel_noise(logits, n_vocab, params.temperature, rng);
-            }
-
-            if (params.algorithm == ORIGIN) {
-                int32_t transfer_count = calculate_transfer_count(
-                    step, steps_per_block, mask_positions.size(), params.schedule, params.eps, num_transfer_tokens);
-                float p_transfer = (float) transfer_count / mask_positions.size();
-
-                for (int32_t pos : mask_positions) {
-                    if (std::uniform_real_distribution<float>(0.0f, 1.0f)(rng) < p_transfer) {
-                        const float * pos_logits = get_logits_for_pos(pos);
-                        for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
-                            candidates[token_id].id    = token_id;
-                            candidates[token_id].logit = pos_logits[token_id];
-                            candidates[token_id].p     = 0.0f;
-                        }
-
-                        llama_token_data_array cur_p = {
-                            candidates.data(),
-                            (size_t) n_vocab,
-                            -1,
-                            false,
-                        };
-
-                        llama_sampler_apply(sampler, &cur_p);
-                        output_tokens[pos] = cur_p.data[cur_p.selected].id;
-                    }
-                }
-            } else {
-                std::vector<std::pair<float, int32_t>> confidences;
-                std::vector<llama_token>               sampled_tokens(mask_positions.size());
-
-                for (size_t i = 0; i < mask_positions.size(); i++) {
-                    int32_t       pos        = mask_positions[i];
-                    const float * pos_logits = get_logits_for_pos(pos);
-
-                    for (int32_t token_id = 0; token_id < n_vocab; token_id++) {
-                        candidates[token_id].logit = pos_logits[token_id];
-                        candidates[token_id].p     = 0.0f;
-                        candidates[token_id].id    = token_id;
-                    }
-
-                    llama_token_data_array cur_p = {
-                        candidates.data(),
-                        candidates.size(),
-                        -1,
-                        false,
-                    };
-
-                    llama_sampler_apply(sampler, &cur_p);
-                    llama_token sampled_token = cur_p.data[cur_p.selected].id;
-
-                    float conf = calculate_confidence(cur_p, params.algorithm, rng);
-
-                    sampled_tokens[i] = sampled_token;
-                    confidences.emplace_back(conf, i);
-                }
-
-                int32_t transfer_count = calculate_transfer_count(
-                    step, steps_per_block, mask_positions.size(), params.schedule, params.eps, num_transfer_tokens);
-
-                if (transfer_count > 0) {
-                    if (params.alg_temp == 0.0f) {
-                        std::partial_sort(confidences.begin(),
-                                          confidences.begin() + std::min(transfer_count, (int32_t) confidences.size()),
-                                          confidences.end(),
-                                          [](const std::pair<float, int32_t> & a, const std::pair<float, int32_t> & b) {
-                                              if (a.first != b.first) {
-                                                  return a.first > b.first;
-                                              }
-                                              return a.second < b.second;
-                                          });
-
-                        for (int32_t i = 0; i < std::min(transfer_count, (int32_t) confidences.size()); i++) {
-                            int32_t mask_idx   = confidences[i].second;
-                            int32_t pos        = mask_positions[mask_idx];
-                            output_tokens[pos] = sampled_tokens[mask_idx];
-                        }
-                    } else {
-                        conf_candidates.clear();
-                        for (size_t i = 0; i < confidences.size(); i++) {
-                            float conf_logit = confidences[i].first / params.alg_temp;
-                            conf_candidates.emplace_back(llama_token_data{ (int32_t) i, conf_logit, 0.0f });
-                        }
-
-                        llama_token_data_array conf_array = {
-                            conf_candidates.data(),
-                            conf_candidates.size(),
-                            -1,
-                            false,
-                        };
-
-                        for (int32_t i = 0; i < std::min(transfer_count, (int32_t) confidences.size()); i++) {
-                            llama_sampler_apply(dist_sampler, &conf_array);
-                            int32_t selected_idx = conf_array.selected;
-                            int32_t mask_idx     = selected_idx;
-                            int32_t pos          = mask_positions[mask_idx];
-                            output_tokens[pos]   = sampled_tokens[mask_idx];
-
-                            conf_candidates[selected_idx].p = 0.0f;
-                            conf_array.selected             = -1;
-                        }
-                    }
-                }
-            }
-
-            int64_t time_end_sampling = ggml_time_us();
-            total_sampling_time += time_end_sampling - time_start_sampling;
-        }
-    }
-
-    int64_t time_end = ggml_time_us();
-    total_time += time_end - time_start;
-
-    LOG_INF("\ntotal time: %0.2fms, time per step: %0.2fms, sampling time per step: %0.2fms\n",
-            total_time / 1000.0,
-            total_time / 1000.0 / params.steps,
-            total_sampling_time / 1000.0 / params.steps);
-
-    llama_batch_free(batch);
-    llama_sampler_free(sampler);
-    llama_sampler_free(dist_sampler);
-
-    n_generated = params.max_length;
-}
-
-static std::string format_input_text(const std::string & 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;
-    inputs.messages.push_back(user_msg);
-
-    auto result = common_chat_templates_apply(chat_templates.get(), inputs);
-
-    return result.prompt;
-}
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DIFFUSION)) {
-        return 1;
-    }
-
-    common_init();
-    llama_backend_init();
-
-    llama_model_params model_params = llama_model_default_params();
-    model_params.n_gpu_layers       = params.n_gpu_layers;
-    model_params.devices            = params.devices.data();
-    model_params.use_mmap           = params.use_mmap;
-    model_params.use_mlock          = params.use_mlock;
-    model_params.check_tensors      = params.check_tensors;
-
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params);
-    if (!model) {
-        LOG_ERR("error: failed to load model '%s'\n", params.model.path.c_str());
-        return 1;
-    }
-
-    if (!llama_model_is_diffusion(model)) {
-        LOG_ERR("error: unsupported model for diffusion");
-        llama_model_free(model);
-        return 1;
-    }
-
-    llama_context_params ctx_params = llama_context_default_params();
-    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.no_perf              = params.no_perf;
-    ctx_params.type_k               = params.cache_type_k;
-    ctx_params.type_v               = params.cache_type_v;
-
-    llama_context * ctx = llama_init_from_model(model, ctx_params);
-    if (!ctx) {
-        LOG_ERR("error: failed to create context\n");
-        llama_model_free(model);
-        return 1;
-    }
-
-    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::vector<llama_token> input_tokens = common_tokenize(vocab,
-                                                            formatted_prompt,
-                                                            /*add special tokens*/ true,
-                                                            /*parse special*/ true);
-
-    int n_input = input_tokens.size();
-
-    if (n_input >= params.n_ctx) {
-        LOG_ERR("error: input too long (%d tokens), max context is %d\n", n_input, params.n_ctx);
-        llama_free(ctx);
-        llama_model_free(model);
-        return 1;
-    }
-
-    llama_token mask_token_id = llama_vocab_mask(vocab);
-    GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
-
-    bool visual_mode = params.diffusion.visual_mode;
-
-    int32_t                  n_generated = 0;
-    std::vector<llama_token> output_tokens(params.n_ubatch);
-
-    struct diffusion_params diff_params;
-
-    char shift_logits_str[8];
-    if (llama_model_meta_val_str(model, "diffusion.shift_logits", shift_logits_str, sizeof(shift_logits_str)) >= 0) {
-        diff_params.shift_logits = (strcmp(shift_logits_str, "true") == 0);
-    } else {
-        diff_params.shift_logits = true;
-    }
-
-    //Use either eps or block length, but not both
-    GGML_ASSERT((params.diffusion.eps == 0) ^ (params.diffusion.block_length == 0));
-
-    if (params.diffusion.eps) {
-        diff_params.schedule = TIMESTEP_BASED;
-        diff_params.eps      = params.diffusion.eps;
-    } else if (params.diffusion.block_length) {
-        diff_params.schedule     = BLOCK_BASED;
-        diff_params.block_length = params.diffusion.block_length;
-    }
-
-    diff_params.mask_token_id    = mask_token_id;
-    diff_params.seed             = params.sampling.seed;
-    diff_params.temperature      = params.sampling.temp;
-    diff_params.steps            = params.diffusion.steps;
-    diff_params.algorithm        = static_cast<diffusion_algorithm>(params.diffusion.algorithm);
-    diff_params.max_length       = params.n_ubatch;
-    diff_params.top_p            = params.sampling.top_p;
-    diff_params.top_k            = params.sampling.top_k;
-    diff_params.visual_mode      = params.diffusion.visual_mode;
-    diff_params.add_gumbel_noise = params.diffusion.add_gumbel_noise;
-
-    diff_params.step_callback           = diffusion_step_callback;
-    callback_data cb_data               = { &diff_params, vocab, n_input };
-    diff_params.step_callback_user_data = &cb_data;
-
-    const char * alg_names[]   = { "ORIGIN", "ENTROPY_BASED", "MARGIN_BASED", "RANDOM", "CONFIDENCE_BASED" };
-    const char * sched_names[] = { "TIMESTEP_BASED", "BLOCK_BASED" };
-    const char * alg_name =
-        (diff_params.algorithm >= 0 && diff_params.algorithm <= 4) ? alg_names[diff_params.algorithm] : "UNKNOWN";
-    const char * sched_name =
-        (diff_params.schedule >= 0 && diff_params.schedule <= 1) ? sched_names[diff_params.schedule] : "UNKNOWN";
-
-    LOG_INF("diffusion_params: - %-25s llama_token      = %d\n", "mask_token_id", mask_token_id);
-    LOG_INF("diffusion_params: - %-25s u32              = %d\n", "steps", diff_params.steps);
-    LOG_INF("diffusion_params: - %-25s u32              = %d\n", "max_length", diff_params.max_length);
-    LOG_INF("diffusion_params: - %-25s enum             = %d (%s)\n", "algorithm", diff_params.algorithm, alg_name);
-    LOG_INF("diffusion_params: - %-25s enum             = %d (%s)\n", "schedule", diff_params.schedule, sched_name);
-    LOG_INF("diffusion_params: - %-25s f32              = %.3f\n", "temperature", diff_params.temperature);
-    if (diff_params.schedule == TIMESTEP_BASED) {
-        LOG_INF("diffusion_params: - %-25s f32              = %.6f\n", "eps", diff_params.eps);
-        LOG_INF("diffusion_params: - %-25s f32              = %.3f\n", "alg_temp", diff_params.alg_temp);
-    }
-    if (diff_params.schedule == BLOCK_BASED) {
-        LOG_INF("diffusion_params: - %-25s u32              = %d\n", "block_length", diff_params.block_length);
-        LOG_INF("diffusion_params: - %-25s f32              = %.3f\n", "cfg_scale", diff_params.cfg_scale);
-    }
-
-    diffusion_generate(ctx, input_tokens.data(), output_tokens.data(), n_input, diff_params, n_generated);
-
-    if (n_generated > 0) {
-        if (visual_mode) {
-            //clear screen and move cursor to top-left
-            LOG_INF("\033[2J\033[H");
-        }
-
-        output_tokens.erase(output_tokens.begin(), output_tokens.begin() + n_input);
-        std::string output_data = common_detokenize(vocab, output_tokens, false);
-        LOG_INF("\n%s\n", output_data.c_str());
-    } else {
-        LOG_INF("Error: diffusion generation failed\n");
-    }
-
-    llama_free(ctx);
-    llama_model_free(model);
-    llama_backend_free();
-
-    return 0;
-}

examples/embedding/embedding.cpp

@@ -81,14 +81,6 @@ int main(int argc, char ** argv) {
 
     params.embedding = true;
 
-    // if the number of prompts that would be encoded is known in advance, it's more efficient to specify the
-    //   --parallel argument accordingly. for convenience, if not specified, we fallback to unified KV cache
-    //   in order to support any number of prompts
-    if (params.n_parallel == 1) {
-        LOG_INF("%s: n_parallel == 1 -> unified KV cache is enabled\n", __func__);
-        params.kv_unified = true;
-    }
-
     // utilize the full context
     if (params.n_batch < params.n_ctx) {
         LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
@@ -115,7 +107,7 @@ int main(int argc, char ** argv) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     const int n_ctx_train = llama_model_n_ctx_train(model);
-    const int n_ctx       = llama_n_ctx(ctx);
+    const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
@@ -141,36 +133,10 @@ int main(int argc, char ** argv) {
     // max batch size
     const uint64_t n_batch = params.n_batch;
 
-    // get added sep and eos token, if any
-    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
-    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
-
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        std::vector<llama_token> inp;
-
-        // split classification pairs and insert expected separator tokens
-        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
-            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
-            std::string final_prompt;
-
-            for (size_t i = 0; i < pairs.size(); i++) {
-                final_prompt += pairs[i];
-                if (i != pairs.size() - 1) {
-                    if (!added_eos_token.empty()) {
-                        final_prompt += added_eos_token;
-                    }
-                    if (!added_sep_token.empty()) {
-                        final_prompt += added_sep_token;
-                    }
-                }
-            }
-
-            inp = common_tokenize(ctx, final_prompt, true, true);
-        } else {
-            inp = common_tokenize(ctx, prompt, true, true);
-        }
+        auto inp = common_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -179,11 +145,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP/EOS
+    // check if the last token is SEP
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
-            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
+        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
+            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }

examples/eval-callback/eval-callback.cpp

@@ -7,7 +7,6 @@
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <numeric>
 
 /**
  * This the arbitrary data which will be passed to each callback.
@@ -56,8 +55,6 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         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) {
@@ -78,12 +75,6 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
         LOG("                                     ]\n");
         LOG("                                     sum = %f\n", sum);
     }
-
-    // TODO: make this abort configurable/optional?
-    if (std::isnan(sum)) {
-        LOG_ERR("encountered NaN - aborting\n");
-        exit(0);
-    }
 }
 
 /**
@@ -143,11 +134,6 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (tokens.empty()) {
-        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
-        return false;
-    }
-
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;

examples/gritlm/gritlm.cpp

@@ -41,11 +41,12 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, true);
+            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
         llama_memory_clear(llama_get_memory(ctx), true);
+        llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -102,6 +103,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
     llama_token eos_token = llama_vocab_eos(vocab);
 
     llama_memory_clear(llama_get_memory(ctx), true);
+    llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -164,8 +166,6 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
-    cparams.embeddings = true;
-
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,8 +213,6 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
-    llama_set_embeddings(ctx, false);
-
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {

examples/jeopardy/jeopardy.sh

@@ -1,4 +1,4 @@
-#!/usr/bin/env bash
+#!/bin/bash
 set -e
 
 MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin

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 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
 "
 "   --batch-size [512, model max context]
 "

examples/lookahead/README.md

@@ -5,9 +5,3 @@ Demonstration of lookahead decoding technique:
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 
 More info: https://github.com/ggml-org/llama.cpp/pull/4207
-
-Sample command:
-
-```bash
-llama-lookahead -hf ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF -p "// network server implemented in C\n// author: Peter Hacker\n\n#include" -e -ngl 99 -t 4 -n 512 -c 4096 -kvu
-```

examples/model-conversion/.gitignore

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

examples/model-conversion/CMakeLists.txt

@@ -1,5 +0,0 @@
-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

@@ -1,172 +0,0 @@
-# Validation functions
-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
-
-###
-### 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-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.sh
-
-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 ${CONVERTED_MODEL}
-
-causal-curl-embedding-endpoint: causal-run-original-embeddings
-	@./scripts/utils/curl-embedding-server.sh | ./scripts/causal/compare-embeddings-logits.sh
-
-causal-quantize-Q8_0: QUANTIZED_TYPE = Q8_0
-causal-quantize-Q8_0: causal-quantize-model
-
-causal-quantize-Q4_0: QUANTIZED_TYPE = Q4_0
-causal-quantize-Q4_0: causal-quantize-model
-
-causal-quantize-model:
-	@CONVERTED_MODEL="$(CONVERTED_MODEL)" QUANTIZED_TYPE="$(QUANTIZED_TYPE)" ./scripts/utils/quantize.sh ${CONVERTED_MODEL} ${QUANTIZED_TYPE}
-	@echo "Export the quantized model path to QUANTIZED_MODEL variable in your environment"
-
-causal-run-quantized-model:
-	@QUANTIZED_MODEL="$(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
-
-embedding-quantize-model:
-	@./scripts/utils/quantize.sh ${CONVERTED_EMBEDDING_MODEL} ${QUANTIZED_TYPE}
-	@echo "Export the quantized model path to QUANTIZED_EMBEDDING_MODEL variable in your environment"
-
-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

@@ -1,343 +0,0 @@
-# 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
-```
-
-
-## 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
-```
-
-## 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

@@ -1,210 +0,0 @@
-#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

@@ -1,5 +0,0 @@
---extra-index-url https://download.pytorch.org/whl/cpu
-torch~=2.6.0
-torchvision~=0.21.0
-transformers~=4.55.0
-huggingface-hub~=0.34.0

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

@@ -1,43 +0,0 @@
-#/bin/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

@@ -1,88 +0,0 @@
-#!/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

@@ -1,22 +0,0 @@
-#!/bin/bash
-
-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}"
-python ../../convert_hf_to_gguf.py --verbose \
-    ${MODEL_PATH} \
-    --outfile ${CONVERTED_MODEL} \
-    --outtype ${TYPE} \
-    --metadata "${METADATA_OVERRIDE}"
-
-echo ""
-echo "The environment variable CONVERTED_MODEL can be set to this path using:"
-echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"

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

@@ -1,13 +0,0 @@
----
-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