llama: fix patches for vendor update to 51fa458a9

- Patch 0020 (no-alloc mode): Add full CUDA implementation including
  graph_reserve, buffer_size, reset, noalloc_buffer functions,
  CUDA_ALIGNMENT define, and conditional allocation logic
- Patch 0015 (GPU UUIDs): Add UUID to device log messages for HIP,
  MUSA, and CUDA platforms
- Update patches 0018, 0022, 0024, 0029 with correct line offsets
  for the new base commit

.github/ISSUE_TEMPLATE/10_bug_report.yml

@@ -13,7 +13,7 @@ body:
     id: logs
     attributes:
       label: Relevant log output
-      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md#how-to-troubleshoot-issues) for details.
+      description: Please copy and paste any relevant log output. See [Troubleshooting Guide](https://github.com/ollama/ollama/blob/main/docs/troubleshooting.mdx#how-to-troubleshoot-issues) for details.
       render: shell
     validations:
       required: false

.github/workflows/release.yaml

@@ -68,6 +68,7 @@ jobs:
           name: bundles-darwin
           path: |
             dist/*.tgz
+            dist/*.tar.zst
             dist/*.zip
             dist/*.dmg
 
@@ -371,13 +372,17 @@ jobs:
           outputs: type=local,dest=dist/${{ matrix.os }}-${{ matrix.arch }}
           cache-from: type=registry,ref=${{ vars.DOCKER_REPO }}:latest
           cache-to: type=inline
+      - name: Deduplicate CUDA libraries
+        run: |
+          ./scripts/deduplicate_cuda_libs.sh dist/${{ matrix.os }}-${{ matrix.arch }}
       - run: |
           for COMPONENT in bin/* lib/ollama/*; do
             case "$COMPONENT" in
-              bin/ollama)                echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              bin/ollama*)               echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/*.so*)          echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_v*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/vulkan*)        echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
+              lib/ollama/mlx*)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}.tar.in ;;
               lib/ollama/cuda_jetpack5)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack5.tar.in ;;
               lib/ollama/cuda_jetpack6)  echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-jetpack6.tar.in ;;
               lib/ollama/rocm)           echo $COMPONENT >>ollama-${{ matrix.os }}-${{ matrix.arch }}-rocm.tar.in ;;
@@ -392,13 +397,13 @@ jobs:
           done
       - run: |
           for ARCHIVE in dist/${{ matrix.os }}-${{ matrix.arch }}/*.tar.in; do
-            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | pigz -9vc >$(basename ${ARCHIVE//.*/}.tgz);
+            tar c -C dist/${{ matrix.os }}-${{ matrix.arch }} -T $ARCHIVE --owner 0 --group 0 | zstd --ultra -22 -T0 >$(basename ${ARCHIVE//.*/}.tar.zst);
           done
       - uses: actions/upload-artifact@v4
         with:
           name: bundles-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.target }}
           path: |
-            *.tgz
+            *.tar.zst
 
   # Build each Docker variant (OS, arch, and flavor) separately. Using QEMU is unreliable and slower.
   docker-build-push:
@@ -531,7 +536,7 @@ jobs:
       - name: Upload release artifacts
         run: |
           pids=()
-          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.exe dist/*.dmg ; do
+          for payload in dist/*.txt dist/*.zip dist/*.tgz dist/*.tar.zst dist/*.exe dist/*.dmg ; do
             echo "Uploading $payload"
             gh release upload ${GITHUB_REF_NAME} $payload --clobber &
             pids[$!]=$!

CMakeLists.txt

@@ -2,6 +2,22 @@ cmake_minimum_required(VERSION 3.21)
 
 project(Ollama C CXX)
 
+# Handle cross-compilation on macOS: when CMAKE_OSX_ARCHITECTURES is set to a
+# single architecture different from the host, override CMAKE_SYSTEM_PROCESSOR
+# to match. This is necessary because CMAKE_SYSTEM_PROCESSOR defaults to the
+# host architecture, but downstream projects (like MLX) use it to detect the
+# target architecture.
+if(CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES ";")
+    # Single architecture specified
+    if(CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+        message(STATUS "Cross-compiling for x86_64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to x86_64")
+        set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+    elseif(CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" AND NOT CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        message(STATUS "Cross-compiling for arm64: overriding CMAKE_SYSTEM_PROCESSOR from ${CMAKE_SYSTEM_PROCESSOR} to arm64")
+        set(CMAKE_SYSTEM_PROCESSOR "arm64")
+    endif()
+endif()
+
 include(CheckLanguage)
 include(GNUInstallDirs)
 
@@ -12,7 +28,7 @@ set(BUILD_SHARED_LIBS ON)
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_EXTENSIONS OFF)
+set(CMAKE_CXX_EXTENSIONS ON) # Recent versions of MLX Requires gnu++17 extensions to compile properly
 
 set(GGML_BUILD ON)
 set(GGML_SHARED ON)
@@ -32,9 +48,10 @@ if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     set(GGML_CPU_ALL_VARIANTS ON)
 endif()
 
-if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+if(APPLE)
     set(CMAKE_BUILD_RPATH "@loader_path")
     set(CMAKE_INSTALL_RPATH "@loader_path")
+    set(CMAKE_BUILD_WITH_INSTALL_RPATH ON)
 endif()
 
 set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
@@ -147,14 +164,56 @@ if(CMAKE_HIP_COMPILER)
     endif()
 endif()
 
-find_package(Vulkan)
-if(Vulkan_FOUND)
-    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
-    install(TARGETS ggml-vulkan
-        RUNTIME_DEPENDENCIES
-            PRE_INCLUDE_REGEXES vulkan
-            PRE_EXCLUDE_REGEXES ".*"
-        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
-    )
+if(NOT APPLE)
+    find_package(Vulkan)
+    if(Vulkan_FOUND)
+        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-vulkan)
+        install(TARGETS ggml-vulkan
+            RUNTIME_DEPENDENCIES
+                PRE_INCLUDE_REGEXES vulkan
+                PRE_EXCLUDE_REGEXES ".*"
+            RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+            LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT Vulkan
+        )
+    endif()
 endif()
+
+option(MLX_ENGINE "Enable MLX backend" OFF)
+
+if(MLX_ENGINE)
+    message(STATUS "Setting up MLX (this takes a while...)")
+    add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/x/ml/backend/mlx)
+
+    # Find CUDA toolkit if MLX is built with CUDA support
+    find_package(CUDAToolkit)
+
+    install(TARGETS mlx mlxc
+        RUNTIME_DEPENDENCIES
+            DIRECTORIES ${CUDAToolkit_BIN_DIR} ${CUDAToolkit_BIN_DIR}/x64 ${CUDAToolkit_LIBRARY_DIR}
+            PRE_INCLUDE_REGEXES cublas cublasLt cudart nvrtc nvrtc-builtins cudnn nccl openblas gfortran
+            PRE_EXCLUDE_REGEXES ".*"
+        RUNTIME DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        LIBRARY DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+        FRAMEWORK DESTINATION ${OLLAMA_INSTALL_DIR} COMPONENT MLX
+    )
+
+    # Install the Metal library for macOS arm64 (must be colocated with the binary)
+    # Metal backend is only built for arm64, not x86_64
+    if(APPLE AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
+        install(FILES ${CMAKE_BINARY_DIR}/_deps/mlx-build/mlx/backend/metal/kernels/mlx.metallib
+            DESTINATION ${OLLAMA_INSTALL_DIR}
+            COMPONENT MLX)
+    endif()
+
+    # Manually install cudart and cublas since they might not be picked up as direct dependencies
+    if(CUDAToolkit_FOUND)
+        file(GLOB CUDART_LIBS
+            "${CUDAToolkit_LIBRARY_DIR}/libcudart.so*"
+            "${CUDAToolkit_LIBRARY_DIR}/libcublas.so*")
+        if(CUDART_LIBS)
+            install(FILES ${CUDART_LIBS}
+                DESTINATION ${OLLAMA_INSTALL_DIR}
+                COMPONENT MLX)
+        endif()
+    endif()
+endif()

CMakePresets.json

@@ -41,7 +41,7 @@
       "inherits": [ "CUDA" ],
       "cacheVariables": {
         "CMAKE_CUDA_ARCHITECTURES": "75-virtual;80-virtual;86-virtual;87-virtual;89-virtual;90-virtual;90a-virtual;100-virtual;103-virtual;110-virtual;120-virtual;121-virtual",
-        "CMAKE_CUDA_FLAGS": "-t 2",
+        "CMAKE_CUDA_FLAGS": "-t 4",
         "OLLAMA_RUNNER_DIR": "cuda_v13"
       }
     },
@@ -83,6 +83,28 @@
       "cacheVariables": {
         "OLLAMA_RUNNER_DIR": "vulkan"
       }
+    },
+    {
+      "name": "MLX",
+      "inherits": [ "Default" ],
+      "cacheVariables": {
+        "MLX_ENGINE": "ON",
+        "OLLAMA_RUNNER_DIR": "mlx"
+      }
+    },
+    {
+      "name": "MLX CUDA 12",
+      "inherits": [ "MLX", "CUDA 12" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v12"
+      }
+    },
+    {
+      "name": "MLX CUDA 13",
+      "inherits": [ "MLX", "CUDA 13" ],
+      "cacheVariables": {
+        "OLLAMA_RUNNER_DIR": "mlx_cuda_v13"
+      }
     }
   ],
   "buildPresets": [
@@ -140,6 +162,21 @@
       "name": "Vulkan",
       "targets": [ "ggml-vulkan" ],
       "configurePreset": "Vulkan"
+    },
+    {
+      "name": "MLX",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX"
+    },
+    {
+      "name": "MLX CUDA 12",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 12"
+    },
+    {
+      "name": "MLX CUDA 13",
+      "targets": [ "mlx", "mlxc" ],
+      "configurePreset": "MLX CUDA 13"
     }
   ]
 }

Dockerfile

@@ -32,7 +32,7 @@ ENV PATH=/${VULKANVERSION}/x86_64/bin:$PATH
 FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
 RUN yum install -y yum-utils epel-release \
-    && dnf install -y clang ccache \
+    && dnf install -y clang ccache git \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
 ENV CC=clang CXX=clang++
 
@@ -131,8 +131,32 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'Vulkan' \
         && cmake --build --parallel --preset 'Vulkan' \
-        && cmake --install build --component Vulkan --strip --parallel 8 
+        && cmake --install build --component Vulkan --strip --parallel 8
 
+FROM base AS mlx
+ARG CUDA13VERSION=13.0
+RUN dnf install -y cuda-toolkit-${CUDA13VERSION//./-} \
+    && dnf install -y openblas-devel lapack-devel \
+    && dnf install -y libcudnn9-cuda-13 libcudnn9-devel-cuda-13 \
+    && dnf install -y libnccl libnccl-devel
+ENV PATH=/usr/local/cuda-13/bin:$PATH
+ENV BLAS_INCLUDE_DIRS=/usr/include/openblas
+ENV LAPACK_INCLUDE_DIRS=/usr/include/openblas
+ENV CGO_LDFLAGS="-L/usr/local/cuda-13/lib64 -L/usr/local/cuda-13/targets/x86_64-linux/lib/stubs"
+ARG PARALLEL
+WORKDIR /go/src/github.com/ollama/ollama
+COPY CMakeLists.txt CMakePresets.json .
+COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
+COPY x/ml/backend/mlx x/ml/backend/mlx
+COPY go.mod go.sum .
+COPY MLX_VERSION .
+RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-$(case $(uname -m) in x86_64) echo amd64 ;; aarch64) echo arm64 ;; esac).tar.gz | tar xz -C /usr/local
+ENV PATH=/usr/local/go/bin:$PATH
+RUN go mod download
+RUN --mount=type=cache,target=/root/.ccache \
+    cmake --preset 'MLX CUDA 13' -DBLAS_INCLUDE_DIRS=/usr/include/openblas -DLAPACK_INCLUDE_DIRS=/usr/include/openblas \
+        && cmake --build --parallel ${PARALLEL} --preset 'MLX CUDA 13' \
+        && cmake --install build --component MLX --strip --parallel ${PARALLEL}
 
 FROM base AS build
 WORKDIR /go/src/github.com/ollama/ollama
@@ -141,18 +165,21 @@ RUN curl -fsSL https://golang.org/dl/go$(awk '/^go/ { print $2 }' go.mod).linux-
 ENV PATH=/usr/local/go/bin:$PATH
 RUN go mod download
 COPY . .
+# Clone mlx-c headers for CGO (version from MLX_VERSION file)
+RUN git clone --depth 1 --branch "$(cat MLX_VERSION)" https://github.com/ml-explore/mlx-c.git build/_deps/mlx-c-src
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
-ARG CGO_CFLAGS
+ENV CGO_CFLAGS="-I/go/src/github.com/ollama/ollama/build/_deps/mlx-c-src"
 ARG CGO_CXXFLAGS
 RUN --mount=type=cache,target=/root/.cache/go-build \
-    go build -trimpath -buildmode=pie -o /bin/ollama .
+    go build -tags mlx -trimpath -buildmode=pie -o /bin/ollama .
 
 FROM --platform=linux/amd64 scratch AS amd64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
 COPY --from=cuda-12 dist/lib/ollama /lib/ollama/
 COPY --from=cuda-13 dist/lib/ollama /lib/ollama/
 COPY --from=vulkan  dist/lib/ollama  /lib/ollama/
+COPY --from=mlx     /go/src/github.com/ollama/ollama/dist/lib/ollama /lib/ollama/
 
 FROM --platform=linux/arm64 scratch AS arm64
 # COPY --from=cuda-11 dist/lib/ollama/ /lib/ollama/
@@ -171,7 +198,7 @@ COPY --from=build /bin/ollama /bin/ollama
 
 FROM ubuntu:24.04
 RUN apt-get update \
-    && apt-get install -y ca-certificates libvulkan1 \
+    && apt-get install -y ca-certificates libvulkan1 libopenblas0 \
     && apt-get clean \
     && rm -rf /var/lib/apt/lists/*
 COPY --from=archive /bin /usr/bin

MLX_VERSION

@@ -0,0 +1 @@
+v0.4.1

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=ec98e2002
+FETCH_HEAD=51fa458a92d6a3f305f8fd76fc8f702e3e87ddb5
 
 .PHONY: help
 help:

README.md

@@ -48,7 +48,7 @@ ollama run gemma3
 
 ## Model library
 
-Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library 'ollama model library')
+Ollama supports a list of models available on [ollama.com/library](https://ollama.com/library "ollama model library")
 
 Here are some example models that can be downloaded:
 
@@ -79,7 +79,7 @@ Here are some example models that can be downloaded:
 | Code Llama         | 7B         | 3.8GB | `ollama run codellama`           |
 | Llama 2 Uncensored | 7B         | 3.8GB | `ollama run llama2-uncensored`   |
 | LLaVA              | 7B         | 4.5GB | `ollama run llava`               |
-| Granite-3.3         | 8B         | 4.9GB | `ollama run granite3.3`          |
+| Granite-3.3        | 8B         | 4.9GB | `ollama run granite3.3`          |
 
 > [!NOTE]
 > You should have at least 8 GB of RAM available to run the 7B models, 16 GB to run the 13B models, and 32 GB to run the 33B models.
@@ -260,6 +260,38 @@ Finally, in a separate shell, run a model:
 ./ollama run llama3.2
 ```
 
+## Building with MLX (experimental)
+
+First build the MLX libraries:
+
+```shell
+cmake --preset MLX
+cmake --build --preset MLX --parallel
+cmake --install build --component MLX
+```
+
+When building with the `-tags mlx` flag, the main `ollama` binary includes MLX support for experimental features like image generation:
+
+```shell
+go build -tags mlx .
+```
+
+Finally, start the server:
+
+```
+./ollama serve
+```
+
+### Building MLX with CUDA
+
+When building with CUDA, use the preset "MLX CUDA 13" or "MLX CUDA 12" to enable CUDA with default architectures:
+
+```shell
+cmake --preset 'MLX CUDA 13'
+cmake --build --preset 'MLX CUDA 13' --parallel
+cmake --install build --component MLX
+```
+
 ## REST API
 
 Ollama has a REST API for running and managing models.
@@ -290,6 +322,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 
 ### Web & Desktop
 
+- [Onyx](https://github.com/onyx-dot-app/onyx)
 - [Open WebUI](https://github.com/open-webui/open-webui)
 - [SwiftChat (macOS with ReactNative)](https://github.com/aws-samples/swift-chat)
 - [Enchanted (macOS native)](https://github.com/AugustDev/enchanted)
@@ -421,7 +454,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 - [Lumina](https://github.com/cushydigit/lumina.git) (A lightweight, minimal React.js frontend for interacting with Ollama servers)
 - [Tiny Notepad](https://pypi.org/project/tiny-notepad) (A lightweight, notepad-like interface to chat with ollama available on PyPI)
-- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.) 
+- [macLlama (macOS native)](https://github.com/hellotunamayo/macLlama) (A native macOS GUI application for interacting with Ollama models, featuring a chat interface.)
 - [GPTranslate](https://github.com/philberndt/GPTranslate) (A fast and lightweight, AI powered desktop translation application written with Rust and Tauri. Features real-time translation with OpenAI/Azure/Ollama.)
 - [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
 - [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
@@ -493,7 +526,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Database
 
 - [pgai](https://github.com/timescale/pgai) - PostgreSQL as a vector database (Create and search embeddings from Ollama models using pgvector)
-   - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
+  - [Get started guide](https://github.com/timescale/pgai/blob/main/docs/vectorizer-quick-start.md)
 - [MindsDB](https://github.com/mindsdb/mindsdb/blob/staging/mindsdb/integrations/handlers/ollama_handler/README.md) (Connects Ollama models with nearly 200 data platforms and apps)
 - [chromem-go](https://github.com/philippgille/chromem-go/blob/v0.5.0/embed_ollama.go) with [example](https://github.com/philippgille/chromem-go/tree/v0.5.0/examples/rag-wikipedia-ollama)
 - [Kangaroo](https://github.com/dbkangaroo/kangaroo) (AI-powered SQL client and admin tool for popular databases)
@@ -636,6 +669,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [llama.cpp](https://github.com/ggml-org/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
+
 - [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native integration to Ollama.
 - [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
@@ -644,4 +678,5 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [MLflow Tracing](https://mlflow.org/docs/latest/llms/tracing/index.html#automatic-tracing) is an open source LLM observability tool with a convenient API to log and visualize traces, making it easy to debug and evaluate GenAI applications.
 
 ### Security
+
 - [Ollama Fortress](https://github.com/ParisNeo/ollama_proxy_server)

anthropic/anthropic.go

@@ -0,0 +1,778 @@
+package anthropic
+
+import (
+	"crypto/rand"
+	"encoding/base64"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"log/slog"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Error types matching Anthropic API
+type Error struct {
+	Type    string `json:"type"`
+	Message string `json:"message"`
+}
+
+type ErrorResponse struct {
+	Type      string `json:"type"` // always "error"
+	Error     Error  `json:"error"`
+	RequestID string `json:"request_id,omitempty"`
+}
+
+// NewError creates a new ErrorResponse with the appropriate error type based on HTTP status code
+func NewError(code int, message string) ErrorResponse {
+	var etype string
+	switch code {
+	case http.StatusBadRequest:
+		etype = "invalid_request_error"
+	case http.StatusUnauthorized:
+		etype = "authentication_error"
+	case http.StatusForbidden:
+		etype = "permission_error"
+	case http.StatusNotFound:
+		etype = "not_found_error"
+	case http.StatusTooManyRequests:
+		etype = "rate_limit_error"
+	case http.StatusServiceUnavailable, 529:
+		etype = "overloaded_error"
+	default:
+		etype = "api_error"
+	}
+
+	return ErrorResponse{
+		Type:      "error",
+		Error:     Error{Type: etype, Message: message},
+		RequestID: generateID("req"),
+	}
+}
+
+// Request types
+
+// MessagesRequest represents an Anthropic Messages API request
+type MessagesRequest struct {
+	Model         string          `json:"model"`
+	MaxTokens     int             `json:"max_tokens"`
+	Messages      []MessageParam  `json:"messages"`
+	System        any             `json:"system,omitempty"` // string or []ContentBlock
+	Stream        bool            `json:"stream,omitempty"`
+	Temperature   *float64        `json:"temperature,omitempty"`
+	TopP          *float64        `json:"top_p,omitempty"`
+	TopK          *int            `json:"top_k,omitempty"`
+	StopSequences []string        `json:"stop_sequences,omitempty"`
+	Tools         []Tool          `json:"tools,omitempty"`
+	ToolChoice    *ToolChoice     `json:"tool_choice,omitempty"`
+	Thinking      *ThinkingConfig `json:"thinking,omitempty"`
+	Metadata      *Metadata       `json:"metadata,omitempty"`
+}
+
+// MessageParam represents a message in the request
+type MessageParam struct {
+	Role    string `json:"role"`    // "user" or "assistant"
+	Content any    `json:"content"` // string or []ContentBlock
+}
+
+// ContentBlock represents a content block in a message.
+// Text and Thinking use pointers so they serialize as the field being present (even if empty)
+// only when set, which is required for SDK streaming accumulation.
+type ContentBlock struct {
+	Type string `json:"type"` // text, image, tool_use, tool_result, thinking
+
+	// For text blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Text *string `json:"text,omitempty"`
+
+	// For image blocks
+	Source *ImageSource `json:"source,omitempty"`
+
+	// For tool_use blocks
+	ID    string `json:"id,omitempty"`
+	Name  string `json:"name,omitempty"`
+	Input any    `json:"input,omitempty"`
+
+	// For tool_result blocks
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   any    `json:"content,omitempty"` // string or []ContentBlock
+	IsError   bool   `json:"is_error,omitempty"`
+
+	// For thinking blocks - pointer so field only appears when set (SDK requires it for accumulation)
+	Thinking  *string `json:"thinking,omitempty"`
+	Signature string  `json:"signature,omitempty"`
+}
+
+// ImageSource represents the source of an image
+type ImageSource struct {
+	Type      string `json:"type"` // "base64" or "url"
+	MediaType string `json:"media_type,omitempty"`
+	Data      string `json:"data,omitempty"`
+	URL       string `json:"url,omitempty"`
+}
+
+// Tool represents a tool definition
+type Tool struct {
+	Type        string          `json:"type,omitempty"` // "custom" for user-defined tools
+	Name        string          `json:"name"`
+	Description string          `json:"description,omitempty"`
+	InputSchema json.RawMessage `json:"input_schema,omitempty"`
+}
+
+// ToolChoice controls how the model uses tools
+type ToolChoice struct {
+	Type                   string `json:"type"` // "auto", "any", "tool", "none"
+	Name                   string `json:"name,omitempty"`
+	DisableParallelToolUse bool   `json:"disable_parallel_tool_use,omitempty"`
+}
+
+// ThinkingConfig controls extended thinking
+type ThinkingConfig struct {
+	Type         string `json:"type"` // "enabled" or "disabled"
+	BudgetTokens int    `json:"budget_tokens,omitempty"`
+}
+
+// Metadata for the request
+type Metadata struct {
+	UserID string `json:"user_id,omitempty"`
+}
+
+// Response types
+
+// MessagesResponse represents an Anthropic Messages API response
+type MessagesResponse struct {
+	ID           string         `json:"id"`
+	Type         string         `json:"type"` // "message"
+	Role         string         `json:"role"` // "assistant"
+	Model        string         `json:"model"`
+	Content      []ContentBlock `json:"content"`
+	StopReason   string         `json:"stop_reason,omitempty"`
+	StopSequence string         `json:"stop_sequence,omitempty"`
+	Usage        Usage          `json:"usage"`
+}
+
+// Usage contains token usage information
+type Usage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens"`
+}
+
+// Streaming event types
+
+// MessageStartEvent is sent at the start of streaming
+type MessageStartEvent struct {
+	Type    string           `json:"type"` // "message_start"
+	Message MessagesResponse `json:"message"`
+}
+
+// ContentBlockStartEvent signals the start of a content block
+type ContentBlockStartEvent struct {
+	Type         string       `json:"type"` // "content_block_start"
+	Index        int          `json:"index"`
+	ContentBlock ContentBlock `json:"content_block"`
+}
+
+// ContentBlockDeltaEvent contains incremental content updates
+type ContentBlockDeltaEvent struct {
+	Type  string `json:"type"` // "content_block_delta"
+	Index int    `json:"index"`
+	Delta Delta  `json:"delta"`
+}
+
+// Delta represents an incremental update
+type Delta struct {
+	Type        string `json:"type"` // "text_delta", "input_json_delta", "thinking_delta", "signature_delta"
+	Text        string `json:"text,omitempty"`
+	PartialJSON string `json:"partial_json,omitempty"`
+	Thinking    string `json:"thinking,omitempty"`
+	Signature   string `json:"signature,omitempty"`
+}
+
+// ContentBlockStopEvent signals the end of a content block
+type ContentBlockStopEvent struct {
+	Type  string `json:"type"` // "content_block_stop"
+	Index int    `json:"index"`
+}
+
+// MessageDeltaEvent contains updates to the message
+type MessageDeltaEvent struct {
+	Type  string       `json:"type"` // "message_delta"
+	Delta MessageDelta `json:"delta"`
+	Usage DeltaUsage   `json:"usage"`
+}
+
+// MessageDelta contains stop information
+type MessageDelta struct {
+	StopReason   string `json:"stop_reason,omitempty"`
+	StopSequence string `json:"stop_sequence,omitempty"`
+}
+
+// DeltaUsage contains cumulative token usage
+type DeltaUsage struct {
+	OutputTokens int `json:"output_tokens"`
+}
+
+// MessageStopEvent signals the end of the message
+type MessageStopEvent struct {
+	Type string `json:"type"` // "message_stop"
+}
+
+// PingEvent is a keepalive event
+type PingEvent struct {
+	Type string `json:"type"` // "ping"
+}
+
+// StreamErrorEvent is an error during streaming
+type StreamErrorEvent struct {
+	Type  string `json:"type"` // "error"
+	Error Error  `json:"error"`
+}
+
+// FromMessagesRequest converts an Anthropic MessagesRequest to an Ollama api.ChatRequest
+func FromMessagesRequest(r MessagesRequest) (*api.ChatRequest, error) {
+	var messages []api.Message
+
+	if r.System != nil {
+		switch sys := r.System.(type) {
+		case string:
+			if sys != "" {
+				messages = append(messages, api.Message{Role: "system", Content: sys})
+			}
+		case []any:
+			// System can be an array of content blocks
+			var content strings.Builder
+			for _, block := range sys {
+				if blockMap, ok := block.(map[string]any); ok {
+					if blockMap["type"] == "text" {
+						if text, ok := blockMap["text"].(string); ok {
+							content.WriteString(text)
+						}
+					}
+				}
+			}
+			if content.Len() > 0 {
+				messages = append(messages, api.Message{Role: "system", Content: content.String()})
+			}
+		}
+	}
+
+	for _, msg := range r.Messages {
+		converted, err := convertMessage(msg)
+		if err != nil {
+			return nil, err
+		}
+		messages = append(messages, converted...)
+	}
+
+	options := make(map[string]any)
+
+	options["num_predict"] = r.MaxTokens
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature
+	}
+
+	if r.TopP != nil {
+		options["top_p"] = *r.TopP
+	}
+
+	if r.TopK != nil {
+		options["top_k"] = *r.TopK
+	}
+
+	if len(r.StopSequences) > 0 {
+		options["stop"] = r.StopSequences
+	}
+
+	var tools api.Tools
+	for _, t := range r.Tools {
+		tool, err := convertTool(t)
+		if err != nil {
+			return nil, err
+		}
+		tools = append(tools, tool)
+	}
+
+	var think *api.ThinkValue
+	if r.Thinking != nil && r.Thinking.Type == "enabled" {
+		think = &api.ThinkValue{Value: true}
+	}
+
+	stream := r.Stream
+
+	return &api.ChatRequest{
+		Model:    r.Model,
+		Messages: messages,
+		Options:  options,
+		Stream:   &stream,
+		Tools:    tools,
+		Think:    think,
+	}, nil
+}
+
+// convertMessage converts an Anthropic MessageParam to Ollama api.Message(s)
+func convertMessage(msg MessageParam) ([]api.Message, error) {
+	var messages []api.Message
+	role := strings.ToLower(msg.Role)
+
+	switch content := msg.Content.(type) {
+	case string:
+		messages = append(messages, api.Message{Role: role, Content: content})
+
+	case []any:
+		var textContent strings.Builder
+		var images []api.ImageData
+		var toolCalls []api.ToolCall
+		var thinking string
+		var toolResults []api.Message
+
+		for _, block := range content {
+			blockMap, ok := block.(map[string]any)
+			if !ok {
+				return nil, errors.New("invalid content block format")
+			}
+
+			blockType, _ := blockMap["type"].(string)
+
+			switch blockType {
+			case "text":
+				if text, ok := blockMap["text"].(string); ok {
+					textContent.WriteString(text)
+				}
+
+			case "image":
+				source, ok := blockMap["source"].(map[string]any)
+				if !ok {
+					return nil, errors.New("invalid image source")
+				}
+
+				sourceType, _ := source["type"].(string)
+				if sourceType == "base64" {
+					data, _ := source["data"].(string)
+					decoded, err := base64.StdEncoding.DecodeString(data)
+					if err != nil {
+						return nil, fmt.Errorf("invalid base64 image data: %w", err)
+					}
+					images = append(images, decoded)
+				} else {
+					return nil, fmt.Errorf("invalid image source type: %s. Only base64 images are supported.", sourceType)
+				}
+				// URL images would need to be fetched - skip for now
+
+			case "tool_use":
+				id, ok := blockMap["id"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'id' field")
+				}
+				name, ok := blockMap["name"].(string)
+				if !ok {
+					return nil, errors.New("tool_use block missing required 'name' field")
+				}
+				tc := api.ToolCall{
+					ID: id,
+					Function: api.ToolCallFunction{
+						Name: name,
+					},
+				}
+				if input, ok := blockMap["input"].(map[string]any); ok {
+					tc.Function.Arguments = mapToArgs(input)
+				}
+				toolCalls = append(toolCalls, tc)
+
+			case "tool_result":
+				toolUseID, _ := blockMap["tool_use_id"].(string)
+				var resultContent string
+
+				switch c := blockMap["content"].(type) {
+				case string:
+					resultContent = c
+				case []any:
+					for _, cb := range c {
+						if cbMap, ok := cb.(map[string]any); ok {
+							if cbMap["type"] == "text" {
+								if text, ok := cbMap["text"].(string); ok {
+									resultContent += text
+								}
+							}
+						}
+					}
+				}
+
+				toolResults = append(toolResults, api.Message{
+					Role:       "tool",
+					Content:    resultContent,
+					ToolCallID: toolUseID,
+				})
+
+			case "thinking":
+				if t, ok := blockMap["thinking"].(string); ok {
+					thinking = t
+				}
+			}
+		}
+
+		if textContent.Len() > 0 || len(images) > 0 || len(toolCalls) > 0 || thinking != "" {
+			m := api.Message{
+				Role:      role,
+				Content:   textContent.String(),
+				Images:    images,
+				ToolCalls: toolCalls,
+				Thinking:  thinking,
+			}
+			messages = append(messages, m)
+		}
+
+		// Add tool results as separate messages
+		messages = append(messages, toolResults...)
+
+	default:
+		return nil, fmt.Errorf("invalid message content type: %T", content)
+	}
+
+	return messages, nil
+}
+
+// convertTool converts an Anthropic Tool to an Ollama api.Tool
+func convertTool(t Tool) (api.Tool, error) {
+	var params api.ToolFunctionParameters
+	if len(t.InputSchema) > 0 {
+		if err := json.Unmarshal(t.InputSchema, &params); err != nil {
+			return api.Tool{}, fmt.Errorf("invalid input_schema for tool %q: %w", t.Name, err)
+		}
+	}
+
+	return api.Tool{
+		Type: "function",
+		Function: api.ToolFunction{
+			Name:        t.Name,
+			Description: t.Description,
+			Parameters:  params,
+		},
+	}, nil
+}
+
+// ToMessagesResponse converts an Ollama api.ChatResponse to an Anthropic MessagesResponse
+func ToMessagesResponse(id string, r api.ChatResponse) MessagesResponse {
+	var content []ContentBlock
+
+	if r.Message.Thinking != "" {
+		content = append(content, ContentBlock{
+			Type:     "thinking",
+			Thinking: ptr(r.Message.Thinking),
+		})
+	}
+
+	if r.Message.Content != "" {
+		content = append(content, ContentBlock{
+			Type: "text",
+			Text: ptr(r.Message.Content),
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		content = append(content, ContentBlock{
+			Type:  "tool_use",
+			ID:    tc.ID,
+			Name:  tc.Function.Name,
+			Input: tc.Function.Arguments,
+		})
+	}
+
+	stopReason := mapStopReason(r.DoneReason, len(r.Message.ToolCalls) > 0)
+
+	return MessagesResponse{
+		ID:         id,
+		Type:       "message",
+		Role:       "assistant",
+		Model:      r.Model,
+		Content:    content,
+		StopReason: stopReason,
+		Usage: Usage{
+			InputTokens:  r.Metrics.PromptEvalCount,
+			OutputTokens: r.Metrics.EvalCount,
+		},
+	}
+}
+
+// mapStopReason converts Ollama done_reason to Anthropic stop_reason
+func mapStopReason(reason string, hasToolCalls bool) string {
+	if hasToolCalls {
+		return "tool_use"
+	}
+
+	switch reason {
+	case "stop":
+		return "end_turn"
+	case "length":
+		return "max_tokens"
+	default:
+		if reason != "" {
+			return "stop_sequence"
+		}
+		return ""
+	}
+}
+
+// StreamConverter manages state for converting Ollama streaming responses to Anthropic format
+type StreamConverter struct {
+	ID              string
+	Model           string
+	firstWrite      bool
+	contentIndex    int
+	inputTokens     int
+	outputTokens    int
+	thinkingStarted bool
+	thinkingDone    bool
+	textStarted     bool
+	toolCallsSent   map[string]bool
+}
+
+func NewStreamConverter(id, model string) *StreamConverter {
+	return &StreamConverter{
+		ID:            id,
+		Model:         model,
+		firstWrite:    true,
+		toolCallsSent: make(map[string]bool),
+	}
+}
+
+// StreamEvent represents a streaming event to be sent to the client
+type StreamEvent struct {
+	Event string
+	Data  any
+}
+
+// Process converts an Ollama ChatResponse to Anthropic streaming events
+func (c *StreamConverter) Process(r api.ChatResponse) []StreamEvent {
+	var events []StreamEvent
+
+	if c.firstWrite {
+		c.firstWrite = false
+		c.inputTokens = r.Metrics.PromptEvalCount
+
+		events = append(events, StreamEvent{
+			Event: "message_start",
+			Data: MessageStartEvent{
+				Type: "message_start",
+				Message: MessagesResponse{
+					ID:      c.ID,
+					Type:    "message",
+					Role:    "assistant",
+					Model:   c.Model,
+					Content: []ContentBlock{},
+					Usage: Usage{
+						InputTokens:  c.inputTokens,
+						OutputTokens: 0,
+					},
+				},
+			},
+		})
+	}
+
+	if r.Message.Thinking != "" && !c.thinkingDone {
+		if !c.thinkingStarted {
+			c.thinkingStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type:     "thinking",
+						Thinking: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:     "thinking_delta",
+					Thinking: r.Message.Thinking,
+				},
+			},
+		})
+	}
+
+	if r.Message.Content != "" {
+		if c.thinkingStarted && !c.thinkingDone {
+			c.thinkingDone = true
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+		}
+
+		if !c.textStarted {
+			c.textStarted = true
+			events = append(events, StreamEvent{
+				Event: "content_block_start",
+				Data: ContentBlockStartEvent{
+					Type:  "content_block_start",
+					Index: c.contentIndex,
+					ContentBlock: ContentBlock{
+						Type: "text",
+						Text: ptr(""),
+					},
+				},
+			})
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type: "text_delta",
+					Text: r.Message.Content,
+				},
+			},
+		})
+	}
+
+	for _, tc := range r.Message.ToolCalls {
+		if c.toolCallsSent[tc.ID] {
+			continue
+		}
+
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+			c.contentIndex++
+			c.textStarted = false
+		}
+
+		argsJSON, err := json.Marshal(tc.Function.Arguments)
+		if err != nil {
+			slog.Error("failed to marshal tool arguments", "error", err, "tool_id", tc.ID)
+			continue
+		}
+
+		events = append(events, StreamEvent{
+			Event: "content_block_start",
+			Data: ContentBlockStartEvent{
+				Type:  "content_block_start",
+				Index: c.contentIndex,
+				ContentBlock: ContentBlock{
+					Type:  "tool_use",
+					ID:    tc.ID,
+					Name:  tc.Function.Name,
+					Input: map[string]any{},
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_delta",
+			Data: ContentBlockDeltaEvent{
+				Type:  "content_block_delta",
+				Index: c.contentIndex,
+				Delta: Delta{
+					Type:        "input_json_delta",
+					PartialJSON: string(argsJSON),
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "content_block_stop",
+			Data: ContentBlockStopEvent{
+				Type:  "content_block_stop",
+				Index: c.contentIndex,
+			},
+		})
+
+		c.toolCallsSent[tc.ID] = true
+		c.contentIndex++
+	}
+
+	if r.Done {
+		if c.textStarted {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		} else if c.thinkingStarted && !c.thinkingDone {
+			events = append(events, StreamEvent{
+				Event: "content_block_stop",
+				Data: ContentBlockStopEvent{
+					Type:  "content_block_stop",
+					Index: c.contentIndex,
+				},
+			})
+		}
+
+		c.outputTokens = r.Metrics.EvalCount
+		stopReason := mapStopReason(r.DoneReason, len(c.toolCallsSent) > 0)
+
+		events = append(events, StreamEvent{
+			Event: "message_delta",
+			Data: MessageDeltaEvent{
+				Type: "message_delta",
+				Delta: MessageDelta{
+					StopReason: stopReason,
+				},
+				Usage: DeltaUsage{
+					OutputTokens: c.outputTokens,
+				},
+			},
+		})
+
+		events = append(events, StreamEvent{
+			Event: "message_stop",
+			Data: MessageStopEvent{
+				Type: "message_stop",
+			},
+		})
+	}
+
+	return events
+}
+
+// generateID generates a unique ID with the given prefix using crypto/rand
+func generateID(prefix string) string {
+	b := make([]byte, 12)
+	if _, err := rand.Read(b); err != nil {
+		// Fallback to time-based ID if crypto/rand fails
+		return fmt.Sprintf("%s_%d", prefix, time.Now().UnixNano())
+	}
+	return fmt.Sprintf("%s_%x", prefix, b)
+}
+
+// GenerateMessageID generates a unique message ID
+func GenerateMessageID() string {
+	return generateID("msg")
+}
+
+// ptr returns a pointer to the given string value
+func ptr(s string) *string {
+	return &s
+}
+
+// mapToArgs converts a map to ToolCallFunctionArguments
+func mapToArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}

anthropic/anthropic_test.go

@@ -0,0 +1,953 @@
+package anthropic
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+
+	"github.com/google/go-cmp/cmp"
+
+	"github.com/ollama/ollama/api"
+)
+
+const (
+	testImage = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
+)
+
+// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests)
+func testArgs(m map[string]any) api.ToolCallFunctionArguments {
+	args := api.NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}
+
+func TestFromMessagesRequest_Basic(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Model != "test-model" {
+		t.Errorf("expected model 'test-model', got %q", result.Model)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "user" || result.Messages[0].Content != "Hello" {
+		t.Errorf("unexpected message: %+v", result.Messages[0])
+	}
+
+	if numPredict, ok := result.Options["num_predict"].(int); !ok || numPredict != 1024 {
+		t.Errorf("expected num_predict 1024, got %v", result.Options["num_predict"])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPrompt(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System:    "You are a helpful assistant.",
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Role != "system" || result.Messages[0].Content != "You are a helpful assistant." {
+		t.Errorf("unexpected system message: %+v", result.Messages[0])
+	}
+}
+
+func TestFromMessagesRequest_WithSystemPromptArray(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		System: []any{
+			map[string]any{"type": "text", "text": "You are helpful."},
+			map[string]any{"type": "text", "text": " Be concise."},
+		},
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "You are helpful. Be concise." {
+		t.Errorf("unexpected system message content: %q", result.Messages[0].Content)
+	}
+}
+
+func TestFromMessagesRequest_WithOptions(t *testing.T) {
+	temp := 0.7
+	topP := 0.9
+	topK := 40
+	req := MessagesRequest{
+		Model:         "test-model",
+		MaxTokens:     2048,
+		Messages:      []MessageParam{{Role: "user", Content: "Hello"}},
+		Temperature:   &temp,
+		TopP:          &topP,
+		TopK:          &topK,
+		StopSequences: []string{"\n", "END"},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Options["temperature"] != 0.7 {
+		t.Errorf("expected temperature 0.7, got %v", result.Options["temperature"])
+	}
+	if result.Options["top_p"] != 0.9 {
+		t.Errorf("expected top_p 0.9, got %v", result.Options["top_p"])
+	}
+	if result.Options["top_k"] != 40 {
+		t.Errorf("expected top_k 40, got %v", result.Options["top_k"])
+	}
+	if diff := cmp.Diff([]string{"\n", "END"}, result.Options["stop"]); diff != "" {
+		t.Errorf("stop sequences mismatch: %s", diff)
+	}
+}
+
+func TestFromMessagesRequest_WithImage(t *testing.T) {
+	imgData, _ := base64.StdEncoding.DecodeString(testImage)
+
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{"type": "text", "text": "What's in this image?"},
+					map[string]any{
+						"type": "image",
+						"source": map[string]any{
+							"type":       "base64",
+							"media_type": "image/png",
+							"data":       testImage,
+						},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	if result.Messages[0].Content != "What's in this image?" {
+		t.Errorf("expected content 'What's in this image?', got %q", result.Messages[0].Content)
+	}
+
+	if len(result.Messages[0].Images) != 1 {
+		t.Fatalf("expected 1 image, got %d", len(result.Messages[0].Images))
+	}
+
+	if string(result.Messages[0].Images[0]) != string(imgData) {
+		t.Error("image data mismatch")
+	}
+}
+
+func TestFromMessagesRequest_WithToolUse(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "What's the weather in Paris?"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":  "tool_use",
+						"id":    "call_123",
+						"name":  "get_weather",
+						"input": map[string]any{"location": "Paris"},
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	if len(result.Messages[1].ToolCalls) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(result.Messages[1].ToolCalls))
+	}
+
+	tc := result.Messages[1].ToolCalls[0]
+	if tc.ID != "call_123" {
+		t.Errorf("expected tool call ID 'call_123', got %q", tc.ID)
+	}
+	if tc.Function.Name != "get_weather" {
+		t.Errorf("expected tool name 'get_weather', got %q", tc.Function.Name)
+	}
+}
+
+func TestFromMessagesRequest_WithToolResult(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "user",
+				Content: []any{
+					map[string]any{
+						"type":        "tool_result",
+						"tool_use_id": "call_123",
+						"content":     "The weather in Paris is sunny, 22°C",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 1 {
+		t.Fatalf("expected 1 message, got %d", len(result.Messages))
+	}
+
+	msg := result.Messages[0]
+	if msg.Role != "tool" {
+		t.Errorf("expected role 'tool', got %q", msg.Role)
+	}
+	if msg.ToolCallID != "call_123" {
+		t.Errorf("expected tool_call_id 'call_123', got %q", msg.ToolCallID)
+	}
+	if msg.Content != "The weather in Paris is sunny, 22°C" {
+		t.Errorf("unexpected content: %q", msg.Content)
+	}
+}
+
+func TestFromMessagesRequest_WithTools(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "get_weather",
+				Description: "Get current weather",
+				InputSchema: json.RawMessage(`{"type":"object","properties":{"location":{"type":"string"}},"required":["location"]}`),
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Tools) != 1 {
+		t.Fatalf("expected 1 tool, got %d", len(result.Tools))
+	}
+
+	tool := result.Tools[0]
+	if tool.Type != "function" {
+		t.Errorf("expected type 'function', got %q", tool.Type)
+	}
+	if tool.Function.Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", tool.Function.Name)
+	}
+	if tool.Function.Description != "Get current weather" {
+		t.Errorf("expected description 'Get current weather', got %q", tool.Function.Description)
+	}
+}
+
+func TestFromMessagesRequest_WithThinking(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Thinking:  &ThinkingConfig{Type: "enabled", BudgetTokens: 1000},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if result.Think == nil {
+		t.Fatal("expected Think to be set")
+	}
+	if v, ok := result.Think.Value.(bool); !ok || !v {
+		t.Errorf("expected Think.Value to be true, got %v", result.Think.Value)
+	}
+}
+
+// TestFromMessagesRequest_ThinkingOnlyBlock verifies that messages containing only
+// a thinking block (no text, images, or tool calls) are preserved and not dropped.
+func TestFromMessagesRequest_ThinkingOnlyBlock(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{Role: "user", Content: "Hello"},
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type":     "thinking",
+						"thinking": "Let me think about this...",
+					},
+				},
+			},
+		},
+	}
+
+	result, err := FromMessagesRequest(req)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if len(result.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(result.Messages))
+	}
+
+	assistantMsg := result.Messages[1]
+	if assistantMsg.Thinking != "Let me think about this..." {
+		t.Errorf("expected thinking content, got %q", assistantMsg.Thinking)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingID(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"name": "get_weather",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use id")
+	}
+	if err.Error() != "tool_use block missing required 'id' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_ToolUseMissingName(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages: []MessageParam{
+			{
+				Role: "assistant",
+				Content: []any{
+					map[string]any{
+						"type": "tool_use",
+						"id":   "call_123",
+					},
+				},
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for missing tool_use name")
+	}
+	if err.Error() != "tool_use block missing required 'name' field" {
+		t.Errorf("unexpected error message: %v", err)
+	}
+}
+
+func TestFromMessagesRequest_InvalidToolSchema(t *testing.T) {
+	req := MessagesRequest{
+		Model:     "test-model",
+		MaxTokens: 1024,
+		Messages:  []MessageParam{{Role: "user", Content: "Hello"}},
+		Tools: []Tool{
+			{
+				Name:        "bad_tool",
+				InputSchema: json.RawMessage(`{invalid json`),
+			},
+		},
+	}
+
+	_, err := FromMessagesRequest(req)
+	if err == nil {
+		t.Fatal("expected error for invalid tool schema")
+	}
+}
+
+func TestToMessagesResponse_Basic(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello there!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics: api.Metrics{
+			PromptEvalCount: 10,
+			EvalCount:       5,
+		},
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if result.ID != "msg_123" {
+		t.Errorf("expected ID 'msg_123', got %q", result.ID)
+	}
+	if result.Type != "message" {
+		t.Errorf("expected type 'message', got %q", result.Type)
+	}
+	if result.Role != "assistant" {
+		t.Errorf("expected role 'assistant', got %q", result.Role)
+	}
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "text" || result.Content[0].Text == nil || *result.Content[0].Text != "Hello there!" {
+		t.Errorf("unexpected content: %+v", result.Content[0])
+	}
+	if result.StopReason != "end_turn" {
+		t.Errorf("expected stop_reason 'end_turn', got %q", result.StopReason)
+	}
+	if result.Usage.InputTokens != 10 || result.Usage.OutputTokens != 5 {
+		t.Errorf("unexpected usage: %+v", result.Usage)
+	}
+}
+
+func TestToMessagesResponse_WithToolCalls(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 1 {
+		t.Fatalf("expected 1 content block, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "tool_use" {
+		t.Errorf("expected type 'tool_use', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].ID != "call_123" {
+		t.Errorf("expected ID 'call_123', got %q", result.Content[0].ID)
+	}
+	if result.Content[0].Name != "get_weather" {
+		t.Errorf("expected name 'get_weather', got %q", result.Content[0].Name)
+	}
+	if result.StopReason != "tool_use" {
+		t.Errorf("expected stop_reason 'tool_use', got %q", result.StopReason)
+	}
+}
+
+func TestToMessagesResponse_WithThinking(t *testing.T) {
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:     "assistant",
+			Content:  "The answer is 42.",
+			Thinking: "Let me think about this...",
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	result := ToMessagesResponse("msg_123", resp)
+
+	if len(result.Content) != 2 {
+		t.Fatalf("expected 2 content blocks, got %d", len(result.Content))
+	}
+	if result.Content[0].Type != "thinking" {
+		t.Errorf("expected first block type 'thinking', got %q", result.Content[0].Type)
+	}
+	if result.Content[0].Thinking == nil || *result.Content[0].Thinking != "Let me think about this..." {
+		t.Errorf("unexpected thinking content: %v", result.Content[0].Thinking)
+	}
+	if result.Content[1].Type != "text" {
+		t.Errorf("expected second block type 'text', got %q", result.Content[1].Type)
+	}
+}
+
+func TestMapStopReason(t *testing.T) {
+	tests := []struct {
+		reason       string
+		hasToolCalls bool
+		want         string
+	}{
+		{"stop", false, "end_turn"},
+		{"length", false, "max_tokens"},
+		{"stop", true, "tool_use"},
+		{"other", false, "stop_sequence"},
+		{"", false, ""},
+	}
+
+	for _, tt := range tests {
+		got := mapStopReason(tt.reason, tt.hasToolCalls)
+		if got != tt.want {
+			t.Errorf("mapStopReason(%q, %v) = %q, want %q", tt.reason, tt.hasToolCalls, got, tt.want)
+		}
+	}
+}
+
+func TestNewError(t *testing.T) {
+	tests := []struct {
+		code int
+		want string
+	}{
+		{400, "invalid_request_error"},
+		{401, "authentication_error"},
+		{403, "permission_error"},
+		{404, "not_found_error"},
+		{429, "rate_limit_error"},
+		{500, "api_error"},
+		{503, "overloaded_error"},
+		{529, "overloaded_error"},
+	}
+
+	for _, tt := range tests {
+		result := NewError(tt.code, "test message")
+		if result.Type != "error" {
+			t.Errorf("NewError(%d) type = %q, want 'error'", tt.code, result.Type)
+		}
+		if result.Error.Type != tt.want {
+			t.Errorf("NewError(%d) error.type = %q, want %q", tt.code, result.Error.Type, tt.want)
+		}
+		if result.Error.Message != "test message" {
+			t.Errorf("NewError(%d) message = %q, want 'test message'", tt.code, result.Error.Message)
+		}
+		if result.RequestID == "" {
+			t.Errorf("NewError(%d) request_id should not be empty", tt.code)
+		}
+	}
+}
+
+func TestGenerateMessageID(t *testing.T) {
+	id1 := GenerateMessageID()
+	id2 := GenerateMessageID()
+
+	if id1 == "" {
+		t.Error("GenerateMessageID returned empty string")
+	}
+	if id1 == id2 {
+		t.Error("GenerateMessageID returned duplicate IDs")
+	}
+	if len(id1) < 10 {
+		t.Errorf("GenerateMessageID returned short ID: %q", id1)
+	}
+	if id1[:4] != "msg_" {
+		t.Errorf("GenerateMessageID should start with 'msg_', got %q", id1[:4])
+	}
+}
+
+func TestStreamConverter_Basic(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// First chunk
+	resp1 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: "Hello",
+		},
+		Metrics: api.Metrics{PromptEvalCount: 10},
+	}
+
+	events1 := conv.Process(resp1)
+	if len(events1) < 3 {
+		t.Fatalf("expected at least 3 events for first chunk, got %d", len(events1))
+	}
+
+	// Should have message_start, content_block_start, content_block_delta
+	if events1[0].Event != "message_start" {
+		t.Errorf("expected first event 'message_start', got %q", events1[0].Event)
+	}
+	if events1[1].Event != "content_block_start" {
+		t.Errorf("expected second event 'content_block_start', got %q", events1[1].Event)
+	}
+	if events1[2].Event != "content_block_delta" {
+		t.Errorf("expected third event 'content_block_delta', got %q", events1[2].Event)
+	}
+
+	// Final chunk
+	resp2 := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role:    "assistant",
+			Content: " world!",
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{EvalCount: 5},
+	}
+
+	events2 := conv.Process(resp2)
+
+	// Should have content_block_delta, content_block_stop, message_delta, message_stop
+	hasStop := false
+	for _, e := range events2 {
+		if e.Event == "message_stop" {
+			hasStop = true
+		}
+	}
+	if !hasStop {
+		t.Error("expected message_stop event in final chunk")
+	}
+}
+
+func TestStreamConverter_WithToolCalls(t *testing.T) {
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_123",
+					Function: api.ToolCallFunction{
+						Name:      "get_weather",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+		Metrics:    api.Metrics{PromptEvalCount: 10, EvalCount: 5},
+	}
+
+	events := conv.Process(resp)
+
+	hasToolStart := false
+	hasToolDelta := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					hasToolDelta = true
+				}
+			}
+		}
+	}
+
+	if !hasToolStart {
+		t.Error("expected tool_use content_block_start event")
+	}
+	if !hasToolDelta {
+		t.Error("expected input_json_delta event")
+	}
+}
+
+func TestStreamConverter_ToolCallWithUnmarshalableArgs(t *testing.T) {
+	// Test that unmarshalable arguments (like channels) are handled gracefully
+	// and don't cause a panic or corrupt stream
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	// Create a channel which cannot be JSON marshaled
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	// Should not panic and should skip the unmarshalable tool call
+	events := conv.Process(resp)
+
+	// Verify no tool_use block was started (since marshal failed before block start)
+	hasToolStart := false
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					hasToolStart = true
+				}
+			}
+		}
+	}
+
+	if hasToolStart {
+		t.Error("expected no tool_use block when arguments cannot be marshaled")
+	}
+}
+
+func TestStreamConverter_MultipleToolCallsWithMixedValidity(t *testing.T) {
+	// Test that valid tool calls still work when mixed with invalid ones
+	conv := NewStreamConverter("msg_123", "test-model")
+
+	unmarshalable := make(chan int)
+	badArgs := api.NewToolCallFunctionArguments()
+	badArgs.Set("channel", unmarshalable)
+
+	resp := api.ChatResponse{
+		Model: "test-model",
+		Message: api.Message{
+			Role: "assistant",
+			ToolCalls: []api.ToolCall{
+				{
+					ID: "call_good",
+					Function: api.ToolCallFunction{
+						Name:      "good_function",
+						Arguments: testArgs(map[string]any{"location": "Paris"}),
+					},
+				},
+				{
+					ID: "call_bad",
+					Function: api.ToolCallFunction{
+						Name:      "bad_function",
+						Arguments: badArgs,
+					},
+				},
+			},
+		},
+		Done:       true,
+		DoneReason: "stop",
+	}
+
+	events := conv.Process(resp)
+
+	// Count tool_use blocks - should only have 1 (the valid one)
+	toolStartCount := 0
+	toolDeltaCount := 0
+	for _, e := range events {
+		if e.Event == "content_block_start" {
+			if start, ok := e.Data.(ContentBlockStartEvent); ok {
+				if start.ContentBlock.Type == "tool_use" {
+					toolStartCount++
+					if start.ContentBlock.Name != "good_function" {
+						t.Errorf("expected tool name 'good_function', got %q", start.ContentBlock.Name)
+					}
+				}
+			}
+		}
+		if e.Event == "content_block_delta" {
+			if delta, ok := e.Data.(ContentBlockDeltaEvent); ok {
+				if delta.Delta.Type == "input_json_delta" {
+					toolDeltaCount++
+				}
+			}
+		}
+	}
+
+	if toolStartCount != 1 {
+		t.Errorf("expected 1 tool_use block, got %d", toolStartCount)
+	}
+	if toolDeltaCount != 1 {
+		t.Errorf("expected 1 input_json_delta, got %d", toolDeltaCount)
+	}
+}
+
+// TestContentBlockJSON_EmptyFieldsPresent verifies that empty text and thinking fields
+// are serialized in JSON output. The Anthropic SDK requires these fields to be present
+// (even when empty) in content_block_start events to properly accumulate streaming deltas.
+// Without these fields, the SDK throws: "TypeError: unsupported operand type(s) for +=: 'NoneType' and 'str'"
+func TestContentBlockJSON_EmptyFieldsPresent(t *testing.T) {
+	tests := []struct {
+		name     string
+		block    ContentBlock
+		wantKeys []string
+	}{
+		{
+			name: "text block includes empty text field",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr(""),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+		{
+			name: "thinking block includes empty thinking field",
+			block: ContentBlock{
+				Type:     "thinking",
+				Thinking: ptr(""),
+			},
+			wantKeys: []string{"type", "thinking"},
+		},
+		{
+			name: "text block with content",
+			block: ContentBlock{
+				Type: "text",
+				Text: ptr("hello"),
+			},
+			wantKeys: []string{"type", "text"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			data, err := json.Marshal(tt.block)
+			if err != nil {
+				t.Fatalf("failed to marshal: %v", err)
+			}
+
+			var result map[string]any
+			if err := json.Unmarshal(data, &result); err != nil {
+				t.Fatalf("failed to unmarshal: %v", err)
+			}
+
+			for _, key := range tt.wantKeys {
+				if _, ok := result[key]; !ok {
+					t.Errorf("expected key %q to be present in JSON output, got: %s", key, string(data))
+				}
+			}
+		})
+	}
+}
+
+// TestStreamConverter_ContentBlockStartIncludesEmptyFields verifies that content_block_start
+// events include the required empty fields for SDK compatibility.
+func TestStreamConverter_ContentBlockStartIncludesEmptyFields(t *testing.T) {
+	t.Run("text block start includes empty text", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Content: "hello"},
+		}
+
+		events := conv.Process(resp)
+
+		var foundTextStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "text" {
+						foundTextStart = true
+						// Marshal and verify the text field is present
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["text"]; !ok {
+							t.Error("content_block_start for text should include 'text' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundTextStart {
+			t.Error("expected text content_block_start event")
+		}
+	})
+
+	t.Run("thinking block start includes empty thinking", func(t *testing.T) {
+		conv := NewStreamConverter("msg_123", "test-model")
+
+		resp := api.ChatResponse{
+			Model:   "test-model",
+			Message: api.Message{Role: "assistant", Thinking: "let me think..."},
+		}
+
+		events := conv.Process(resp)
+
+		var foundThinkingStart bool
+		for _, e := range events {
+			if e.Event == "content_block_start" {
+				if start, ok := e.Data.(ContentBlockStartEvent); ok {
+					if start.ContentBlock.Type == "thinking" {
+						foundThinkingStart = true
+						data, _ := json.Marshal(start)
+						var result map[string]any
+						json.Unmarshal(data, &result)
+						cb := result["content_block"].(map[string]any)
+						if _, ok := cb["thinking"]; !ok {
+							t.Error("content_block_start for thinking should include 'thinking' field")
+						}
+					}
+				}
+			}
+		}
+
+		if !foundThinkingStart {
+			t.Error("expected thinking content_block_start event")
+		}
+	})
+}

api/client.go

@@ -165,7 +165,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 	return nil
 }
 
-const maxBufferSize = 512 * format.KiloByte
+const maxBufferSize = 8 * format.MegaByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
 	var buf io.Reader

api/types.go

@@ -3,6 +3,7 @@ package api
 import (
 	"encoding/json"
 	"fmt"
+	"iter"
 	"log/slog"
 	"math"
 	"os"
@@ -14,6 +15,7 @@ import (
 	"github.com/google/uuid"
 
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/internal/orderedmap"
 	"github.com/ollama/ollama/types/model"
 )
 
@@ -125,6 +127,20 @@ type GenerateRequest struct {
 	// each with an associated log probability. Only applies when Logprobs is true.
 	// Valid values are 0-20. Default is 0 (only return the selected token's logprob).
 	TopLogprobs int `json:"top_logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Width is the width of the generated image in pixels.
+	// Only used for image generation models.
+	Width int32 `json:"width,omitempty"`
+
+	// Height is the height of the generated image in pixels.
+	// Only used for image generation models.
+	Height int32 `json:"height,omitempty"`
+
+	// Steps is the number of diffusion steps for image generation.
+	// Only used for image generation models.
+	Steps int32 `json:"steps,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -227,13 +243,79 @@ type ToolCallFunction struct {
 	Arguments ToolCallFunctionArguments `json:"arguments"`
 }
 
-type ToolCallFunctionArguments map[string]any
+// ToolCallFunctionArguments holds tool call arguments in insertion order.
+type ToolCallFunctionArguments struct {
+	om *orderedmap.Map[string, any]
+}
+
+// NewToolCallFunctionArguments creates a new empty ToolCallFunctionArguments.
+func NewToolCallFunctionArguments() ToolCallFunctionArguments {
+	return ToolCallFunctionArguments{om: orderedmap.New[string, any]()}
+}
+
+// Get retrieves a value by key.
+func (t *ToolCallFunctionArguments) Get(key string) (any, bool) {
+	if t == nil || t.om == nil {
+		return nil, false
+	}
+	return t.om.Get(key)
+}
+
+// Set sets a key-value pair, preserving insertion order.
+func (t *ToolCallFunctionArguments) Set(key string, value any) {
+	if t == nil {
+		return
+	}
+	if t.om == nil {
+		t.om = orderedmap.New[string, any]()
+	}
+	t.om.Set(key, value)
+}
+
+// Len returns the number of arguments.
+func (t *ToolCallFunctionArguments) Len() int {
+	if t == nil || t.om == nil {
+		return 0
+	}
+	return t.om.Len()
+}
+
+// All returns an iterator over all key-value pairs in insertion order.
+func (t *ToolCallFunctionArguments) All() iter.Seq2[string, any] {
+	if t == nil || t.om == nil {
+		return func(yield func(string, any) bool) {}
+	}
+	return t.om.All()
+}
+
+// ToMap returns a regular map (order not preserved).
+func (t *ToolCallFunctionArguments) ToMap() map[string]any {
+	if t == nil || t.om == nil {
+		return nil
+	}
+	return t.om.ToMap()
+}
 
 func (t *ToolCallFunctionArguments) String() string {
-	bts, _ := json.Marshal(t)
+	if t == nil || t.om == nil {
+		return "{}"
+	}
+	bts, _ := json.Marshal(t.om)
 	return string(bts)
 }
 
+func (t *ToolCallFunctionArguments) UnmarshalJSON(data []byte) error {
+	t.om = orderedmap.New[string, any]()
+	return json.Unmarshal(data, t.om)
+}
+
+func (t ToolCallFunctionArguments) MarshalJSON() ([]byte, error) {
+	if t.om == nil {
+		return []byte("{}"), nil
+	}
+	return json.Marshal(t.om)
+}
+
 type Tool struct {
 	Type     string       `json:"type"`
 	Items    any          `json:"items,omitempty"`
@@ -282,13 +364,78 @@ func (pt PropertyType) String() string {
 	return fmt.Sprintf("%v", []string(pt))
 }
 
+// ToolPropertiesMap holds tool properties in insertion order.
+type ToolPropertiesMap struct {
+	om *orderedmap.Map[string, ToolProperty]
+}
+
+// NewToolPropertiesMap creates a new empty ToolPropertiesMap.
+func NewToolPropertiesMap() *ToolPropertiesMap {
+	return &ToolPropertiesMap{om: orderedmap.New[string, ToolProperty]()}
+}
+
+// Get retrieves a property by name.
+func (t *ToolPropertiesMap) Get(key string) (ToolProperty, bool) {
+	if t == nil || t.om == nil {
+		return ToolProperty{}, false
+	}
+	return t.om.Get(key)
+}
+
+// Set sets a property, preserving insertion order.
+func (t *ToolPropertiesMap) Set(key string, value ToolProperty) {
+	if t == nil {
+		return
+	}
+	if t.om == nil {
+		t.om = orderedmap.New[string, ToolProperty]()
+	}
+	t.om.Set(key, value)
+}
+
+// Len returns the number of properties.
+func (t *ToolPropertiesMap) Len() int {
+	if t == nil || t.om == nil {
+		return 0
+	}
+	return t.om.Len()
+}
+
+// All returns an iterator over all properties in insertion order.
+func (t *ToolPropertiesMap) All() iter.Seq2[string, ToolProperty] {
+	if t == nil || t.om == nil {
+		return func(yield func(string, ToolProperty) bool) {}
+	}
+	return t.om.All()
+}
+
+// ToMap returns a regular map (order not preserved).
+func (t *ToolPropertiesMap) ToMap() map[string]ToolProperty {
+	if t == nil || t.om == nil {
+		return nil
+	}
+	return t.om.ToMap()
+}
+
+func (t ToolPropertiesMap) MarshalJSON() ([]byte, error) {
+	if t.om == nil {
+		return []byte("null"), nil
+	}
+	return json.Marshal(t.om)
+}
+
+func (t *ToolPropertiesMap) UnmarshalJSON(data []byte) error {
+	t.om = orderedmap.New[string, ToolProperty]()
+	return json.Unmarshal(data, t.om)
+}
+
 type ToolProperty struct {
-	AnyOf       []ToolProperty          `json:"anyOf,omitempty"`
-	Type        PropertyType            `json:"type,omitempty"`
-	Items       any                     `json:"items,omitempty"`
-	Description string                  `json:"description,omitempty"`
-	Enum        []any                   `json:"enum,omitempty"`
-	Properties  map[string]ToolProperty `json:"properties,omitempty"`
+	AnyOf       []ToolProperty     `json:"anyOf,omitempty"`
+	Type        PropertyType       `json:"type,omitempty"`
+	Items       any                `json:"items,omitempty"`
+	Description string             `json:"description,omitempty"`
+	Enum        []any              `json:"enum,omitempty"`
+	Properties  *ToolPropertiesMap `json:"properties,omitempty"`
 }
 
 // ToTypeScriptType converts a ToolProperty to a TypeScript type string
@@ -337,11 +484,11 @@ func mapToTypeScriptType(jsonType string) string {
 }
 
 type ToolFunctionParameters struct {
-	Type       string                  `json:"type"`
-	Defs       any                     `json:"$defs,omitempty"`
-	Items      any                     `json:"items,omitempty"`
-	Required   []string                `json:"required,omitempty"`
-	Properties map[string]ToolProperty `json:"properties"`
+	Type       string             `json:"type"`
+	Defs       any                `json:"$defs,omitempty"`
+	Items      any                `json:"items,omitempty"`
+	Required   []string           `json:"required,omitempty"`
+	Properties *ToolPropertiesMap `json:"properties"`
 }
 
 func (t *ToolFunctionParameters) String() string {
@@ -602,7 +749,7 @@ type ShowResponse struct {
 	Messages      []Message          `json:"messages,omitempty"`
 	RemoteModel   string             `json:"remote_model,omitempty"`
 	RemoteHost    string             `json:"remote_host,omitempty"`
-	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info"`
 	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
 	Tensors       []Tensor           `json:"tensors,omitempty"`
 	Capabilities  []model.Capability `json:"capabilities,omitempty"`
@@ -727,6 +874,20 @@ type GenerateResponse struct {
 	// Logprobs contains log probability information for the generated tokens,
 	// if requested via the Logprobs parameter.
 	Logprobs []Logprob `json:"logprobs,omitempty"`
+
+	// Experimental: Image generation fields (may change or be removed)
+
+	// Image contains a base64-encoded generated image.
+	// Only present for image generation models.
+	Image string `json:"image,omitempty"`
+
+	// Completed is the number of completed steps in image generation.
+	// Only present for image generation models during streaming.
+	Completed int64 `json:"completed,omitempty"`
+
+	// Total is the total number of steps for image generation.
+	// Only present for image generation models during streaming.
+	Total int64 `json:"total,omitempty"`
 }
 
 // ModelDetails provides details about a model.

api/types_test.go

@@ -11,6 +11,24 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
+// testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests, order not preserved)
+func testPropsMap(m map[string]ToolProperty) *ToolPropertiesMap {
+	props := NewToolPropertiesMap()
+	for k, v := range m {
+		props.Set(k, v)
+	}
+	return props
+}
+
+// testArgs creates ToolCallFunctionArguments from a map (convenience function for tests, order not preserved)
+func testArgs(m map[string]any) ToolCallFunctionArguments {
+	args := NewToolCallFunctionArguments()
+	for k, v := range m {
+		args.Set(k, v)
+	}
+	return args
+}
+
 func TestKeepAliveParsingFromJSON(t *testing.T) {
 	tests := []struct {
 		name string
@@ -309,9 +327,9 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 			input: ToolFunctionParameters{
 				Type:     "object",
 				Required: []string{"name"},
-				Properties: map[string]ToolProperty{
+				Properties: testPropsMap(map[string]ToolProperty{
 					"name": {Type: PropertyType{"string"}},
-				},
+				}),
 			},
 			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string"}}}`,
 		},
@@ -319,9 +337,9 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 			name: "no required",
 			input: ToolFunctionParameters{
 				Type: "object",
-				Properties: map[string]ToolProperty{
+				Properties: testPropsMap(map[string]ToolProperty{
 					"name": {Type: PropertyType{"string"}},
-				},
+				}),
 			},
 			expected: `{"type":"object","properties":{"name":{"type":"string"}}}`,
 		},
@@ -339,7 +357,7 @@ func TestToolFunctionParameters_MarshalJSON(t *testing.T) {
 func TestToolCallFunction_IndexAlwaysMarshals(t *testing.T) {
 	fn := ToolCallFunction{
 		Name:      "echo",
-		Arguments: ToolCallFunctionArguments{"message": "hi"},
+		Arguments: testArgs(map[string]any{"message": "hi"}),
 	}
 
 	data, err := json.Marshal(fn)
@@ -529,7 +547,7 @@ func TestToolPropertyNestedProperties(t *testing.T) {
 			expected: ToolProperty{
 				Type:        PropertyType{"object"},
 				Description: "Location details",
-				Properties: map[string]ToolProperty{
+				Properties: testPropsMap(map[string]ToolProperty{
 					"address": {
 						Type:        PropertyType{"string"},
 						Description: "Street address",
@@ -538,7 +556,7 @@ func TestToolPropertyNestedProperties(t *testing.T) {
 						Type:        PropertyType{"string"},
 						Description: "City name",
 					},
-				},
+				}),
 			},
 		},
 		{
@@ -566,22 +584,22 @@ func TestToolPropertyNestedProperties(t *testing.T) {
 			expected: ToolProperty{
 				Type:        PropertyType{"object"},
 				Description: "Event",
-				Properties: map[string]ToolProperty{
+				Properties: testPropsMap(map[string]ToolProperty{
 					"location": {
 						Type:        PropertyType{"object"},
 						Description: "Location",
-						Properties: map[string]ToolProperty{
+						Properties: testPropsMap(map[string]ToolProperty{
 							"coordinates": {
 								Type:        PropertyType{"object"},
 								Description: "GPS coordinates",
-								Properties: map[string]ToolProperty{
+								Properties: testPropsMap(map[string]ToolProperty{
 									"lat": {Type: PropertyType{"number"}, Description: "Latitude"},
 									"lng": {Type: PropertyType{"number"}, Description: "Longitude"},
-								},
+								}),
 							},
-						},
+						}),
 					},
-				},
+				}),
 			},
 		},
 	}
@@ -591,7 +609,13 @@ func TestToolPropertyNestedProperties(t *testing.T) {
 			var prop ToolProperty
 			err := json.Unmarshal([]byte(tt.input), &prop)
 			require.NoError(t, err)
-			assert.Equal(t, tt.expected, prop)
+
+			// Compare JSON representations since pointer comparison doesn't work
+			expectedJSON, err := json.Marshal(tt.expected)
+			require.NoError(t, err)
+			actualJSON, err := json.Marshal(prop)
+			require.NoError(t, err)
+			assert.JSONEq(t, string(expectedJSON), string(actualJSON))
 
 			// Round-trip test: marshal and unmarshal again
 			data, err := json.Marshal(prop)
@@ -600,7 +624,10 @@ func TestToolPropertyNestedProperties(t *testing.T) {
 			var prop2 ToolProperty
 			err = json.Unmarshal(data, &prop2)
 			require.NoError(t, err)
-			assert.Equal(t, tt.expected, prop2)
+
+			prop2JSON, err := json.Marshal(prop2)
+			require.NoError(t, err)
+			assert.JSONEq(t, string(expectedJSON), string(prop2JSON))
 		})
 	}
 }
@@ -616,12 +643,12 @@ func TestToolFunctionParameters_String(t *testing.T) {
 			params: ToolFunctionParameters{
 				Type:     "object",
 				Required: []string{"name"},
-				Properties: map[string]ToolProperty{
+				Properties: testPropsMap(map[string]ToolProperty{
 					"name": {
 						Type:        PropertyType{"string"},
 						Description: "The name of the person",
 					},
-				},
+				}),
 			},
 			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string","description":"The name of the person"}}}`,
 		},
@@ -638,7 +665,7 @@ func TestToolFunctionParameters_String(t *testing.T) {
 					s.Self = s
 					return s
 				}(),
-				Properties: map[string]ToolProperty{},
+				Properties: testPropsMap(map[string]ToolProperty{}),
 			},
 			expected: "",
 		},
@@ -651,3 +678,235 @@ func TestToolFunctionParameters_String(t *testing.T) {
 		})
 	}
 }
+
+func TestToolCallFunctionArguments_OrderPreservation(t *testing.T) {
+	t.Run("marshal preserves insertion order", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		args.Set("zebra", "z")
+		args.Set("apple", "a")
+		args.Set("mango", "m")
+
+		data, err := json.Marshal(args)
+		require.NoError(t, err)
+
+		// Should preserve insertion order, not alphabetical
+		assert.Equal(t, `{"zebra":"z","apple":"a","mango":"m"}`, string(data))
+	})
+
+	t.Run("unmarshal preserves JSON order", func(t *testing.T) {
+		jsonData := `{"zebra":"z","apple":"a","mango":"m"}`
+
+		var args ToolCallFunctionArguments
+		err := json.Unmarshal([]byte(jsonData), &args)
+		require.NoError(t, err)
+
+		// Verify iteration order matches JSON order
+		var keys []string
+		for k := range args.All() {
+			keys = append(keys, k)
+		}
+		assert.Equal(t, []string{"zebra", "apple", "mango"}, keys)
+	})
+
+	t.Run("round trip preserves order", func(t *testing.T) {
+		original := `{"z":1,"a":2,"m":3,"b":4}`
+
+		var args ToolCallFunctionArguments
+		err := json.Unmarshal([]byte(original), &args)
+		require.NoError(t, err)
+
+		data, err := json.Marshal(args)
+		require.NoError(t, err)
+
+		assert.Equal(t, original, string(data))
+	})
+
+	t.Run("String method returns ordered JSON", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		args.Set("c", 3)
+		args.Set("a", 1)
+		args.Set("b", 2)
+
+		assert.Equal(t, `{"c":3,"a":1,"b":2}`, args.String())
+	})
+
+	t.Run("Get retrieves correct values", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		args.Set("key1", "value1")
+		args.Set("key2", 42)
+
+		v, ok := args.Get("key1")
+		assert.True(t, ok)
+		assert.Equal(t, "value1", v)
+
+		v, ok = args.Get("key2")
+		assert.True(t, ok)
+		assert.Equal(t, 42, v)
+
+		_, ok = args.Get("nonexistent")
+		assert.False(t, ok)
+	})
+
+	t.Run("Len returns correct count", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		assert.Equal(t, 0, args.Len())
+
+		args.Set("a", 1)
+		assert.Equal(t, 1, args.Len())
+
+		args.Set("b", 2)
+		assert.Equal(t, 2, args.Len())
+	})
+
+	t.Run("empty args marshal to empty object", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		data, err := json.Marshal(args)
+		require.NoError(t, err)
+		assert.Equal(t, `{}`, string(data))
+	})
+
+	t.Run("zero value args marshal to empty object", func(t *testing.T) {
+		var args ToolCallFunctionArguments
+		assert.Equal(t, "{}", args.String())
+	})
+}
+
+func TestToolPropertiesMap_OrderPreservation(t *testing.T) {
+	t.Run("marshal preserves insertion order", func(t *testing.T) {
+		props := NewToolPropertiesMap()
+		props.Set("zebra", ToolProperty{Type: PropertyType{"string"}})
+		props.Set("apple", ToolProperty{Type: PropertyType{"number"}})
+		props.Set("mango", ToolProperty{Type: PropertyType{"boolean"}})
+
+		data, err := json.Marshal(props)
+		require.NoError(t, err)
+
+		// Should preserve insertion order, not alphabetical
+		expected := `{"zebra":{"type":"string"},"apple":{"type":"number"},"mango":{"type":"boolean"}}`
+		assert.Equal(t, expected, string(data))
+	})
+
+	t.Run("unmarshal preserves JSON order", func(t *testing.T) {
+		jsonData := `{"zebra":{"type":"string"},"apple":{"type":"number"},"mango":{"type":"boolean"}}`
+
+		var props ToolPropertiesMap
+		err := json.Unmarshal([]byte(jsonData), &props)
+		require.NoError(t, err)
+
+		// Verify iteration order matches JSON order
+		var keys []string
+		for k := range props.All() {
+			keys = append(keys, k)
+		}
+		assert.Equal(t, []string{"zebra", "apple", "mango"}, keys)
+	})
+
+	t.Run("round trip preserves order", func(t *testing.T) {
+		original := `{"z":{"type":"string"},"a":{"type":"number"},"m":{"type":"boolean"}}`
+
+		var props ToolPropertiesMap
+		err := json.Unmarshal([]byte(original), &props)
+		require.NoError(t, err)
+
+		data, err := json.Marshal(props)
+		require.NoError(t, err)
+
+		assert.Equal(t, original, string(data))
+	})
+
+	t.Run("Get retrieves correct values", func(t *testing.T) {
+		props := NewToolPropertiesMap()
+		props.Set("name", ToolProperty{Type: PropertyType{"string"}, Description: "The name"})
+		props.Set("age", ToolProperty{Type: PropertyType{"integer"}, Description: "The age"})
+
+		v, ok := props.Get("name")
+		assert.True(t, ok)
+		assert.Equal(t, "The name", v.Description)
+
+		v, ok = props.Get("age")
+		assert.True(t, ok)
+		assert.Equal(t, "The age", v.Description)
+
+		_, ok = props.Get("nonexistent")
+		assert.False(t, ok)
+	})
+
+	t.Run("Len returns correct count", func(t *testing.T) {
+		props := NewToolPropertiesMap()
+		assert.Equal(t, 0, props.Len())
+
+		props.Set("a", ToolProperty{})
+		assert.Equal(t, 1, props.Len())
+
+		props.Set("b", ToolProperty{})
+		assert.Equal(t, 2, props.Len())
+	})
+
+	t.Run("nil props marshal to null", func(t *testing.T) {
+		var props *ToolPropertiesMap
+		data, err := json.Marshal(props)
+		require.NoError(t, err)
+		assert.Equal(t, `null`, string(data))
+	})
+
+	t.Run("ToMap returns regular map", func(t *testing.T) {
+		props := NewToolPropertiesMap()
+		props.Set("a", ToolProperty{Type: PropertyType{"string"}})
+		props.Set("b", ToolProperty{Type: PropertyType{"number"}})
+
+		m := props.ToMap()
+		assert.Equal(t, 2, len(m))
+		assert.Equal(t, PropertyType{"string"}, m["a"].Type)
+		assert.Equal(t, PropertyType{"number"}, m["b"].Type)
+	})
+}
+
+func TestToolCallFunctionArguments_ComplexValues(t *testing.T) {
+	t.Run("nested objects preserve order", func(t *testing.T) {
+		jsonData := `{"outer":{"z":1,"a":2},"simple":"value"}`
+
+		var args ToolCallFunctionArguments
+		err := json.Unmarshal([]byte(jsonData), &args)
+		require.NoError(t, err)
+
+		// Outer keys should be in order
+		var keys []string
+		for k := range args.All() {
+			keys = append(keys, k)
+		}
+		assert.Equal(t, []string{"outer", "simple"}, keys)
+	})
+
+	t.Run("arrays as values", func(t *testing.T) {
+		args := NewToolCallFunctionArguments()
+		args.Set("items", []string{"a", "b", "c"})
+		args.Set("numbers", []int{1, 2, 3})
+
+		data, err := json.Marshal(args)
+		require.NoError(t, err)
+
+		assert.Equal(t, `{"items":["a","b","c"],"numbers":[1,2,3]}`, string(data))
+	})
+}
+
+func TestToolPropertiesMap_NestedProperties(t *testing.T) {
+	t.Run("nested properties preserve order", func(t *testing.T) {
+		props := NewToolPropertiesMap()
+
+		nestedProps := NewToolPropertiesMap()
+		nestedProps.Set("z_field", ToolProperty{Type: PropertyType{"string"}})
+		nestedProps.Set("a_field", ToolProperty{Type: PropertyType{"number"}})
+
+		props.Set("outer", ToolProperty{
+			Type:       PropertyType{"object"},
+			Properties: nestedProps,
+		})
+
+		data, err := json.Marshal(props)
+		require.NoError(t, err)
+
+		// Both outer and inner should preserve order
+		expected := `{"outer":{"type":"object","properties":{"z_field":{"type":"string"},"a_field":{"type":"number"}}}}`
+		assert.Equal(t, expected, string(data))
+	})
+}

app/cmd/app/app_darwin.m

@@ -14,6 +14,7 @@ extern NSString *SystemWidePath;
 @interface AppDelegate () <NSWindowDelegate, WKNavigationDelegate, WKUIDelegate>
 @property(strong, nonatomic) NSStatusItem *statusItem;
 @property(assign, nonatomic) BOOL updateAvailable;
+@property(assign, nonatomic) BOOL systemShutdownInProgress;
 @end
 
 @implementation AppDelegate
@@ -40,6 +41,13 @@ bool firstTimeRun,startHidden; // Set in run before initialization
 }
 
 - (void)applicationDidFinishLaunching:(NSNotification *)aNotification {
+    // Register for system shutdown/restart notification so we can allow termination
+    [[[NSWorkspace sharedWorkspace] notificationCenter]
+        addObserver:self
+           selector:@selector(systemWillPowerOff:)
+               name:NSWorkspaceWillPowerOffNotification
+             object:nil];
+
     // if we're in development mode, set the app icon
     NSString *bundlePath = [[NSBundle mainBundle] bundlePath];
     if (![bundlePath hasSuffix:@".app"]) {
@@ -278,7 +286,18 @@ bool firstTimeRun,startHidden; // Set in run before initialization
     [NSApp activateIgnoringOtherApps:YES];
 }
 
+- (void)systemWillPowerOff:(NSNotification *)notification {
+    // Set flag so applicationShouldTerminate: knows to allow termination.
+    // The system will call applicationShouldTerminate: after posting this notification.
+    self.systemShutdownInProgress = YES;
+}
+
 - (NSApplicationTerminateReply)applicationShouldTerminate:(NSApplication *)sender {
+    // Allow termination if the system is shutting down or restarting
+    if (self.systemShutdownInProgress) {
+        return NSTerminateNow;
+    }
+    // Otherwise just hide the app (for Cmd+Q, close button, etc.)
     [NSApp hide:nil];
     [NSApp setActivationPolicy:NSApplicationActivationPolicyAccessory];
     return NSTerminateCancel;

app/ui/app/src/lib/highlighter.ts

@@ -147,6 +147,7 @@ export const highlighterPromise = createHighlighter({
     "c",
     "cpp",
     "sql",
+    "swift",
     "yaml",
     "markdown",
   ],

app/ui/ui.go

@@ -997,7 +997,7 @@ func (s *Server) chat(w http.ResponseWriter, r *http.Request) error {
 				for _, toolCall := range res.Message.ToolCalls {
 					// continues loop as tools were executed
 					toolsExecuted = true
-					result, content, err := registry.Execute(ctx, toolCall.Function.Name, toolCall.Function.Arguments)
+					result, content, err := registry.Execute(ctx, toolCall.Function.Name, toolCall.Function.Arguments.ToMap())
 					if err != nil {
 						errContent := fmt.Sprintf("Error: %v", err)
 						toolErrMsg := store.NewMessage("tool", errContent, nil)
@@ -1558,13 +1558,13 @@ func convertToOllamaTool(toolSchema map[string]any) api.Tool {
 
 	tool.Function.Parameters.Type = "object"
 	tool.Function.Parameters.Required = []string{}
-	tool.Function.Parameters.Properties = make(map[string]api.ToolProperty)
+	tool.Function.Parameters.Properties = api.NewToolPropertiesMap()
 
 	if schemaProps, ok := toolSchema["schema"].(map[string]any); ok {
 		tool.Function.Parameters.Type = getStringFromMap(schemaProps, "type", "object")
 
 		if props, ok := schemaProps["properties"].(map[string]any); ok {
-			tool.Function.Parameters.Properties = make(map[string]api.ToolProperty)
+			tool.Function.Parameters.Properties = api.NewToolPropertiesMap()
 
 			for propName, propDef := range props {
 				if propMap, ok := propDef.(map[string]any); ok {
@@ -1572,7 +1572,7 @@ func convertToOllamaTool(toolSchema map[string]any) api.Tool {
 						Type:        api.PropertyType{getStringFromMap(propMap, "type", "string")},
 						Description: getStringFromMap(propMap, "description", ""),
 					}
-					tool.Function.Parameters.Properties[propName] = prop
+					tool.Function.Parameters.Properties.Set(propName, prop)
 				}
 			}
 		}

cmd/cmd.go

@@ -45,6 +45,10 @@ import (
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/types/syncmap"
 	"github.com/ollama/ollama/version"
+	xcmd "github.com/ollama/ollama/x/cmd"
+	"github.com/ollama/ollama/x/create"
+	xcreateclient "github.com/ollama/ollama/x/create/client"
+	"github.com/ollama/ollama/x/imagegen"
 )
 
 const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
@@ -90,11 +94,88 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
+	// Validate model name early to fail fast
+	modelName := args[0]
+	name := model.ParseName(modelName)
+	if !name.IsValid() {
+		return fmt.Errorf("invalid model name: %s", modelName)
+	}
+
+	// Check for --experimental flag for safetensors model creation
+	experimental, _ := cmd.Flags().GetBool("experimental")
+	if experimental {
+		// Get Modelfile content - either from -f flag or default to "FROM ."
+		var reader io.Reader
+		filename, err := getModelfileName(cmd)
+		if os.IsNotExist(err) || filename == "" {
+			// No Modelfile specified or found - use default
+			reader = strings.NewReader("FROM .\n")
+		} else if err != nil {
+			return err
+		} else {
+			f, err := os.Open(filename)
+			if err != nil {
+				return err
+			}
+			defer f.Close()
+			reader = f
+		}
+
+		// Parse the Modelfile
+		modelfile, err := parser.ParseFile(reader)
+		if err != nil {
+			return fmt.Errorf("failed to parse Modelfile: %w", err)
+		}
+
+		// Extract FROM path and configuration
+		var modelDir string
+		mfConfig := &xcreateclient.ModelfileConfig{}
+
+		for _, cmd := range modelfile.Commands {
+			switch cmd.Name {
+			case "model":
+				modelDir = cmd.Args
+			case "template":
+				mfConfig.Template = cmd.Args
+			case "system":
+				mfConfig.System = cmd.Args
+			case "license":
+				mfConfig.License = cmd.Args
+			}
+		}
+
+		if modelDir == "" {
+			modelDir = "."
+		}
+
+		// Resolve relative paths based on Modelfile location
+		if !filepath.IsAbs(modelDir) && filename != "" {
+			modelDir = filepath.Join(filepath.Dir(filename), modelDir)
+		}
+
+		quantize, _ := cmd.Flags().GetString("quantize")
+		return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+			ModelName: modelName,
+			ModelDir:  modelDir,
+			Quantize:  quantize,
+			Modelfile: mfConfig,
+		}, p)
+	}
+
 	var reader io.Reader
 
 	filename, err := getModelfileName(cmd)
 	if os.IsNotExist(err) {
 		if filename == "" {
+			// No Modelfile found - check if current directory is an image gen model
+			if create.IsTensorModelDir(".") {
+				quantize, _ := cmd.Flags().GetString("quantize")
+				return xcreateclient.CreateModel(xcreateclient.CreateOptions{
+					ModelName: modelName,
+					ModelDir:  ".",
+					Quantize:  quantize,
+				}, p)
+			}
 			reader = strings.NewReader("FROM .\n")
 		} else {
 			return errModelfileNotFound
@@ -126,7 +207,7 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	}
 	spinner.Stop()
 
-	req.Model = args[0]
+	req.Model = modelName
 	quantize, _ := cmd.Flags().GetString("quantize")
 	if quantize != "" {
 		req.Quantize = quantize
@@ -456,6 +537,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	name := args[0]
+
 	info, err := func() (*api.ShowResponse, error) {
 		showReq := &api.ShowRequest{Name: name}
 		info, err := client.Show(cmd.Context(), showReq)
@@ -517,6 +599,19 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return generateEmbedding(cmd, name, opts.Prompt, opts.KeepAlive, truncate, dimensions)
 	}
 
+	// Check if this is an image generation model
+	if slices.Contains(info.Capabilities, model.CapabilityImage) {
+		if opts.Prompt == "" && !interactive {
+			return errors.New("image generation models require a prompt. Usage: ollama run " + name + " \"your prompt here\"")
+		}
+		return imagegen.RunCLI(cmd, name, opts.Prompt, interactive, opts.KeepAlive)
+	}
+
+	// Check for experimental flag
+	isExperimental, _ := cmd.Flags().GetBool("experimental")
+	yoloMode, _ := cmd.Flags().GetBool("experimental-yolo")
+	enableWebsearch, _ := cmd.Flags().GetBool("experimental-websearch")
+
 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
 			var sErr api.AuthorizationError
@@ -543,6 +638,11 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 			}
 		}
 
+		// Use experimental agent loop with tools
+		if isExperimental {
+			return xcmd.GenerateInteractive(cmd, opts.Model, opts.WordWrap, opts.Options, opts.Think, opts.HideThinking, opts.KeepAlive, yoloMode, enableWebsearch)
+		}
+
 		return generateInteractive(cmd, opts)
 	}
 	return generate(cmd, opts)
@@ -646,7 +746,11 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pushing %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				msg := resp.Status
+				if msg == "" {
+					msg = fmt.Sprintf("pushing %s...", resp.Digest[7:19])
+				}
+				bar = progress.NewBar(msg, resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -795,11 +899,11 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 	for _, arg := range args {
 		// Unload the model if it's running before deletion
 		if err := loadOrUnloadModel(cmd, &runOptions{
-			Model:     args[0],
+			Model:     arg,
 			KeepAlive: &api.Duration{Duration: 0},
 		}); err != nil {
 			if !strings.Contains(strings.ToLower(err.Error()), "not found") {
-				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", arg)
 			}
 		}
 
@@ -1711,15 +1815,22 @@ func NewCLI() *cobra.Command {
 	rootCmd.Flags().BoolP("version", "v", false, "Show version information")
 
 	createCmd := &cobra.Command{
-		Use:     "create MODEL",
-		Short:   "Create a model",
-		Args:    cobra.ExactArgs(1),
-		PreRunE: checkServerHeartbeat,
-		RunE:    CreateHandler,
+		Use:   "create MODEL",
+		Short: "Create a model",
+		Args:  cobra.ExactArgs(1),
+		PreRunE: func(cmd *cobra.Command, args []string) error {
+			// Skip server check for experimental mode (writes directly to disk)
+			if experimental, _ := cmd.Flags().GetBool("experimental"); experimental {
+				return nil
+			}
+			return checkServerHeartbeat(cmd, args)
+		},
+		RunE: CreateHandler,
 	}
 
 	createCmd.Flags().StringP("file", "f", "", "Name of the Modelfile (default \"Modelfile\")")
 	createCmd.Flags().StringP("quantize", "q", "", "Quantize model to this level (e.g. q4_K_M)")
+	createCmd.Flags().Bool("experimental", false, "Enable experimental safetensors model creation")
 
 	showCmd := &cobra.Command{
 		Use:     "show MODEL",
@@ -1754,6 +1865,12 @@ func NewCLI() *cobra.Command {
 	runCmd.Flags().Bool("hidethinking", false, "Hide thinking output (if provided)")
 	runCmd.Flags().Bool("truncate", false, "For embedding models: truncate inputs exceeding context length (default: true). Set --truncate=false to error instead")
 	runCmd.Flags().Int("dimensions", 0, "Truncate output embeddings to specified dimension (embedding models only)")
+	runCmd.Flags().Bool("experimental", false, "Enable experimental agent loop with tools")
+	runCmd.Flags().Bool("experimental-yolo", false, "Skip all tool approval prompts (use with caution)")
+	runCmd.Flags().Bool("experimental-websearch", false, "Enable web search tool in experimental mode")
+
+	// Image generation flags (width, height, steps, seed, etc.)
+	imagegen.RegisterFlags(runCmd)
 
 	stopCmd := &cobra.Command{
 		Use:     "stop MODEL",
@@ -1868,6 +1985,7 @@ func NewCLI() *cobra.Command {
 	} {
 		switch cmd {
 		case runCmd:
+			imagegen.AppendFlagsDocs(cmd)
 			appendEnvDocs(cmd, []envconfig.EnvVar{envVars["OLLAMA_HOST"], envVars["OLLAMA_NOHISTORY"]})
 		case serveCmd:
 			appendEnvDocs(cmd, []envconfig.EnvVar{

cmd/cmd_test.go

@@ -1547,6 +1547,79 @@ func TestRunOptions_Copy_ThinkValueVariants(t *testing.T) {
 	}
 }
 
+func TestShowInfoImageGen(t *testing.T) {
+	var b bytes.Buffer
+	err := showInfo(&api.ShowResponse{
+		Details: api.ModelDetails{
+			Family:            "ZImagePipeline",
+			ParameterSize:     "10.3B",
+			QuantizationLevel: "FP8",
+		},
+		Capabilities: []model.Capability{model.CapabilityImage},
+		Requires:     "0.14.0",
+	}, false, &b)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	expect := "  Model\n" +
+		"    architecture    ZImagePipeline    \n" +
+		"    parameters      10.3B             \n" +
+		"    quantization    FP8               \n" +
+		"    requires        0.14.0            \n" +
+		"\n" +
+		"  Capabilities\n" +
+		"    image    \n" +
+		"\n"
+	if diff := cmp.Diff(expect, b.String()); diff != "" {
+		t.Errorf("unexpected output (-want +got):\n%s", diff)
+	}
+}
+
+func TestPushProgressMessage(t *testing.T) {
+	tests := []struct {
+		name    string
+		status  string
+		digest  string
+		wantMsg string
+	}{
+		{
+			name:    "uses status when provided",
+			status:  "uploading model",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "uploading model",
+		},
+		{
+			name:    "falls back to digest when status empty",
+			status:  "",
+			digest:  "sha256:abc123456789def",
+			wantMsg: "pushing abc123456789...",
+		},
+		{
+			name:    "handles short digest gracefully",
+			status:  "",
+			digest:  "sha256:abc",
+			wantMsg: "pushing sha256:abc...",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			msg := tt.status
+			if msg == "" {
+				if len(tt.digest) >= 19 {
+					msg = fmt.Sprintf("pushing %s...", tt.digest[7:19])
+				} else {
+					msg = fmt.Sprintf("pushing %s...", tt.digest)
+				}
+			}
+			if msg != tt.wantMsg {
+				t.Errorf("got %q, want %q", msg, tt.wantMsg)
+			}
+		})
+	}
+}
+
 func TestRunOptions_Copy_Independence(t *testing.T) {
 	// Test that modifications to original don't affect copy
 	originalThink := &api.ThinkValue{Value: "original"}

cmd/interactive.go

@@ -40,6 +40,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		fmt.Fprintln(os.Stderr, "  /bye            Exit")
 		fmt.Fprintln(os.Stderr, "  /?, /help       Help for a command")
 		fmt.Fprintln(os.Stderr, "  /? shortcuts    Help for keyboard shortcuts")
+
 		fmt.Fprintln(os.Stderr, "")
 		fmt.Fprintln(os.Stderr, "Use \"\"\" to begin a multi-line message.")
 
@@ -115,7 +116,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 		Prompt:         ">>> ",
 		AltPrompt:      "... ",
 		Placeholder:    "Send a message (/? for help)",
-		AltPlaceholder: `Use """ to end multi-line input`,
+		AltPlaceholder: "Press Enter to send",
 	})
 	if err != nil {
 		return err

convert/convert.go

@@ -6,11 +6,14 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"iter"
 	"log/slog"
+	"maps"
 	"os"
 	"slices"
 	"strings"
 
+	ofs "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,8 +21,13 @@ type ModelParameters struct {
 	Architectures []string `json:"architectures"`
 	VocabSize     uint32   `json:"vocab_size"`
 
+	// TODO is this needed?
+	ModelType string `json:"model_type"`
+
 	TextModel struct {
-		VocabSize uint32 `json:"vocab_size"`
+		VocabSize  uint32 `json:"vocab_size"`
+		HiddenSize uint32 `json:"hidden_size"`
+		ModelType  string `json:"model_type"`
 	} `json:"text_config"`
 }
 
@@ -33,8 +41,94 @@ type AdapterParameters struct {
 	} `json:"lora_parameters"`
 }
 
-func (ModelParameters) KV(t *Tokenizer) ggml.KV {
-	kv := ggml.KV{
+type KV map[string]any
+
+func (kv KV) Architecture() string {
+	return kv.String("general.architecture", "unknown")
+}
+
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	[]uint8 | []int8 | []uint16 | []int16 |
+		[]uint32 | []int32 | []uint64 | []int64 |
+		[]string | []float32 | []float64 | []bool
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) (T, bool) {
+	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
+		key = kv.Architecture() + "." + key
+	}
+
+	if val, ok := kv[key].(T); ok {
+		return val, true
+	}
+	return defaultValue[0], false
+}
+
+func (kv KV) String(key string, defaultValue ...string) string {
+	val, _ := keyValue(kv, key, append(defaultValue, "")...)
+	return val
+}
+
+func (kv KV) Uint(key string, defaultValue ...uint32) uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Float(key string, defaultValue ...float32) float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, 0)...)
+	return val
+}
+
+func (kv KV) Bool(key string, defaultValue ...bool) bool {
+	val, _ := keyValue(kv, key, append(defaultValue, false)...)
+	return val
+}
+
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	val, _ := keyValue(kv, key, append(defaultValue, []string{""})...)
+	return val
+}
+
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []int32{0})...)
+	return val
+}
+
+func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []uint32{0})...)
+	return val
+}
+
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	val, _ := keyValue(kv, key, append(defaultValue, []float32{0})...)
+	return val
+}
+
+func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
+	val, _ := keyValue(kv, key, append(defaultValue, []bool{false})...)
+	return val
+}
+
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
+func (ModelParameters) KV(t *Tokenizer) KV {
+	kv := KV{
 		"general.file_type":            uint32(1),
 		"general.quantization_version": uint32(2),
 		"tokenizer.ggml.pre":           t.Pre,
@@ -63,7 +157,7 @@ func (ModelParameters) KV(t *Tokenizer) ggml.KV {
 	return kv
 }
 
-func (p AdapterParameters) KV() ggml.KV {
+func (p AdapterParameters) KV() KV {
 	var alpha float32
 	if p.LoraParameters.Alpha == 0 {
 		alpha = float32(p.Alpha)
@@ -71,7 +165,7 @@ func (p AdapterParameters) KV() ggml.KV {
 		alpha = p.LoraParameters.Alpha
 	}
 
-	kv := ggml.KV{
+	kv := KV{
 		"adapter.lora.alpha": alpha,
 		"adapter.type":       "lora",
 		"general.file_type":  uint32(1),
@@ -88,9 +182,14 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-type ModelConverter interface {
+type ModelKV interface {
 	// KV maps parameters to LLM key-values
-	KV(*Tokenizer) ggml.KV
+	KV(*Tokenizer) KV
+}
+
+type ModelConverter interface {
+	ModelKV
+
 	// Tensors maps input tensors to LLM tensors. Model specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -107,7 +206,7 @@ type moreParser interface {
 
 type AdapterConverter interface {
 	// KV maps parameters to LLM key-values
-	KV(ggml.KV) ggml.KV
+	KV(ofs.Config) KV
 	// Tensors maps input tensors to LLM tensors. Adapter specific modifications can be done here.
 	Tensors([]Tensor) []*ggml.Tensor
 	// Replacements returns a list of string pairs to replace in tensor names.
@@ -115,7 +214,7 @@ type AdapterConverter interface {
 	Replacements() []string
 }
 
-func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
+func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ofs.Config) error {
 	bts, err := fs.ReadFile(fsys, "adapter_config.json")
 	if err != nil {
 		return err
@@ -126,8 +225,8 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 		return err
 	}
 
-	arch, ok := baseKV["general.architecture"]
-	if !ok {
+	arch := baseKV.Architecture()
+	if arch == "" {
 		return errors.New("architecture not set for the base model")
 	}
 
@@ -153,23 +252,19 @@ func ConvertAdapter(fsys fs.FS, f *os.File, baseKV ggml.KV) error {
 	return writeFile(f, conv.KV(baseKV), conv.Tensors(ts))
 }
 
-// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
-// and files it finds in the input path.
-// Supported input model formats include safetensors.
-// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
-func ConvertModel(fsys fs.FS, f *os.File) error {
+func LoadModelMetadata(fsys fs.FS) (ModelKV, *Tokenizer, error) {
 	bts, err := fs.ReadFile(fsys, "config.json")
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	var p ModelParameters
 	if err := json.Unmarshal(bts, &p); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if len(p.Architectures) < 1 {
-		return errors.New("unknown architecture")
+		return nil, nil, errors.New("unknown architecture")
 	}
 
 	var conv ModelConverter
@@ -216,23 +311,27 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 		conv = &deepseekocr{}
 	case "DeepseekV3ForCausalLM":
 		conv = &deepseek2Model{}
+	case "Glm4MoeLiteForCausalLM":
+		conv = &glm4MoeLiteModel{}
+	case "Lfm2ForCausalLM":
+		conv = &lfm2Model{}
 	default:
-		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
+		return nil, nil, fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	if t, ok := conv.(moreParser); ok {
 		if err := t.parseMore(fsys); err != nil {
-			return err
+			return nil, nil, err
 		}
 	}
 
 	t, err := parseTokenizer(fsys, conv.specialTokenTypes())
 	if err != nil {
-		return err
+		return nil, nil, err
 	}
 
 	vocabSize := int(cmp.Or(p.VocabSize, p.TextModel.VocabSize))
@@ -254,6 +353,19 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	default:
 		slog.Debug("vocabulary", "size", len(t.Vocabulary.Tokens))
 	}
+	return conv, t, nil
+}
+
+// Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
+// and files it finds in the input path.
+// Supported input model formats include safetensors.
+// Supported input tokenizers files include tokenizer.json (preferred) and tokenizer.model.
+func ConvertModel(fsys fs.FS, f *os.File) error {
+	kv, t, err := LoadModelMetadata(fsys)
+	if err != nil {
+		return err
+	}
+	conv := kv.(ModelConverter)
 
 	ts, err := parseTensors(fsys, strings.NewReplacer(conv.Replacements()...))
 	if err != nil {
@@ -263,7 +375,7 @@ func ConvertModel(fsys fs.FS, f *os.File) error {
 	return writeFile(f, conv.KV(t), conv.Tensors(ts))
 }
 
-func writeFile(f *os.File, kv ggml.KV, ts []*ggml.Tensor) error {
+func writeFile(f *os.File, kv KV, ts []*ggml.Tensor) error {
 	for i := range ts {
 		ts[i].Shape = slices.Clone(ts[i].Shape)
 		slices.Reverse(ts[i].Shape)

convert/convert_bert.go

@@ -88,7 +88,7 @@ func (p *bertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *bertModel) KV(t *Tokenizer) ggml.KV {
+func (p *bertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "bert"
 	kv["bert.attention.causal"] = false

convert/convert_commandr.go

@@ -24,7 +24,7 @@ type commandrModel struct {
 
 var _ ModelConverter = (*commandrModel)(nil)
 
-func (p *commandrModel) KV(t *Tokenizer) ggml.KV {
+func (p *commandrModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "command-r"
 	kv["general.name"] = "command-r"

convert/convert_deepseek2.go

@@ -47,7 +47,7 @@ type deepseek2Model struct {
 	Architecture string
 }
 
-func (p *deepseek2Model) KV(t *Tokenizer) ggml.KV {
+func (p *deepseek2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseek2"
 	kv["general.type"] = "model"

convert/convert_deepseekocr.go

@@ -41,7 +41,7 @@ type deepseekocr struct {
 	} `json:"vision_config"`
 }
 
-func (m *deepseekocr) KV(t *Tokenizer) ggml.KV {
+func (m *deepseekocr) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "deepseekocr"
 	kv["block_count"] = m.LanguageConfig.HiddenLayers

convert/convert_gemma.go

@@ -23,7 +23,7 @@ type gemmaModel struct {
 
 var _ ModelConverter = (*gemmaModel)(nil)
 
-func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
+func (p *gemmaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma"
 	kv["gemma.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2.go

@@ -1,7 +1,5 @@
 package convert
 
-import "github.com/ollama/ollama/fs/ggml"
-
 type gemma2Model struct {
 	gemmaModel
 	SlidingWindow         uint32  `json:"sliding_window"`
@@ -9,7 +7,7 @@ type gemma2Model struct {
 	FinalLogitSoftcap     float32 `json:"final_logit_softcapping"`
 }
 
-func (p *gemma2Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma2Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma2"
 	kv["gemma2.context_length"] = p.MaxPositionEmbeddings

convert/convert_gemma2_adapter.go

@@ -6,6 +6,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -15,7 +16,7 @@ type gemma2Adapter struct {
 
 var _ AdapterConverter = (*gemma2Adapter)(nil)
 
-func (p *gemma2Adapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *gemma2Adapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "gemma2"
 	return kv

convert/convert_gemma3.go

@@ -3,8 +3,6 @@ package convert
 import (
 	"cmp"
 	"slices"
-
-	"github.com/ollama/ollama/fs/ggml"
 )
 
 type gemma3Model struct {
@@ -55,7 +53,7 @@ const (
 	gemma27BLayerCount = 62
 )
 
-func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+func (p *gemma3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3"
 

convert/convert_gemma3n.go

@@ -38,7 +38,7 @@ type gemma3nModel struct {
 	VisionModel struct{} `json:"vision_config"`
 }
 
-func (m *gemma3nModel) KV(t *Tokenizer) ggml.KV {
+func (m *gemma3nModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gemma3n"
 	kv["gemma3n.activation_sparsity_scale"] = slices.Collect(func(yield func(float32) bool) {

convert/convert_glm4moelite.go

@@ -0,0 +1,150 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"log/slog"
+	"regexp"
+	"strconv"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type glm4MoeLiteModel struct {
+	ModelParameters
+	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+	HiddenSize            uint32  `json:"hidden_size"`
+	HiddenLayers          uint32  `json:"num_hidden_layers"`
+	IntermediateSize      uint32  `json:"intermediate_size"`
+	NumAttentionHeads     uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+	RMSNormEPS            float32 `json:"rms_norm_eps"`
+
+	RopeTheta     float32 `json:"rope_theta"`
+	QKNopeHeadDim uint32  `json:"qk_nope_head_dim"`
+	QKRopeHeadDim uint32  `json:"qk_rope_head_dim"`
+	KVLoraRank    uint32  `json:"kv_lora_rank"`
+	QLoraRank     uint32  `json:"q_lora_rank"`
+	VHeadDim      uint32  `json:"v_head_dim"`
+
+	ExpertCount            uint32  `json:"n_routed_experts"`
+	ExpertSharedCount      uint32  `json:"n_shared_experts"`
+	ExpertIntermediateSize uint32  `json:"moe_intermediate_size"`
+	ExpertUsedCount        uint32  `json:"num_experts_per_tok"`
+	ExpertWeightsNorm      bool    `json:"norm_topk_prob"`
+	ExpertWeightsScale     float32 `json:"routed_scaling_factor"`
+
+	LeadingDenseBlockCount uint32 `json:"first_k_dense_replace"`
+}
+
+func (p *glm4MoeLiteModel) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "glm4moelite"
+	kv["general.type"] = "model"
+	kv["glm4moelite.block_count"] = p.HiddenLayers
+
+	numHeads := p.NumAttentionHeads
+	numKVHeads := p.NumKeyValueHeads
+
+	kv["glm4moelite.attention.head_count"] = numHeads
+	kv["glm4moelite.attention.head_count_kv"] = numKVHeads
+	kv["glm4moelite.attention.key_length"] = p.QKNopeHeadDim + p.QKRopeHeadDim
+	kv["glm4moelite.attention.kv_lora_rank"] = p.KVLoraRank
+	kv["glm4moelite.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+	kv["glm4moelite.attention.q_lora_rank"] = p.QLoraRank
+	kv["glm4moelite.attention.value_length"] = p.VHeadDim
+	kv["glm4moelite.context_length"] = p.MaxPositionEmbeddings
+	kv["glm4moelite.embedding_length"] = p.HiddenSize
+	kv["glm4moelite.expert_count"] = p.ExpertCount
+	kv["glm4moelite.expert_feed_forward_length"] = p.ExpertIntermediateSize
+	kv["glm4moelite.expert_shared_count"] = p.ExpertSharedCount
+
+	kv["glm4moelite.expert_gating_func"] = uint32(2)
+	kv["glm4moelite.expert_used_count"] = p.ExpertUsedCount
+	kv["glm4moelite.expert_weights_norm"] = p.ExpertWeightsNorm
+	kv["glm4moelite.expert_weights_scale"] = p.ExpertWeightsScale
+	kv["glm4moelite.feed_forward_length"] = p.IntermediateSize
+	kv["glm4moelite.leading_dense_block_count"] = p.LeadingDenseBlockCount
+
+	kv["glm4moelite.rope.dimension_count"] = p.QKRopeHeadDim
+	kv["glm4moelite.rope.freq_base"] = cmp.Or(p.RopeTheta, float32(1000000.0))
+
+	kv["tokenizer.ggml.pre"] = "glm4"
+
+	return kv
+}
+
+func (p *glm4MoeLiteModel) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"model.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.kv_a_proj_with_mqa", "attn_kv_a_mqa",
+		"self_attn.kv_a_layernorm", "attn_kv_a_norm",
+		"self_attn.kv_b_proj", "attn_kv_b",
+		"self_attn.q_a_proj", "attn_q_a",
+		"self_attn.q_a_layernorm", "attn_q_a_norm",
+		"self_attn.q_b_proj", "attn_q_b",
+		"self_attn.o_proj", "attn_output",
+		"post_attention_layernorm", "ffn_norm",
+		"mlp.shared_experts.down_proj", "ffn_down_shexp",
+		"mlp.shared_experts.gate_proj", "ffn_gate_shexp",
+		"mlp.shared_experts.up_proj", "ffn_up_shexp",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"mlp.gate.e_score_correction_bias", "exp_probs_b.bias",
+		"mlp.gate", "ffn_gate_inp",
+	}
+}
+
+func (p *glm4MoeLiteModel) Tensors(s []Tensor) (out []*ggml.Tensor) {
+	merges := make([]merge, p.HiddenLayers*3)
+	for i := range p.HiddenLayers {
+		merges[i*3+0] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.gate_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_gate_exps.weight", i),
+		}
+		merges[i*3+1] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.up_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_up_exps.weight", i),
+		}
+		merges[i*3+2] = merge{
+			fmt.Sprintf("blk.%d.mlp.experts.*.down_proj.weight", i),
+			fmt.Sprintf("blk.%d.ffn_down_exps.weight", i),
+		}
+	}
+
+	skipLayer := func(n string, minValue uint32) bool {
+		re := regexp.MustCompile(`^blk\.(\d+)`)
+		matches := re.FindStringSubmatch(n)
+		if matches == nil {
+			return false
+		}
+
+		blkNum, err := strconv.Atoi(matches[1])
+		if err != nil {
+			return false
+		}
+
+		return uint32(blkNum) >= minValue
+	}
+
+	out, s = mergeTensors(s, merges...)
+	for _, t := range s {
+		// skip any additional layers (such as the Multi-Token Prediction layer)
+		if skipLayer(t.Name(), p.HiddenLayers) {
+			slog.Debug("skipping layer", "name", t.Name())
+			continue
+		}
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+	return out
+}

convert/convert_gptoss.go

@@ -37,7 +37,7 @@ type gptossModel struct {
 
 var _ ModelConverter = (*gptossModel)(nil)
 
-func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
+func (m *gptossModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gptoss"
 	kv["general.file_type"] = uint32(4)

convert/convert_lfm2.go

@@ -0,0 +1,100 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type lfm2Model struct {
+	ModelParameters
+	HiddenSize            uint32   `json:"hidden_size"`
+	NumHiddenLayers       uint32   `json:"num_hidden_layers"`
+	MaxPositionEmbeddings uint32   `json:"max_position_embeddings"`
+	IntermediateSize      uint32   `json:"intermediate_size"`
+	NumAttentionHeads     uint32   `json:"num_attention_heads"`
+	NumKeyValueHeads      uint32   `json:"num_key_value_heads"`
+	RopeTheta             float32  `json:"rope_theta"`
+	NormEps               float32  `json:"norm_eps"`
+	ConvLCache            uint32   `json:"conv_L_cache"`
+	LayerTypes            []string `json:"layer_types"`
+	TieEmbedding          bool     `json:"tie_embedding"`
+}
+
+var _ ModelConverter = (*lfm2Model)(nil)
+
+func (p *lfm2Model) KV(t *Tokenizer) KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "lfm2"
+	kv["lfm2.vocab_size"] = p.VocabSize
+	kv["lfm2.block_count"] = p.NumHiddenLayers
+	kv["lfm2.embedding_length"] = p.HiddenSize
+	kv["lfm2.feed_forward_length"] = p.IntermediateSize
+	kv["lfm2.context_length"] = p.MaxPositionEmbeddings
+
+	// Build per-layer KV head count array based on layer_types
+	// (0 = shortconv layer, non-zero = attention layer with that many KV heads)
+	kvHeadCounts := make([]uint32, p.NumHiddenLayers)
+	for i := range p.NumHiddenLayers {
+		if int(i) < len(p.LayerTypes) && p.LayerTypes[i] == "full_attention" {
+			kvHeadCounts[i] = p.NumKeyValueHeads
+		}
+	}
+
+	kv["lfm2.attention.head_count"] = p.NumAttentionHeads
+	kv["lfm2.attention.head_count_kv"] = kvHeadCounts
+	kv["lfm2.attention.key_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.value_length"] = p.HiddenSize / p.NumAttentionHeads
+	kv["lfm2.attention.layer_norm_rms_epsilon"] = p.NormEps
+	kv["lfm2.rope.freq_base"] = p.RopeTheta
+	kv["lfm2.shortconv.l_cache"] = p.ConvLCache
+
+	return kv
+}
+
+func (p *lfm2Model) Tensors(ts []Tensor) []*ggml.Tensor {
+	var out []*ggml.Tensor
+
+	for _, t := range ts {
+		shape := t.Shape()
+
+		// Squeeze conv weights: [D, 1, K] -> [D, K]
+		if strings.HasSuffix(t.Name(), "shortconv.conv.weight") {
+			if len(shape) == 3 && shape[1] == 1 {
+				shape = []uint64{shape[0], shape[2]}
+			}
+		}
+
+		out = append(out, &ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    slices.Clone(shape),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *lfm2Model) Replacements() []string {
+	return []string{
+		"model.embed_tokens", "token_embd",
+		"model.embedding_norm", "output_norm",
+		"model.layers", "blk",
+		"operator_norm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.out_proj", "attn_output",
+		"self_attn.q_layernorm", "attn_q_norm",
+		"self_attn.k_layernorm", "attn_k_norm",
+		"conv.conv", "shortconv.conv",
+		"conv.in_proj", "shortconv.in_proj",
+		"conv.out_proj", "shortconv.out_proj",
+		"feed_forward.w1", "ffn_gate",
+		"feed_forward.w2", "ffn_down",
+		"feed_forward.w3", "ffn_up",
+		"ffn_norm", "ffn_norm",
+	}
+}

convert/convert_llama.go

@@ -48,7 +48,7 @@ type llamaModel struct {
 
 var _ ModelConverter = (*llamaModel)(nil)
 
-func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
+func (p *llamaModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama"
 	kv["llama.vocab_size"] = p.VocabSize

convert/convert_llama4.go

@@ -35,7 +35,7 @@ type llama4Model struct {
 }
 
 // KV implements ModelConverter.
-func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
+func (p *llama4Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "llama4"
 

convert/convert_llama_adapter.go

@@ -7,6 +7,7 @@ import (
 	"github.com/pdevine/tensor"
 	"github.com/pdevine/tensor/native"
 
+	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -18,13 +19,13 @@ type llamaAdapter struct {
 
 var _ AdapterConverter = (*llamaAdapter)(nil)
 
-func (p *llamaAdapter) KV(baseKV ggml.KV) ggml.KV {
+func (p *llamaAdapter) KV(baseKV fs.Config) KV {
 	kv := p.AdapterParameters.KV()
 	kv["general.architecture"] = "llama"
-	kv["llama.attention.head_count"] = baseKV["llama.attention.head_count"]
-	kv["llama.attention.head_count_kv"] = baseKV["llama.attention.head_count_kv"]
+	kv["llama.attention.head_count"] = baseKV.Value("llama.attention.head_count")
+	kv["llama.attention.head_count_kv"] = baseKV.Value("llama.attention.head_count_kv")
 
-	p.NumAttentionHeads = baseKV["llama.attention.head_count"].(uint32)
+	p.NumAttentionHeads = baseKV.Value("llama.attention.head_count").(uint32)
 
 	return kv
 }

convert/convert_mistral.go

@@ -60,7 +60,7 @@ type mistral3Model struct {
 	ProjectorHiddenAct      string `json:"projector_hidden_act"`
 }
 
-func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.TextModel.VocabSize

convert/convert_mistral_causal.go

@@ -39,7 +39,7 @@ type mistral3CausalModel struct {
 	} `json:"rope_parameters"`
 }
 
-func (p *mistral3CausalModel) KV(t *Tokenizer) ggml.KV {
+func (p *mistral3CausalModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "mistral3"
 	kv["mistral3.vocab_size"] = p.VocabSize

convert/convert_mixtral.go

@@ -12,7 +12,7 @@ type mixtralModel struct {
 	NumExpertsPerToken uint32 `json:"num_experts_per_tok"`
 }
 
-func (p *mixtralModel) KV(t *Tokenizer) ggml.KV {
+func (p *mixtralModel) KV(t *Tokenizer) KV {
 	kv := p.llamaModel.KV(t)
 
 	if p.NumLocalExperts > 0 {

convert/convert_mllama.go

@@ -34,7 +34,7 @@ type mllamaModel struct {
 	} `json:"vision_config"`
 }
 
-func (m *mllamaModel) KV(t *Tokenizer) ggml.KV {
+func (m *mllamaModel) KV(t *Tokenizer) KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "mllama"
 

convert/convert_nomicbert.go

@@ -87,7 +87,7 @@ func (p *nomicbertModel) parseMore(fsys fs.FS) error {
 	return nil
 }
 
-func (p *nomicbertModel) KV(t *Tokenizer) ggml.KV {
+func (p *nomicbertModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 
 	// Determine architecture based on MoE parameters (following qwen3 pattern)

convert/convert_olmo.go

@@ -34,7 +34,7 @@ type olmoModel struct {
 
 var _ ModelConverter = (*olmoModel)(nil)
 
-func (p *olmoModel) KV(t *Tokenizer) ggml.KV {
+func (p *olmoModel) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "olmo3"
 	kv["olmo3.block_count"] = p.NumHiddenLayers

convert/convert_phi3.go

@@ -37,7 +37,7 @@ type phi3Model struct {
 
 var _ ModelConverter = (*phi3Model)(nil)
 
-func (p *phi3Model) KV(t *Tokenizer) ggml.KV {
+func (p *phi3Model) KV(t *Tokenizer) KV {
 	kv := p.ModelParameters.KV(t)
 	kv["general.architecture"] = "phi3"
 	kv["phi3.context_length"] = p.MaxPositionEmbeddings

convert/convert_qwen2.go

@@ -22,7 +22,7 @@ type qwen2Model struct {
 
 var _ ModelConverter = (*qwen2Model)(nil)
 
-func (q *qwen2Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen2Model) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen2"
 	kv["qwen2.block_count"] = q.HiddenLayers

convert/convert_qwen25vl.go

@@ -29,7 +29,7 @@ type qwen25VLModel struct {
 
 var _ ModelConverter = (*qwen25VLModel)(nil)
 
-func (q *qwen25VLModel) KV(t *Tokenizer) ggml.KV {
+func (q *qwen25VLModel) KV(t *Tokenizer) KV {
 	kv := q.ModelParameters.KV(t)
 	kv["general.architecture"] = "qwen25vl"
 

convert/convert_qwen3.go

@@ -32,7 +32,7 @@ type qwen3Model struct {
 }
 
 // KV implements ModelConverter.
-func (q *qwen3Model) KV(t *Tokenizer) ggml.KV {
+func (q *qwen3Model) KV(t *Tokenizer) KV {
 	arch := "qwen3"
 	if q.NumExperts > 0 {
 		arch += "moe"

convert/convert_qwen3vl.go

@@ -45,7 +45,7 @@ func (m *qwen3VLModel) parseMore(fsys fs.FS) error {
 	return json.Unmarshal(bts, &m.VisionModel)
 }
 
-func (m *qwen3VLModel) KV(t *Tokenizer) ggml.KV {
+func (m *qwen3VLModel) KV(t *Tokenizer) KV {
 	kv := m.qwen3Model.KV(t)
 
 	arch := "qwen3vl"

convert/convert_test.go

@@ -19,6 +19,7 @@ import (
 	"testing"
 
 	"github.com/google/go-cmp/cmp"
+	fsc "github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/fs/ggml"
 )
 
@@ -28,7 +29,7 @@ type tensorData struct {
 	Shape   []int  `json:"shape"`
 }
 
-func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
+func convertFull(t *testing.T, fsys fs.FS) (*os.File, fsc.Config, ggml.Tensors) {
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), "f16")
@@ -59,9 +60,10 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	return r, m.KV(), m.Tensors()
 }
 
-func generateResultsJSON(t *testing.T, f *os.File, kv ggml.KV, tensors ggml.Tensors) map[string]string {
+func generateResultsJSON(t *testing.T, f *os.File, kv fsc.Config, tensors ggml.Tensors) map[string]string {
 	actual := make(map[string]string)
-	for k, v := range kv {
+	for k := range kv.Keys() {
+		v := kv.Value(k)
 		if s, ok := v.(json.Marshaler); !ok {
 			actual[k] = fmt.Sprintf("%v", v)
 		} else {
@@ -277,7 +279,7 @@ func generateSafetensorTestData(t *testing.T, tempDir string, tensorData map[str
 func TestConvertAdapter(t *testing.T) {
 	type AdapterCase struct {
 		Name     string
-		BaseKV   map[string]any
+		BaseKV   KV
 		Expected map[string]string
 	}
 

convert/reader.go

@@ -40,6 +40,7 @@ const (
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
 		strings.HasSuffix(t.name, ".bias") ||
+		strings.HasSuffix(t.name, ".shortconv.conv.weight") ||
 		t.name == "token_types.weight" ||
 		t.name == "v.positional_embedding_vlm" ||
 		t.name == "v.tile_position_embd.weight" ||

docs/README.md

@@ -14,6 +14,7 @@
 * [API Reference](https://docs.ollama.com/api)
 * [Modelfile Reference](https://docs.ollama.com/modelfile)
 * [OpenAI Compatibility](https://docs.ollama.com/api/openai-compatibility)
+* [Anthropic Compatibility](./api/anthropic-compatibility.mdx)
 
 ### Resources
 

docs/api.md

@@ -16,6 +16,7 @@
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
 - [Version](#version)
+- [Experimental: Image Generation](#image-generation-experimental)
 
 ## Conventions
 
@@ -58,6 +59,15 @@ Advanced parameters (optional):
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 - `context` (deprecated): the context parameter returned from a previous request to `/generate`, this can be used to keep a short conversational memory
 
+Experimental image generation parameters (for image generation models only):
+
+> [!WARNING]
+> These parameters are experimental and may change in future versions.
+
+- `width`: width of the generated image in pixels
+- `height`: height of the generated image in pixels
+- `steps`: number of diffusion steps
+
 #### Structured outputs
 
 Structured outputs are supported by providing a JSON schema in the `format` parameter. The model will generate a response that matches the schema. See the [structured outputs](#request-structured-outputs) example below.
@@ -895,11 +905,11 @@ curl http://localhost:11434/api/chat -d '{
       "tool_calls": [
         {
           "function": {
-            "name": "get_temperature",
+            "name": "get_weather",
             "arguments": {
               "city": "Toronto"
             }
-          },
+          }
         }
       ]
     },
@@ -907,7 +917,7 @@ curl http://localhost:11434/api/chat -d '{
     {
       "role": "tool",
       "content": "11 degrees celsius",
-      "tool_name": "get_temperature",
+      "tool_name": "get_weather"
     }
   ],
   "stream": false,
@@ -1867,3 +1877,55 @@ curl http://localhost:11434/api/version
   "version": "0.5.1"
 }
 ```
+
+## Experimental Features
+
+### Image Generation (Experimental)
+
+> [!WARNING]
+> Image generation is experimental and may change in future versions.
+
+Image generation is now supported through the standard `/api/generate` endpoint when using image generation models. The API automatically detects when an image generation model is being used.
+
+See the [Generate a completion](#generate-a-completion) section for the full API documentation. The experimental image generation parameters (`width`, `height`, `steps`) are documented there.
+
+#### Example
+
+##### Request
+
+```shell
+curl http://localhost:11434/api/generate -d '{
+  "model": "x/z-image-turbo",
+  "prompt": "a sunset over mountains",
+  "width": 1024,
+  "height": 768
+}'
+```
+
+##### Response (streaming)
+
+Progress updates during generation:
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:00.000000Z",
+  "completed": 5,
+  "total": 20,
+  "done": false
+}
+```
+
+##### Final Response
+
+```json
+{
+  "model": "x/z-image-turbo",
+  "created_at": "2024-01-15T10:30:15.000000Z",
+  "image": "iVBORw0KGgoAAAANSUhEUg...",
+  "done": true,
+  "done_reason": "stop",
+  "total_duration": 15000000000,
+  "load_duration": 2000000000
+}
+```

docs/api/anthropic-compatibility.mdx

@@ -0,0 +1,408 @@
+---
+title: Anthropic compatibility
+---
+
+Ollama provides compatibility with the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) to help connect existing applications to Ollama, including tools like Claude Code.
+
+## Recommended models
+
+For coding use cases, models like `glm-4.7:cloud`, `minimax-m2.1:cloud`, and `qwen3-coder` are recommended.
+
+Pull a model before use:
+```shell
+ollama pull qwen3-coder
+ollama pull glm-4.7:cloud
+```
+
+## Usage
+
+### Environment variables
+
+To use Ollama with tools that expect the Anthropic API (like Claude Code), set these environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama  # required but ignored
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama  # required but ignored
+```
+
+### Simple `/v1/messages` example
+
+<CodeGroup dropdown>
+
+```python basic.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',  # required but ignored
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[
+        {'role': 'user', 'content': 'Hello, how are you?'}
+    ]
+)
+print(message.content[0].text)
+```
+
+```javascript basic.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama", // required but ignored
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Hello, how are you?" }],
+});
+
+console.log(message.content[0].text);
+```
+
+```shell basic.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-H "x-api-key: ollama" \
+-H "anthropic-version: 2023-06-01" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "messages": [{ "role": "user", "content": "Hello, how are you?" }]
+}'
+```
+
+</CodeGroup>
+
+### Streaming example
+
+<CodeGroup dropdown>
+
+```python streaming.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+with client.messages.stream(
+    model='qwen3-coder',
+    max_tokens=1024,
+    messages=[{'role': 'user', 'content': 'Count from 1 to 10'}]
+) as stream:
+    for text in stream.text_stream:
+        print(text, end='', flush=True)
+```
+
+```javascript streaming.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const stream = await anthropic.messages.stream({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  messages: [{ role: "user", content: "Count from 1 to 10" }],
+});
+
+for await (const event of stream) {
+  if (
+    event.type === "content_block_delta" &&
+    event.delta.type === "text_delta"
+  ) {
+    process.stdout.write(event.delta.text);
+  }
+}
+```
+
+```shell streaming.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "stream": true,
+  "messages": [{ "role": "user", "content": "Count from 1 to 10" }]
+}'
+```
+
+</CodeGroup>
+
+### Tool calling example
+
+<CodeGroup dropdown>
+
+```python tools.py
+import anthropic
+
+client = anthropic.Anthropic(
+    base_url='http://localhost:11434',
+    api_key='ollama',
+)
+
+message = client.messages.create(
+    model='qwen3-coder',
+    max_tokens=1024,
+    tools=[
+        {
+            'name': 'get_weather',
+            'description': 'Get the current weather in a location',
+            'input_schema': {
+                'type': 'object',
+                'properties': {
+                    'location': {
+                        'type': 'string',
+                        'description': 'The city and state, e.g. San Francisco, CA'
+                    }
+                },
+                'required': ['location']
+            }
+        }
+    ],
+    messages=[{'role': 'user', 'content': "What's the weather in San Francisco?"}]
+)
+
+for block in message.content:
+    if block.type == 'tool_use':
+        print(f'Tool: {block.name}')
+        print(f'Input: {block.input}')
+```
+
+```javascript tools.js
+import Anthropic from "@anthropic-ai/sdk";
+
+const anthropic = new Anthropic({
+  baseURL: "http://localhost:11434",
+  apiKey: "ollama",
+});
+
+const message = await anthropic.messages.create({
+  model: "qwen3-coder",
+  max_tokens: 1024,
+  tools: [
+    {
+      name: "get_weather",
+      description: "Get the current weather in a location",
+      input_schema: {
+        type: "object",
+        properties: {
+          location: {
+            type: "string",
+            description: "The city and state, e.g. San Francisco, CA",
+          },
+        },
+        required: ["location"],
+      },
+    },
+  ],
+  messages: [{ role: "user", content: "What's the weather in San Francisco?" }],
+});
+
+for (const block of message.content) {
+  if (block.type === "tool_use") {
+    console.log("Tool:", block.name);
+    console.log("Input:", block.input);
+  }
+}
+```
+
+```shell tools.sh
+curl -X POST http://localhost:11434/v1/messages \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "qwen3-coder",
+  "max_tokens": 1024,
+  "tools": [
+    {
+      "name": "get_weather",
+      "description": "Get the current weather in a location",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "location": {
+            "type": "string",
+            "description": "The city and state"
+          }
+        },
+        "required": ["location"]
+      }
+    }
+  ],
+  "messages": [{ "role": "user", "content": "What is the weather in San Francisco?" }]
+}'
+```
+
+</CodeGroup>
+
+## Using with Claude Code
+
+[Claude Code](https://code.claude.com/docs/en/overview) can be configured to use Ollama as its backend:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 ANTHROPIC_API_KEY=ollama claude --model qwen3-coder
+```
+
+Or set the environment variables in your shell profile:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+export ANTHROPIC_API_KEY=ollama
+```
+
+Then run Claude Code with any Ollama model:
+
+```shell
+# Local models
+claude --model qwen3-coder
+claude --model gpt-oss:20b
+
+# Cloud models
+claude --model glm-4.7:cloud
+claude --model minimax-m2.1:cloud
+```
+
+## Endpoints
+
+### `/v1/messages`
+
+#### Supported features
+
+- [x] Messages
+- [x] Streaming
+- [x] System prompts
+- [x] Multi-turn conversations
+- [x] Vision (images)
+- [x] Tools (function calling)
+- [x] Tool results
+- [x] Thinking/extended thinking
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `max_tokens`
+- [x] `messages`
+  - [x] Text `content`
+  - [x] Image `content` (base64)
+  - [x] Array of content blocks
+  - [x] `tool_use` blocks
+  - [x] `tool_result` blocks
+  - [x] `thinking` blocks
+- [x] `system` (string or array)
+- [x] `stream`
+- [x] `temperature`
+- [x] `top_p`
+- [x] `top_k`
+- [x] `stop_sequences`
+- [x] `tools`
+- [x] `thinking`
+- [ ] `tool_choice`
+- [ ] `metadata`
+
+#### Supported response fields
+
+- [x] `id`
+- [x] `type`
+- [x] `role`
+- [x] `model`
+- [x] `content` (text, tool_use, thinking blocks)
+- [x] `stop_reason` (end_turn, max_tokens, tool_use)
+- [x] `usage` (input_tokens, output_tokens)
+
+#### Streaming events
+
+- [x] `message_start`
+- [x] `content_block_start`
+- [x] `content_block_delta` (text_delta, input_json_delta, thinking_delta)
+- [x] `content_block_stop`
+- [x] `message_delta`
+- [x] `message_stop`
+- [x] `ping`
+- [x] `error`
+
+## Models
+
+Ollama supports both local and cloud models.
+
+### Local models
+
+Pull a local model before use:
+
+```shell
+ollama pull qwen3-coder
+```
+
+Recommended local models:
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+
+### Cloud models
+
+Cloud models are available immediately without pulling:
+
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+
+### Default model names
+
+For tooling that relies on default Anthropic model names such as `claude-3-5-sonnet`, use `ollama cp` to copy an existing model name:
+
+```shell
+ollama cp qwen3-coder claude-3-5-sonnet
+```
+
+Afterwards, this new model name can be specified in the `model` field:
+
+```shell
+curl http://localhost:11434/v1/messages \
+    -H "Content-Type: application/json" \
+    -d '{
+        "model": "claude-3-5-sonnet",
+        "max_tokens": 1024,
+        "messages": [
+            {
+                "role": "user",
+                "content": "Hello!"
+            }
+        ]
+    }'
+```
+
+## Differences from the Anthropic API
+
+### Behavior differences
+
+- API key is accepted but not validated
+- `anthropic-version` header is accepted but not used
+- Token counts are approximations based on the underlying model's tokenizer
+
+### Not supported
+
+The following Anthropic API features are not currently supported:
+
+| Feature | Description |
+|---------|-------------|
+| `/v1/messages/count_tokens` | Token counting endpoint |
+| `tool_choice` | Forcing specific tool use or disabling tools |
+| `metadata` | Request metadata (user_id) |
+| Prompt caching | `cache_control` blocks for caching prefixes |
+| Batches API | `/v1/messages/batches` for async batch processing |
+| Citations | `citations` content blocks |
+| PDF support | `document` content blocks with PDF files |
+| Server-sent errors | `error` events during streaming (errors return HTTP status) |
+
+### Partial support
+
+| Feature | Status |
+|---------|--------|
+| Image content | Base64 images supported; URL images not supported |
+| Extended thinking | Basic support; `budget_tokens` accepted but not enforced |

docs/api/openai-compatibility.mdx

@@ -275,8 +275,77 @@ curl -X POST http://localhost:11434/v1/chat/completions \
 - [x] `dimensions`
 - [ ] `user`
 
+### `/v1/images/generations` (experimental)
+
+> Note: This endpoint is experimental and may change or be removed in future versions.
+
+Generate images using image generation models.
+
+<CodeGroup dropdown>
+
+```python images.py
+from openai import OpenAI
+
+client = OpenAI(
+    base_url='http://localhost:11434/v1/',
+    api_key='ollama',  # required but ignored
+)
+
+response = client.images.generate(
+    model='x/z-image-turbo',
+    prompt='A cute robot learning to paint',
+    size='1024x1024',
+    response_format='b64_json',
+)
+print(response.data[0].b64_json[:50] + '...')
+```
+
+```javascript images.js
+import OpenAI from "openai";
+
+const openai = new OpenAI({
+  baseURL: "http://localhost:11434/v1/",
+  apiKey: "ollama", // required but ignored
+});
+
+const response = await openai.images.generate({
+  model: "x/z-image-turbo",
+  prompt: "A cute robot learning to paint",
+  size: "1024x1024",
+  response_format: "b64_json",
+});
+
+console.log(response.data[0].b64_json.slice(0, 50) + "...");
+```
+
+```shell images.sh
+curl -X POST http://localhost:11434/v1/images/generations \
+-H "Content-Type: application/json" \
+-d '{
+  "model": "x/z-image-turbo",
+  "prompt": "A cute robot learning to paint",
+  "size": "1024x1024",
+  "response_format": "b64_json"
+}'
+```
+
+</CodeGroup>
+
+#### Supported request fields
+
+- [x] `model`
+- [x] `prompt`
+- [x] `size` (e.g. "1024x1024")
+- [x] `response_format` (only `b64_json` supported)
+- [ ] `n`
+- [ ] `quality`
+- [ ] `style`
+- [ ] `user`
+
 ### `/v1/responses`
 
+> Note: Added in Ollama v0.13.3
+
 Ollama supports the [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses). Only the non-stateful flavor is supported (i.e., there is no `previous_response_id` or `conversation` support).
 
 #### Supported features

docs/capabilities/vision.mdx

@@ -36,7 +36,6 @@ Provide an `images` array. SDKs accept file paths, URLs or raw bytes while the R
         }],
         "stream": false
     }'
-    "
     ```
   </Tab>
   <Tab title="Python">

docs/capabilities/web-search.mdx

@@ -110,7 +110,7 @@ More Ollama [Python example](https://github.com/ollama/ollama-python/blob/main/e
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const results = await client.webSearch({ query: "what is ollama?" });
+const results = await client.webSearch("what is ollama?");
 console.log(JSON.stringify(results, null, 2));
 ```
 
@@ -213,7 +213,7 @@ models](https://ollama.com/models)\n\nAvailable for macOS, Windows, and Linux',
 import { Ollama } from "ollama";
 
 const client = new Ollama();
-const fetchResult = await client.webFetch({ url: "https://ollama.com" });
+const fetchResult = await client.webFetch("https://ollama.com");
 console.log(JSON.stringify(fetchResult, null, 2));
 ```
 

docs/docs.json

@@ -32,7 +32,9 @@
     "codeblocks": "system"
   },
   "contextual": {
-    "options": ["copy"]
+    "options": [
+      "copy"
+    ]
   },
   "navbar": {
     "links": [
@@ -52,7 +54,9 @@
       "display": "simple"
     },
     "examples": {
-      "languages": ["curl"]
+      "languages": [
+        "curl"
+      ]
     }
   },
   "redirects": [
@@ -97,6 +101,7 @@
           {
             "group": "Integrations",
             "pages": [
+              "/integrations/claude-code",
               "/integrations/vscode",
               "/integrations/jetbrains",
               "/integrations/codex",
@@ -106,7 +111,9 @@
               "/integrations/zed",
               "/integrations/roo-code",
               "/integrations/n8n",
-              "/integrations/xcode"
+              "/integrations/xcode",
+              "/integrations/onyx",
+              "/integrations/marimo"
             ]
           },
           {
@@ -139,7 +146,8 @@
               "/api/streaming",
               "/api/usage",
               "/api/errors",
-              "/api/openai-compatibility"
+              "/api/openai-compatibility",
+              "/api/anthropic-compatibility"
             ]
           },
           {

docs/faq.mdx

@@ -14,15 +14,15 @@ curl -fsSL https://ollama.com/install.sh | sh
 
 ## How can I view the logs?
 
-Review the [Troubleshooting](./troubleshooting.md) docs for more about using logs.
+Review the [Troubleshooting](./troubleshooting) docs for more about using logs.
 
 ## Is my GPU compatible with Ollama?
 
-Please refer to the [GPU docs](./gpu.md).
+Please refer to the [GPU docs](./gpu).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 2048 tokens.
+By default, Ollama uses a context window size of 4096 tokens.
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use:
 

docs/gpu.mdx

@@ -33,7 +33,7 @@ Check your compute compatibility to see if your card is supported:
 | 5.0                | GeForce GTX         | `GTX 750 Ti` `GTX 750` `NVS 810`                                                                                               |
 |                    | Quadro              | `K2200` `K1200` `K620` `M1200` `M520` `M5000M` `M4000M` `M3000M` `M2000M` `M1000M` `K620M` `M600M` `M500M`                     |
 
-For building locally to support older GPUs, see [developer.md](./development.md#linux-cuda-nvidia)
+For building locally to support older GPUs, see [developer](./development#linux-cuda-nvidia)
 
 ### GPU Selection
 
@@ -54,7 +54,7 @@ sudo modprobe nvidia_uvm`
 
 Ollama supports the following AMD GPUs via the ROCm library:
 
-> [!NOTE]
+> **NOTE:**
 > Additional AMD GPU support is provided by the Vulkan Library - see below.
 
 
@@ -132,9 +132,9 @@ Ollama supports GPU acceleration on Apple devices via the Metal API.
 
 ## Vulkan GPU Support
 
-> [!NOTE]
+> **NOTE:**
 > Vulkan is currently an Experimental feature.  To enable, you must set OLLAMA_VULKAN=1 for the Ollama server as
-described in the [FAQ](faq.md#how-do-i-configure-ollama-server)
+described in the [FAQ](faq#how-do-i-configure-ollama-server)
 
 Additional GPU support on Windows and Linux is provided via
 [Vulkan](https://www.vulkan.org/). On Windows most GPU vendors drivers come
@@ -161,6 +161,6 @@ sudo setcap cap_perfmon+ep /usr/local/bin/ollama
 
 To select specific Vulkan GPU(s), you can set the environment variable
 `GGML_VK_VISIBLE_DEVICES` to one or more numeric IDs on the Ollama server as
-described in the [FAQ](faq.md#how-do-i-configure-ollama-server). If you
+described in the [FAQ](faq#how-do-i-configure-ollama-server). If you
 encounter any problems with Vulkan based GPUs, you can disable all Vulkan GPUs
 by setting `GGML_VK_VISIBLE_DEVICES=-1` 

docs/integrations/claude-code.mdx

@@ -0,0 +1,78 @@
+---
+title: Claude Code
+---
+
+Claude Code is Anthropic's agentic coding tool that can read, modify, and execute code in your working directory. 
+
+Open models can be used with Claude Code through Ollama's Anthropic-compatible API, enabling you to use models such as `qwen3-coder`, `gpt-oss:20b`, or other models.
+
+![Claude Code with Ollama](https://files.ollama.com/claude-code.png)
+
+## Install
+
+Install [Claude Code](https://code.claude.com/docs/en/overview):
+
+<CodeGroup>
+
+```shell macOS / Linux
+curl -fsSL https://claude.ai/install.sh | bash
+```
+
+```powershell Windows
+irm https://claude.ai/install.ps1 | iex
+```
+
+</CodeGroup>
+
+## Usage with Ollama
+
+Claude Code connects to Ollama using the Anthropic-compatible API.
+
+1. Set the environment variables:
+
+```shell
+export ANTHROPIC_AUTH_TOKEN=ollama
+export ANTHROPIC_BASE_URL=http://localhost:11434
+```
+
+2. Run Claude Code with an Ollama model:
+
+```shell
+claude --model gpt-oss:20b
+```
+
+Or run with environment variables inline:
+
+```shell
+ANTHROPIC_AUTH_TOKEN=ollama ANTHROPIC_BASE_URL=http://localhost:11434 claude --model gpt-oss:20b
+```
+
+**Note:** Claude Code requires a large context window. We recommend at least 32K tokens. See the [context length documentation](/context-length) for how to adjust context length in Ollama.
+
+## Connecting to ollama.com
+
+1. Create an [API key](https://ollama.com/settings/keys) on ollama.com
+2. Set the environment variables:
+
+```shell
+export ANTHROPIC_BASE_URL=https://ollama.com
+export ANTHROPIC_API_KEY=<your-api-key>
+```
+
+3. Run Claude Code with a cloud model:
+
+```shell
+claude --model glm-4.7:cloud
+```
+
+## Recommended Models
+
+### Cloud models
+- `glm-4.7:cloud` - High-performance cloud model
+- `minimax-m2.1:cloud` - Fast cloud model
+- `qwen3-coder:480b` - Large coding model
+
+### Local models
+- `qwen3-coder` - Excellent for coding tasks
+- `gpt-oss:20b` - Strong general-purpose model
+- `gpt-oss:120b` - Larger general-purpose model for more complex tasks

docs/integrations/marimo.mdx

@@ -0,0 +1,73 @@
+---
+title: marimo
+---
+
+## Install
+
+Install [marimo](https://marimo.io). You can use `pip` or `uv` for this. You 
+can also use `uv` to create a sandboxed environment for marimo by running:
+
+```
+uvx marimo edit --sandbox notebook.py
+```
+
+## Usage with Ollama
+
+1. In marimo, go to the user settings and go to the AI tab. From here
+you can find and configure Ollama as an AI provider. For local use you
+would typically point the base url to `http://localhost:11434/v1`.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-settings.png" 
+    alt="Ollama settings in marimo"
+    width="50%"
+  />
+</div>
+
+2. Once the AI provider is set up, you can turn on/off specific AI models you'd like to access. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-models.png" 
+    alt="Selecting an Ollama model"
+    width="50%"
+  />
+</div>
+
+3. You can also add a model to the list of available models by scrolling to the bottom and using the UI there. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-add-model.png" 
+    alt="Adding a new Ollama model"
+    width="50%"
+  />
+</div>
+
+4. Once configured, you can now use Ollama for AI chats in marimo.
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-chat.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+4. Alternatively, you can now use Ollama for **inline code completion** in marimo. This can be configured in the "AI Features" tab. 
+
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/marimo-code-completion.png" 
+    alt="Configure code completion"
+    width="50%"
+  />
+</div>
+
+
+## Connecting to ollama.com
+
+1. Sign in to ollama cloud via `ollama signin` 
+2. In the ollama model settings add a model that ollama hosts, like `gpt-oss:120b`.
+3. You can now refer to this model in marimo!

docs/integrations/onyx.mdx

@@ -0,0 +1,63 @@
+---
+title: Onyx
+---
+
+## Overview
+[Onyx](http://onyx.app/) is a self-hostable Chat UI that integrates with all Ollama models. Features include:
+- Creating custom Agents
+- Web search
+- Deep Research
+- RAG over uploaded documents and connected apps
+- Connectors to applications like Google Drive, Email, Slack, etc.
+- MCP and OpenAPI Actions support
+- Image generation
+- User/Groups management, RBAC, SSO, etc.
+
+Onyx can be deployed for single users or large organizations.
+
+## Install Onyx
+
+Deploy Onyx with the [quickstart guide](https://docs.onyx.app/deployment/getting_started/quickstart).
+
+<Info>
+Resourcing/scaling docs [here](https://docs.onyx.app/deployment/getting_started/resourcing).
+</Info>
+
+## Usage with Ollama 
+
+1. Login to your Onyx deployment (create an account first).
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-login.png" 
+    alt="Onyx Login Page"
+    width="75%"
+  />
+</div>
+2. In the set-up process select `Ollama` as the LLM provider.
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-llm.png" 
+    alt="Onyx Set Up Form"
+    width="75%"
+  />
+</div>
+3. Provide your **Ollama API URL** and select your models.
+<Note>If you're running Onyx in Docker, to access your computer's local network use `http://host.docker.internal` instead of `http://127.0.0.1`.</Note>
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-ollama-form.png" 
+    alt="Selecting Ollama Models"
+    width="75%"
+  />
+</div>
+
+You can also easily connect up Onyx Cloud with the `Ollama Cloud` tab of the setup.
+
+## Send your first query
+<div style={{ display: 'flex', justifyContent: 'center' }}>
+  <img 
+    src="/images/onyx-query.png" 
+    alt="Onyx Query Example"
+    width="75%"
+  />
+</div>

docs/linux.mdx

@@ -20,8 +20,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Download and extract the package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 Start Ollama:
@@ -41,8 +41,8 @@ ollama -v
 If you have an AMD GPU, also download and extract the additional ROCm package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### ARM64 install
@@ -50,8 +50,8 @@ curl -fsSL https://ollama.com/download/ollama-linux-amd64-rocm.tgz \
 Download and extract the ARM64-specific package:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-arm64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-arm64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ### Adding Ollama as a startup service (recommended)
@@ -146,8 +146,8 @@ curl -fsSL https://ollama.com/install.sh | sh
 Or by re-downloading Ollama:
 
 ```shell
-curl -fsSL https://ollama.com/download/ollama-linux-amd64.tgz \
-    | sudo tar zx -C /usr
+curl -fsSL https://ollama.com/download/ollama-linux-amd64.tar.zst \
+    | sudo tar x -C /usr
 ```
 
 ## Installing specific versions

docs/troubleshooting.md

@@ -1,3 +0,0 @@
-# Troubleshooting
-
-For troubleshooting, see [https://docs.ollama.com/troubleshooting](https://docs.ollama.com/troubleshooting)

docs/troubleshooting.mdx

@@ -87,7 +87,7 @@ When Ollama starts up, it takes inventory of the GPUs present in the system to d
 
 ### Linux NVIDIA Troubleshooting
 
-If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker](./docker)
 
 Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
 

fs/config.go

@@ -1,5 +1,7 @@
 package fs
 
+import "iter"
+
 type Config interface {
 	Architecture() string
 	String(string, ...string) string
@@ -11,4 +13,8 @@ type Config interface {
 	Ints(string, ...[]int32) []int32
 	Floats(string, ...[]float32) []float32
 	Bools(string, ...[]bool) []bool
+
+	Len() int
+	Keys() iter.Seq[string]
+	Value(key string) any
 }

fs/ggml/ggml.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"iter"
 	"log/slog"
+	"maps"
 	"math"
 	"slices"
 	"strings"
@@ -239,6 +241,18 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
 	return val.values
 }
 
+func (kv KV) Len() int {
+	return len(kv)
+}
+
+func (kv KV) Keys() iter.Seq[string] {
+	return maps.Keys(kv)
+}
+
+func (kv KV) Value(key string) any {
+	return kv[key]
+}
+
 func (kv KV) OllamaEngineRequired() bool {
 	return slices.Contains([]string{
 		"bert",
@@ -255,6 +269,8 @@ func (kv KV) OllamaEngineRequired() bool {
 		"qwen25vl",
 		"qwen3", "qwen3moe",
 		"qwen3vl", "qwen3vlmoe",
+		"glm4moelite",
+		"lfm2",
 	}, kv.Architecture())
 }
 
@@ -842,7 +858,9 @@ func (f GGML) FlashAttention() bool {
 	return slices.Contains([]string{
 		"bert",
 		"gemma3",
+		"glm4moelite",
 		"gptoss", "gpt-oss",
+		"lfm2",
 		"mistral3",
 		"olmo3",
 		"qwen3", "qwen3moe",

fs/ggml/gguf.go

@@ -8,12 +8,12 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"maps"
 	"os"
 	"runtime"
 	"slices"
 	"strings"
 
+	"github.com/ollama/ollama/fs"
 	"golang.org/x/sync/errgroup"
 )
 
@@ -508,7 +508,7 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 	return binary.Write(w, binary.LittleEndian, s)
 }
 
-func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
+func WriteGGUF(f *os.File, kv fs.Config, ts []*Tensor) error {
 	arch := kv.String("general.architecture")
 	if arch == "" {
 		return fmt.Errorf("architecture not set")
@@ -526,12 +526,12 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		return err
 	}
 
-	if err := binary.Write(f, binary.LittleEndian, uint64(len(kv))); err != nil {
+	if err := binary.Write(f, binary.LittleEndian, uint64(kv.Len())); err != nil {
 		return err
 	}
 
-	for _, key := range slices.Sorted(maps.Keys(kv)) {
-		if err := ggufWriteKV(f, arch, key, kv[key]); err != nil {
+	for _, key := range slices.Sorted(kv.Keys()) {
+		if err := ggufWriteKV(f, arch, key, kv.Value(key)); err != nil {
 			return err
 		}
 	}

go.mod

@@ -28,6 +28,7 @@ require (
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	github.com/tkrajina/typescriptify-golang-structs v0.2.0
+	github.com/wk8/go-ordered-map/v2 v2.1.8
 	golang.org/x/image v0.22.0
 	golang.org/x/mod v0.30.0
 	golang.org/x/tools v0.38.0
@@ -36,6 +37,8 @@ require (
 
 require (
 	github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 // indirect
+	github.com/bahlo/generic-list-go v0.2.0 // indirect
+	github.com/buger/jsonparser v1.1.1 // indirect
 	github.com/bytedance/sonic/loader v0.1.1 // indirect
 	github.com/chewxy/hm v1.0.0 // indirect
 	github.com/chewxy/math32 v1.11.0 // indirect
@@ -45,6 +48,7 @@ require (
 	github.com/gogo/protobuf v1.3.2 // indirect
 	github.com/google/flatbuffers v24.3.25+incompatible // indirect
 	github.com/kr/text v0.2.0 // indirect
+	github.com/mailru/easyjson v0.7.7 // indirect
 	github.com/pkg/errors v0.9.1 // indirect
 	github.com/pmezard/go-difflib v1.0.0 // indirect
 	github.com/rivo/uniseg v0.2.0 // indirect

go.sum

@@ -14,7 +14,11 @@ github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40 h1:q4dksr6IC
 github.com/apache/arrow/go/arrow v0.0.0-20211112161151-bc219186db40/go.mod h1:Q7yQnSMnLvcXlZ8RV+jwz/6y1rQTqbX6C82SndT52Zs=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0 h1:jfIu9sQUG6Ig+0+Ap1h4unLjW6YQJpKZVmUzxsD4E/Q=
 github.com/arbovm/levenshtein v0.0.0-20160628152529-48b4e1c0c4d0/go.mod h1:t2tdKJDJF9BV14lnkjHmOQgcvEKgtqs5a1N3LNdJhGE=
+github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
+github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
 github.com/boombuler/barcode v1.0.0/go.mod h1:paBWMcWSl3LHKBqUq+rly7CNSldXjb2rDl3JlRe0mD8=
+github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=
+github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
 github.com/bytedance/sonic v1.11.6 h1:oUp34TzMlL+OY1OUWxHqsdkgC/Zfc85zGqw9siXjrc0=
 github.com/bytedance/sonic v1.11.6/go.mod h1:LysEHSvpvDySVdC2f87zGWf6CIKJcAvqab1ZaiQtds4=
 github.com/bytedance/sonic/loader v0.1.1 h1:c+e5Pt1k/cy5wMveRDyk2X4B9hF4g7an8N3zCYjJFNM=
@@ -123,6 +127,7 @@ github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+
 github.com/grpc-ecosystem/grpc-gateway v1.16.0/go.mod h1:BDjrQk3hbvj6Nolgz8mAMFbcEtjT1g+wF4CSlocrBnw=
 github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
 github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
+github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
 github.com/jung-kurt/gofpdf v1.0.0/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes=
@@ -143,6 +148,8 @@ github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728 h1:QwWKgMY28TAXaDl+
 github.com/ledongthuc/pdf v0.0.0-20250511090121-5959a4027728/go.mod h1:1fEHWurg7pvf5SG6XNE5Q8UZmOwex51Mkx3SLhrW5B4=
 github.com/leodido/go-urn v1.4.0 h1:WT9HwE9SGECu3lg4d/dIA+jxlljEa1/ffXKmRjqdmIQ=
 github.com/leodido/go-urn v1.4.0/go.mod h1:bvxc+MVxLKB4z00jd1z+Dvzr47oO32F/QSNjSBOlFxI=
+github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=
+github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
 github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=
 github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=
 github.com/mattn/go-runewidth v0.0.9/go.mod h1:H031xJmbD/WCDINGzjvQ9THkh0rPKHF+m2gUSrubnMI=
@@ -207,6 +214,8 @@ github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
 github.com/ugorji/go/codec v1.2.12/go.mod h1:UNopzCgEMSXjBc6AOMqYvWC1ktqTAfzJZUZgYf6w6lg=
+github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=
+github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
 github.com/xtgo/set v1.0.0 h1:6BCNBRv3ORNDQ7fyoJXRv+tstJz3m1JVFQErfeZz2pY=

integration/imagegen_test.go

@@ -0,0 +1,148 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"encoding/base64"
+	"fmt"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestImageGeneration(t *testing.T) {
+	skipUnderMinVRAM(t, 8)
+
+	type testCase struct {
+		imageGenModel string
+		visionModel   string
+		prompt        string
+		expectedWords []string
+	}
+
+	testCases := []testCase{
+		{
+			imageGenModel: "jmorgan/z-image-turbo",
+			visionModel:   "llama3.2-vision",
+			prompt:        "A cartoon style llama flying like a superhero through the air with clouds in the background",
+			expectedWords: []string{"llama", "flying", "cartoon", "cloud", "sky", "superhero", "air", "animal", "camelid"},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(fmt.Sprintf("%s->%s", tc.imageGenModel, tc.visionModel), func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 10*time.Minute)
+			defer cancel()
+
+			client, _, cleanup := InitServerConnection(ctx, t)
+			defer cleanup()
+
+			// Pull both models
+			if err := PullIfMissing(ctx, client, tc.imageGenModel); err != nil {
+				t.Fatalf("failed to pull image gen model: %v", err)
+			}
+			if err := PullIfMissing(ctx, client, tc.visionModel); err != nil {
+				t.Fatalf("failed to pull vision model: %v", err)
+			}
+
+			// Generate the image
+			t.Logf("Generating image with prompt: %s", tc.prompt)
+			imageBase64, err := generateImage(ctx, client, tc.imageGenModel, tc.prompt)
+			if err != nil {
+				if strings.Contains(err.Error(), "image generation not available") {
+					t.Skip("Target system does not support image generation")
+				} else if strings.Contains(err.Error(), "executable file not found in") { // Windows pattern, not yet supported
+					t.Skip("Windows does not support image generation yet")
+				} else if strings.Contains(err.Error(), "CUDA driver version is insufficient") {
+					t.Skip("Driver is too old")
+				} else if strings.Contains(err.Error(), "insufficient memory for image generation") {
+					t.Skip("insufficient memory for image generation")
+				} else if strings.Contains(err.Error(), "error while loading shared libraries: libcuda.so.1") { // AMD GPU or CPU
+					t.Skip("CUDA GPU is not available")
+				} else if strings.Contains(err.Error(), "ollama-mlx: no such file or directory") {
+					// most likely linux arm - not supported yet
+					t.Skip("unsupported architecture")
+				}
+				t.Fatalf("failed to generate image: %v", err)
+			}
+
+			imageData, err := base64.StdEncoding.DecodeString(imageBase64)
+			if err != nil {
+				t.Fatalf("failed to decode image: %v", err)
+			}
+			t.Logf("Generated image: %d bytes", len(imageData))
+
+			// Preload vision model and check GPU loading
+			err = client.Generate(ctx, &api.GenerateRequest{Model: tc.visionModel}, func(response api.GenerateResponse) error { return nil })
+			if err != nil {
+				t.Fatalf("failed to load vision model: %v", err)
+			}
+
+			// Use vision model to describe the image
+			chatReq := api.ChatRequest{
+				Model: tc.visionModel,
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "Describe this image in detail. What is shown? What style is it? What is the main subject doing?",
+						Images:  []api.ImageData{imageData},
+					},
+				},
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+			}
+
+			// Verify the vision model's response contains expected keywords
+			response := DoChat(ctx, t, client, chatReq, tc.expectedWords, 240*time.Second, 30*time.Second)
+			if response != nil {
+				t.Logf("Vision model response: %s", response.Content)
+
+				// Additional detailed check for keywords
+				content := strings.ToLower(response.Content)
+				foundWords := []string{}
+				missingWords := []string{}
+				for _, word := range tc.expectedWords {
+					if strings.Contains(content, word) {
+						foundWords = append(foundWords, word)
+					} else {
+						missingWords = append(missingWords, word)
+					}
+				}
+				t.Logf("Found keywords: %v", foundWords)
+				if len(missingWords) > 0 {
+					t.Logf("Missing keywords (at least one was found so test passed): %v", missingWords)
+				}
+			}
+		})
+	}
+}
+
+// generateImage calls the Ollama API to generate an image and returns the base64 image data
+func generateImage(ctx context.Context, client *api.Client, model, prompt string) (string, error) {
+	var imageBase64 string
+
+	err := client.Generate(ctx, &api.GenerateRequest{
+		Model:  model,
+		Prompt: prompt,
+	}, func(resp api.GenerateResponse) error {
+		if resp.Image != "" {
+			imageBase64 = resp.Image
+		}
+		return nil
+	})
+	if err != nil {
+		return "", fmt.Errorf("failed to generate image: %w", err)
+	}
+
+	if imageBase64 == "" {
+		return "", fmt.Errorf("no image data in response")
+	}
+
+	return imageBase64, nil
+}

integration/tools_test.go

@@ -11,6 +11,15 @@ import (
 	"github.com/ollama/ollama/api"
 )
 
+// testPropsMap creates a ToolPropertiesMap from a map (convenience function for tests)
+func testPropsMap(m map[string]api.ToolProperty) *api.ToolPropertiesMap {
+	props := api.NewToolPropertiesMap()
+	for k, v := range m {
+		props.Set(k, v)
+	}
+	return props
+}
+
 func TestAPIToolCalling(t *testing.T) {
 	initialTimeout := 60 * time.Second
 	streamTimeout := 60 * time.Second
@@ -57,12 +66,12 @@ func TestAPIToolCalling(t *testing.T) {
 						Parameters: api.ToolFunctionParameters{
 							Type:     "object",
 							Required: []string{"location"},
-							Properties: map[string]api.ToolProperty{
+							Properties: testPropsMap(map[string]api.ToolProperty{
 								"location": {
 									Type:        api.PropertyType{"string"},
 									Description: "The city and state, e.g. San Francisco, CA",
 								},
-							},
+							}),
 						},
 					},
 				},
@@ -122,7 +131,7 @@ func TestAPIToolCalling(t *testing.T) {
 					t.Errorf("unexpected tool called: got %q want %q", lastToolCall.Function.Name, "get_weather")
 				}
 
-				if _, ok := lastToolCall.Function.Arguments["location"]; !ok {
+				if _, ok := lastToolCall.Function.Arguments.Get("location"); !ok {
 					t.Errorf("expected tool arguments to include 'location', got: %s", lastToolCall.Function.Arguments.String())
 				}
 			case <-ctx.Done():

integration/utils_test.go

@@ -38,6 +38,7 @@ var (
 
 	// Note: add newer models at the top of the list to test them first
 	ollamaEngineChatModels = []string{
+		"lfm2.5-thinking",
 		"ministral-3",
 		"qwen3-coder:30b",
 		"gpt-oss:20b",
@@ -143,6 +144,7 @@ var (
 		"granite3.3",
 		"hermes3",
 		"internlm2",
+		"lfm2.5-thinking",
 		"llama-guard3",
 		"llama-pro",
 		"llama2-chinese",
@@ -263,6 +265,7 @@ var (
 		"snowflake-arctic-embed2",
 	}
 	libraryToolsModels = []string{
+		"lfm2.5-thinking",
 		"qwen3-vl",
 		"gpt-oss:20b",
 		"gpt-oss:120b",

internal/orderedmap/orderedmap.go

@@ -0,0 +1,94 @@
+// Package orderedmap provides a generic ordered map that maintains insertion order.
+// It wraps github.com/wk8/go-ordered-map/v2 to encapsulate the dependency.
+package orderedmap
+
+import (
+	"encoding/json"
+	"iter"
+
+	orderedmap "github.com/wk8/go-ordered-map/v2"
+)
+
+// Map is a generic ordered map that maintains insertion order.
+type Map[K comparable, V any] struct {
+	om *orderedmap.OrderedMap[K, V]
+}
+
+// New creates a new empty ordered map.
+func New[K comparable, V any]() *Map[K, V] {
+	return &Map[K, V]{
+		om: orderedmap.New[K, V](),
+	}
+}
+
+// Get retrieves a value by key.
+func (m *Map[K, V]) Get(key K) (V, bool) {
+	if m == nil || m.om == nil {
+		var zero V
+		return zero, false
+	}
+	return m.om.Get(key)
+}
+
+// Set sets a key-value pair. If the key already exists, its value is updated
+// but its position in the iteration order is preserved. If the key is new,
+// it is appended to the end.
+func (m *Map[K, V]) Set(key K, value V) {
+	if m == nil {
+		return
+	}
+	if m.om == nil {
+		m.om = orderedmap.New[K, V]()
+	}
+	m.om.Set(key, value)
+}
+
+// Len returns the number of entries.
+func (m *Map[K, V]) Len() int {
+	if m == nil || m.om == nil {
+		return 0
+	}
+	return m.om.Len()
+}
+
+// All returns an iterator over all key-value pairs in insertion order.
+func (m *Map[K, V]) All() iter.Seq2[K, V] {
+	return func(yield func(K, V) bool) {
+		if m == nil || m.om == nil {
+			return
+		}
+		for pair := m.om.Oldest(); pair != nil; pair = pair.Next() {
+			if !yield(pair.Key, pair.Value) {
+				return
+			}
+		}
+	}
+}
+
+// ToMap converts to a regular Go map.
+// Note: The resulting map does not preserve order.
+func (m *Map[K, V]) ToMap() map[K]V {
+	if m == nil || m.om == nil {
+		return nil
+	}
+	result := make(map[K]V, m.om.Len())
+	for pair := m.om.Oldest(); pair != nil; pair = pair.Next() {
+		result[pair.Key] = pair.Value
+	}
+	return result
+}
+
+// MarshalJSON implements json.Marshaler. The JSON output preserves key order.
+func (m *Map[K, V]) MarshalJSON() ([]byte, error) {
+	if m == nil || m.om == nil {
+		return []byte("null"), nil
+	}
+	return json.Marshal(m.om)
+}
+
+// UnmarshalJSON implements json.Unmarshaler. The insertion order matches the
+// order of keys in the JSON input.
+func (m *Map[K, V]) UnmarshalJSON(data []byte) error {
+	m.om = orderedmap.New[K, V]()
+	return json.Unmarshal(data, &m.om)
+}

internal/orderedmap/orderedmap_test.go

@@ -0,0 +1,348 @@
+package orderedmap
+
+import (
+	"encoding/json"
+	"slices"
+	"testing"
+)
+
+func TestMap_BasicOperations(t *testing.T) {
+	m := New[string, int]()
+
+	// Test empty map
+	if m.Len() != 0 {
+		t.Errorf("expected Len() = 0, got %d", m.Len())
+	}
+	v, ok := m.Get("a")
+	if ok {
+		t.Error("expected Get on empty map to return false")
+	}
+	if v != 0 {
+		t.Errorf("expected zero value, got %d", v)
+	}
+
+	// Test Set and Get
+	m.Set("a", 1)
+	m.Set("b", 2)
+	m.Set("c", 3)
+
+	if m.Len() != 3 {
+		t.Errorf("expected Len() = 3, got %d", m.Len())
+	}
+
+	v, ok = m.Get("a")
+	if !ok || v != 1 {
+		t.Errorf("expected Get(a) = (1, true), got (%d, %v)", v, ok)
+	}
+
+	v, ok = m.Get("b")
+	if !ok || v != 2 {
+		t.Errorf("expected Get(b) = (2, true), got (%d, %v)", v, ok)
+	}
+
+	v, ok = m.Get("c")
+	if !ok || v != 3 {
+		t.Errorf("expected Get(c) = (3, true), got (%d, %v)", v, ok)
+	}
+
+	// Test updating existing key preserves position
+	m.Set("a", 10)
+	v, ok = m.Get("a")
+	if !ok || v != 10 {
+		t.Errorf("expected Get(a) = (10, true), got (%d, %v)", v, ok)
+	}
+	if m.Len() != 3 {
+		t.Errorf("expected Len() = 3 after update, got %d", m.Len())
+	}
+}
+
+func TestMap_InsertionOrderPreserved(t *testing.T) {
+	m := New[string, int]()
+
+	// Insert in non-alphabetical order
+	m.Set("z", 1)
+	m.Set("a", 2)
+	m.Set("m", 3)
+	m.Set("b", 4)
+
+	// Verify iteration order matches insertion order
+	var keys []string
+	var values []int
+	for k, v := range m.All() {
+		keys = append(keys, k)
+		values = append(values, v)
+	}
+
+	expectedKeys := []string{"z", "a", "m", "b"}
+	expectedValues := []int{1, 2, 3, 4}
+
+	if !slices.Equal(keys, expectedKeys) {
+		t.Errorf("expected keys %v, got %v", expectedKeys, keys)
+	}
+	if !slices.Equal(values, expectedValues) {
+		t.Errorf("expected values %v, got %v", expectedValues, values)
+	}
+}
+
+func TestMap_UpdatePreservesPosition(t *testing.T) {
+	m := New[string, int]()
+
+	m.Set("first", 1)
+	m.Set("second", 2)
+	m.Set("third", 3)
+
+	// Update middle element
+	m.Set("second", 20)
+
+	var keys []string
+	for k := range m.All() {
+		keys = append(keys, k)
+	}
+
+	// Order should still be first, second, third
+	expected := []string{"first", "second", "third"}
+	if !slices.Equal(keys, expected) {
+		t.Errorf("expected keys %v, got %v", expected, keys)
+	}
+}
+
+func TestMap_MarshalJSON_PreservesOrder(t *testing.T) {
+	m := New[string, int]()
+
+	// Insert in non-alphabetical order
+	m.Set("z", 1)
+	m.Set("a", 2)
+	m.Set("m", 3)
+
+	data, err := json.Marshal(m)
+	if err != nil {
+		t.Fatalf("Marshal failed: %v", err)
+	}
+
+	// JSON should preserve insertion order, not alphabetical
+	expected := `{"z":1,"a":2,"m":3}`
+	if string(data) != expected {
+		t.Errorf("expected %s, got %s", expected, string(data))
+	}
+}
+
+func TestMap_UnmarshalJSON_PreservesOrder(t *testing.T) {
+	// JSON with non-alphabetical key order
+	jsonData := `{"z":1,"a":2,"m":3}`
+
+	m := New[string, int]()
+	if err := json.Unmarshal([]byte(jsonData), m); err != nil {
+		t.Fatalf("Unmarshal failed: %v", err)
+	}
+
+	// Verify iteration order matches JSON order
+	var keys []string
+	for k := range m.All() {
+		keys = append(keys, k)
+	}
+
+	expected := []string{"z", "a", "m"}
+	if !slices.Equal(keys, expected) {
+		t.Errorf("expected keys %v, got %v", expected, keys)
+	}
+}
+
+func TestMap_JSONRoundTrip(t *testing.T) {
+	// Test that unmarshal -> marshal produces identical JSON
+	original := `{"zebra":"z","apple":"a","mango":"m","banana":"b"}`
+
+	m := New[string, string]()
+	if err := json.Unmarshal([]byte(original), m); err != nil {
+		t.Fatalf("Unmarshal failed: %v", err)
+	}
+
+	data, err := json.Marshal(m)
+	if err != nil {
+		t.Fatalf("Marshal failed: %v", err)
+	}
+
+	if string(data) != original {
+		t.Errorf("round trip failed: expected %s, got %s", original, string(data))
+	}
+}
+
+func TestMap_ToMap(t *testing.T) {
+	m := New[string, int]()
+	m.Set("a", 1)
+	m.Set("b", 2)
+
+	regular := m.ToMap()
+
+	if len(regular) != 2 {
+		t.Errorf("expected len 2, got %d", len(regular))
+	}
+	if regular["a"] != 1 {
+		t.Errorf("expected regular[a] = 1, got %d", regular["a"])
+	}
+	if regular["b"] != 2 {
+		t.Errorf("expected regular[b] = 2, got %d", regular["b"])
+	}
+}
+
+func TestMap_NilSafety(t *testing.T) {
+	var m *Map[string, int]
+
+	// All operations should be safe on nil
+	if m.Len() != 0 {
+		t.Errorf("expected Len() = 0 on nil map, got %d", m.Len())
+	}
+
+	v, ok := m.Get("a")
+	if ok {
+		t.Error("expected Get on nil map to return false")
+	}
+	if v != 0 {
+		t.Errorf("expected zero value from nil map, got %d", v)
+	}
+
+	// Set on nil is a no-op
+	m.Set("a", 1)
+	if m.Len() != 0 {
+		t.Errorf("expected Len() = 0 after Set on nil, got %d", m.Len())
+	}
+
+	// All returns empty iterator
+	var keys []string
+	for k := range m.All() {
+		keys = append(keys, k)
+	}
+	if len(keys) != 0 {
+		t.Errorf("expected empty iteration on nil map, got %v", keys)
+	}
+
+	// ToMap returns nil
+	if m.ToMap() != nil {
+		t.Error("expected ToMap to return nil on nil map")
+	}
+
+	// MarshalJSON returns null
+	data, err := json.Marshal(m)
+	if err != nil {
+		t.Fatalf("Marshal failed: %v", err)
+	}
+	if string(data) != "null" {
+		t.Errorf("expected null, got %s", string(data))
+	}
+}
+
+func TestMap_EmptyMapMarshal(t *testing.T) {
+	m := New[string, int]()
+
+	data, err := json.Marshal(m)
+	if err != nil {
+		t.Fatalf("Marshal failed: %v", err)
+	}
+	if string(data) != "{}" {
+		t.Errorf("expected {}, got %s", string(data))
+	}
+}
+
+func TestMap_NestedValues(t *testing.T) {
+	m := New[string, any]()
+	m.Set("string", "hello")
+	m.Set("number", 42)
+	m.Set("bool", true)
+	m.Set("nested", map[string]int{"x": 1})
+
+	data, err := json.Marshal(m)
+	if err != nil {
+		t.Fatalf("Marshal failed: %v", err)
+	}
+
+	expected := `{"string":"hello","number":42,"bool":true,"nested":{"x":1}}`
+	if string(data) != expected {
+		t.Errorf("expected %s, got %s", expected, string(data))
+	}
+}
+
+func TestMap_AllIteratorEarlyExit(t *testing.T) {
+	m := New[string, int]()
+	m.Set("a", 1)
+	m.Set("b", 2)
+	m.Set("c", 3)
+	m.Set("d", 4)
+
+	// Collect only first 2
+	var keys []string
+	for k := range m.All() {
+		keys = append(keys, k)
+		if len(keys) == 2 {
+			break
+		}
+	}
+
+	expected := []string{"a", "b"}
+	if !slices.Equal(keys, expected) {
+		t.Errorf("expected %v, got %v", expected, keys)
+	}
+}
+
+func TestMap_IntegerKeys(t *testing.T) {
+	m := New[int, string]()
+	m.Set(3, "three")
+	m.Set(1, "one")
+	m.Set(2, "two")
+
+	var keys []int
+	for k := range m.All() {
+		keys = append(keys, k)
+	}
+
+	// Should preserve insertion order, not numerical order
+	expected := []int{3, 1, 2}
+	if !slices.Equal(keys, expected) {
+		t.Errorf("expected %v, got %v", expected, keys)
+	}
+}
+
+func TestMap_UnmarshalIntoExisting(t *testing.T) {
+	m := New[string, int]()
+	m.Set("existing", 999)
+
+	// Unmarshal should replace contents
+	if err := json.Unmarshal([]byte(`{"new":1}`), m); err != nil {
+		t.Fatalf("Unmarshal failed: %v", err)
+	}
+
+	_, ok := m.Get("existing")
+	if ok {
+		t.Error("existing key should be gone after unmarshal")
+	}
+
+	v, ok := m.Get("new")
+	if !ok || v != 1 {
+		t.Errorf("expected Get(new) = (1, true), got (%d, %v)", v, ok)
+	}
+}
+
+func TestMap_LargeOrderPreservation(t *testing.T) {
+	m := New[string, int]()
+
+	// Create many keys in specific order
+	keys := make([]string, 100)
+	for i := range 100 {
+		keys[i] = string(rune('a' + (99 - i))) // reverse order: 'd', 'c', 'b', 'a' (extended)
+		if i >= 26 {
+			keys[i] = string(rune('A'+i-26)) + string(rune('a'+i%26))
+		}
+	}
+
+	for i, k := range keys {
+		m.Set(k, i)
+	}
+
+	// Verify order preserved
+	var resultKeys []string
+	for k := range m.All() {
+		resultKeys = append(resultKeys, k)
+	}
+
+	if !slices.Equal(keys, resultKeys) {
+		t.Error("large map should preserve insertion order")
+	}
+}

llama/README.md

@@ -14,25 +14,28 @@ make -f Makefile.sync apply-patches
 
 ### Updating Base Commit
 
-**Pin to new base commit**
+To update to a new base commit:
 
-To change the base commit, update `FETCH_HEAD` in Makefile.sync.
+1. **Update FETCH_HEAD** in `Makefile.sync` to the new commit hash.
 
-When updating to a newer base commit, the existing patches may not apply cleanly and require manual merge resolution.
+2. **Check for upstreamed patches**: Before applying, review if any patches have been merged upstream. Remove those patches from `./patches/` to avoid conflicts.
 
-Start by applying the patches. If any of the patches have conflicts, the `git am` will stop at the first failure.
+3. **Apply patches**:
+   ```shell
+   make -f Makefile.sync apply-patches
+   ```
 
-```shell
-make -f Makefile.sync apply-patches
-```
+4. **Resolve conflicts** (if any): When `git am` fails on a patch:
+   - Fix conflicts in `./vendor/`
+   - Stage the resolved files: `git -C llama/vendor add <file>`
+   - Continue: `git -C llama/vendor am --continue`
+   - Re-run: `make -f Makefile.sync apply-patches`
+   - Repeat until all patches are applied.
 
-If there are conflicts, you will see an error message. Resolve the conflicts in `./vendor/`, and continue the patch series with `git am --continue` and rerun `make -f Makefile.sync apply-patches`. Repeat until all patches are successfully applied.
-
-Once all patches are applied, commit the changes to the tracking repository.
-
-```shell
-make -f Makefile.sync format-patches sync
-```
+5. **Regenerate patches and sync**:
+   ```shell
+   make -f Makefile.sync format-patches sync
+   ```
 
 ### Generating Patches
 

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "ec98e2002";
+char const *LLAMA_COMMIT = "51fa458a92d6a3f305f8fd76fc8f702e3e87ddb5";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/common.cpp

@@ -251,7 +251,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
         case GGML_SCHED_PRIO_REALTIME: p = -20; break;
     }
 
-    if (!setpriority(PRIO_PROCESS, 0, p)) {
+    if (setpriority(PRIO_PROCESS, 0, p) != 0) {
         LOG_WRN("failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
         return false;
     }
@@ -1078,12 +1078,15 @@ struct common_init_result::impl {
     impl() = default;
     ~impl() = default;
 
+    // note: the order in which model, context, etc. are declared matters because their destructors will be called bottom-to-top
+
     llama_model_ptr   model;
     llama_context_ptr context;
 
     std::vector<llama_adapter_lora_ptr> lora;
 
     std::vector<common_sampler_ptr> samplers;
+    std::vector<llama_sampler_seq_config> samplers_seq_config;
 };
 
 common_init_result::common_init_result(common_params & params) :
@@ -1092,9 +1095,9 @@ common_init_result::common_init_result(common_params & params) :
     auto cparams = common_context_params_to_llama(params);
 
     if (params.fit_params) {
-        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1107,6 +1110,25 @@ common_init_result::common_init_result(common_params & params) :
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
+    // load and optionally apply lora adapters (must be loaded before context creation)
+    for (auto & la : params.lora_adapters) {
+        llama_adapter_lora_ptr lora;
+        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
+        if (lora == nullptr) {
+            LOG_ERR("%s: failed to load lora adapter '%s'\n", __func__, la.path.c_str());
+            pimpl->model.reset(model);
+            return;
+        }
+
+        char buf[1024];
+        la.ptr = lora.get();
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
+        la.task_name = buf;
+        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
+        la.prompt_prefix = buf;
+        pimpl->lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+    }
+
     // updates params.sampling
     // TODO: fix naming
     common_init_sampler_from_model(model, params.sampling);
@@ -1141,10 +1163,18 @@ common_init_result::common_init_result(common_params & params) :
     //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
     //}
 
+    // init the backend samplers as part of the context creation
     pimpl->samplers.resize(cparams.n_seq_max);
+    pimpl->samplers_seq_config.resize(cparams.n_seq_max);
 
     for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
         pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+        pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
+    }
+
+    if (params.sampling.backend_sampling) {
+        cparams.samplers   = pimpl->samplers_seq_config.data();
+        cparams.n_samplers = pimpl->samplers_seq_config.size();
     }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -1168,6 +1198,12 @@ common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
     return pimpl->samplers[seq_id].get();
 }
 
+void common_init_result::reset_samplers() {
+    for (int i = 0; i < (int) pimpl->samplers.size(); ++i) {
+        llama_sampler_reset(common_sampler_get(pimpl->samplers[i].get()));
+    }
+}
+
 std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
@@ -1243,24 +1279,6 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         }
     }
 
-    // load and optionally apply lora adapters
-    for (auto & la : params.lora_adapters) {
-        llama_adapter_lora_ptr lora;
-        lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
-        if (lora == nullptr) {
-            LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
-            return res;
-        }
-
-        char buf[1024];
-        la.ptr = lora.get();
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.task_name", buf, sizeof(buf));
-        la.task_name = buf;
-        llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
-        la.prompt_prefix = buf;
-        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
-    }
-
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
@@ -1301,6 +1319,9 @@ common_init_result_ptr common_init_from_params(common_params & params) {
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
+
+        // reset samplers to reset RNG state after warmup to the seeded state
+        res->reset_samplers();
     }
 
     return res;
@@ -1339,14 +1360,12 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.devices = params.devices.data();
     }
 
-    if (params.n_gpu_layers != -1) {
-        mparams.n_gpu_layers = params.n_gpu_layers;
-    }
-
+    mparams.n_gpu_layers    = params.n_gpu_layers;
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
+    mparams.use_direct_io   = params.use_direct_io;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

llama/llama.cpp/common/common.h

@@ -57,6 +57,8 @@ extern const char * LLAMA_COMMIT;
 extern const char * LLAMA_COMPILER;
 extern const char * LLAMA_BUILD_TARGET;
 
+const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
+
 struct common_control_vector_load_info;
 
 //
@@ -80,6 +82,8 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
+    LLAMA_EXAMPLE_BATCHED,
+    LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_COMPLETION,
@@ -117,6 +121,7 @@ enum common_sampler_type {
     COMMON_SAMPLER_TYPE_INFILL      = 9,
     COMMON_SAMPLER_TYPE_PENALTIES   = 10,
     COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11,
+    COMMON_SAMPLER_TYPE_ADAPTIVE_P  = 12,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -164,32 +169,34 @@ enum common_params_sampling_config : uint64_t {
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev             = 64;    // number of previous tokens to remember
-    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k              = 40;    // <= 0 to use vocab size
-    float   top_p              = 0.95f; // 1.0 = disabled
-    float   min_p              = 0.05f; // 0.0 = disabled
-    float   xtc_probability    = 0.00f; // 0.0 = disabled
-    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
-    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
-    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range     = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat     = 1.00f; // 1.0 = disabled
-    float   penalty_freq       = 0.00f; // 0.0 = disabled
-    float   penalty_present    = 0.00f; // 0.0 = disabled
-    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
-    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
-    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
-    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   top_n_sigma        = -1.00f;// -1.0 = disabled
-    float   mirostat_tau       = 5.00f; // target entropy
-    float   mirostat_eta       = 0.10f; // learning rate
+    int32_t n_prev             = 64;     // number of previous tokens to remember
+    int32_t n_probs            = 0;      // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep           = 0;      // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k              = 40;     // <= 0 to use vocab size
+    float   top_p              = 0.95f;  // 1.0 = disabled
+    float   min_p              = 0.05f;  // 0.0 = disabled
+    float   xtc_probability    = 0.00f;  // 0.0 = disabled
+    float   xtc_threshold      = 0.10f;  // > 0.5 disables XTC
+    float   typ_p              = 1.00f;  // typical_p, 1.0 = disabled
+    float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range     = 0.00f;  // 0.0 = disabled
+    float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat     = 1.00f;  // 1.0 = disabled
+    float   penalty_freq       = 0.00f;  // 0.0 = disabled
+    float   penalty_present    = 0.00f;  // 0.0 = disabled
+    float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
+    float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
+    int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
+    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
+    float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
+    int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   top_n_sigma        = -1.00f; // -1.0 = disabled
+    float   mirostat_tau       = 5.00f;  // target entropy
+    float   mirostat_eta       = 0.10f;  // learning rate
     bool    ignore_eos         = false;
-    bool    no_perf            = false; // disable performance metrics
+    bool    no_perf            = false;  // disable performance metrics
     bool    timing_per_token   = false;
 
     uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers
@@ -216,6 +223,8 @@ struct common_params_sampling {
     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
 
+    bool backend_sampling = false;
+
     bool has_logit_bias() const {
         return !logit_bias.empty();
     }
@@ -277,6 +286,7 @@ struct common_params_diffusion {
 };
 
 // reasoning API response format (not to be confused as chat template's reasoning format)
+// only used by server
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
@@ -329,12 +339,14 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
-    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
-    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
-    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
+    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
+    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+
+    // margin per device in bytes for fitting parameters to free memory:
+    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -370,6 +382,11 @@ struct common_params {
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
+    // llama-debug specific options
+    std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
+    bool        save_logits       = false;  // whether to save logits to files                              // NOLINT
+    std::vector<std::string> tensor_filter; // filter tensor names for debug output (regex)                 // NOLINT
+
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -420,7 +437,8 @@ struct common_params {
     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
+    bool use_mmap          = true;  // enable mmap to use filesystem cache
+    bool use_direct_io     = true;  // read from disk without buffering for faster model loading
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
@@ -464,6 +482,7 @@ struct common_params {
     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
+    bool    cache_prompt      = true;         // whether to enable prompt caching
     int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
     int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
@@ -475,7 +494,8 @@ struct common_params {
     bool enable_chat_template = true;
     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
+    bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
+    int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
     std::vector<std::string> api_keys;
 
@@ -484,8 +504,11 @@ struct common_params {
 
     std::map<std::string, std::string> default_template_kwargs;
 
+    // webui configs
+    bool webui = true;
+    std::string webui_config_json;
+
     // "advanced" endpoints are disabled by default for better security
-    bool webui            = true;
     bool endpoint_slots   = true;
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
@@ -685,7 +708,9 @@ struct common_init_result {
 
     llama_model * model();
     llama_context * context();
+
     common_sampler * sampler(llama_seq_id seq_id);
+    void reset_samplers();
 
     std::vector<llama_adapter_lora_ptr> & lora();
 

llama/llama.cpp/common/sampling.cpp

@@ -104,10 +104,9 @@ struct ring_buffer {
 struct common_sampler {
     common_params_sampling params;
 
+    struct llama_sampler * grmr;
     struct llama_sampler * chain;
 
-    bool grammar;
-
     ring_buffer<llama_token> prev;
 
     std::vector<llama_token_data> cur;
@@ -121,17 +120,34 @@ struct common_sampler {
     }
 
     void set_logits(struct llama_context * ctx, int idx) {
-        const auto * logits = llama_get_logits_ith(ctx, idx);
+        const float *       sampled_probs  = llama_get_sampled_probs_ith     (ctx, idx);
+        const float *       sampled_logits = llama_get_sampled_logits_ith    (ctx, idx);
+        const llama_token * sampled_ids    = llama_get_sampled_candidates_ith(ctx, idx);
 
         const llama_model * model = llama_get_model(ctx);
         const llama_vocab * vocab = llama_model_get_vocab(model);
 
         const int n_vocab = llama_vocab_n_tokens(vocab);
 
-        cur.resize(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        if (sampled_probs) {
+            const uint32_t sampled_probs_count = llama_get_sampled_probs_count_ith(ctx, idx);
+            cur.resize(sampled_probs_count);
+            for (uint32_t i = 0; i < sampled_probs_count; ++i) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], sampled_probs[i]};
+            }
+        } else if (sampled_logits) {
+            const uint32_t sampled_logits_count = llama_get_sampled_logits_count_ith(ctx, idx);
+            cur.resize(sampled_logits_count);
+            for (uint32_t i = 0; i < sampled_logits_count; i++) {
+                cur[i] = llama_token_data{sampled_ids[i], sampled_logits[i], 0.0f};
+            }
+        } else {
+            const auto * logits = llama_get_logits_ith(ctx, idx);
+            GGML_ASSERT(logits != nullptr);
+            cur.resize(n_vocab);
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+            }
         }
 
         cur_p = { cur.data(), cur.size(), -1, false };
@@ -151,54 +167,59 @@ std::string common_params_sampling::print() const {
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
             "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
             "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, top_n_sigma = %.3f, temp = %.3f\n"
-            "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
+            "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f, adaptive_target = %.3f, adaptive_decay = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
             dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
             top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp,
-            mirostat, mirostat_eta, mirostat_tau);
+            mirostat, mirostat_eta, mirostat_tau, adaptive_target, adaptive_decay);
 
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
 
+    llama_sampler * grmr = nullptr;
     llama_sampler * chain = llama_sampler_chain_init(lparams);
 
-    bool grammar = false;
     std::vector<llama_sampler *> samplers;
 
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
-        grammar = true;
+        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
         std::vector<std::string> trigger_patterns;
-        std::vector<std::string> patterns_anywhere;
         std::vector<llama_token> trigger_tokens;
         for (const auto & trigger : params.grammar_triggers) {
             switch (trigger.type) {
                 case COMMON_GRAMMAR_TRIGGER_TYPE_WORD:
                 {
                     const auto & word = trigger.value;
-                    patterns_anywhere.push_back(regex_escape(word));
+                    trigger_patterns.push_back(regex_escape(word));
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
                 {
-                    patterns_anywhere.push_back(trigger.value);
+                    trigger_patterns.push_back(trigger.value);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL:
                 {
-                    trigger_patterns.push_back(trigger.value);
+                    const auto & pattern = trigger.value;
+                    std::string anchored = "^$";
+                    if (!pattern.empty()) {
+                        anchored = (pattern.front() != '^' ? "^" : "")
+                            + pattern
+                            + (pattern.back() != '$' ? "$" : "");
+                    }
+                    trigger_patterns.push_back(anchored);
                     break;
                 }
                 case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
@@ -212,10 +233,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             }
         }
 
-        if (!patterns_anywhere.empty()) {
-            trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*");
-        }
-
         std::vector<const char *> trigger_patterns_c;
         trigger_patterns_c.reserve(trigger_patterns.size());
         for (const auto & regex : trigger_patterns) {
@@ -224,15 +241,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
         if (!params.grammar.empty()) {
              if (params.grammar_lazy) {
-                 samplers.push_back(
-                         llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
-                             trigger_patterns_c.data(), trigger_patterns_c.size(),
-                             trigger_tokens.data(),     trigger_tokens.size()));
+                 grmr = llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                         trigger_patterns_c.data(), trigger_patterns_c.size(),
+                         trigger_tokens.data(), trigger_tokens.size());
              } else {
-                 samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
+                 grmr = llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
              }
-
-             grammar = true;
         }
     }
 
@@ -241,6 +255,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
     }
 
     if (params.mirostat == 0) {
+
+        bool use_adaptive_p = false; // see below
+
         for (const auto & cnstr : params.samplers) {
             switch (cnstr) {
                 case COMMON_SAMPLER_TYPE_DRY:
@@ -250,43 +267,54 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                         for (const auto & str : params.dry_sequence_breakers) {
                             c_breakers.push_back(str.c_str());
                         }
-
-                        samplers.push_back(llama_sampler_init_dry    (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        samplers.push_back(llama_sampler_init_dry(vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                     }
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
-                    samplers.push_back(llama_sampler_init_top_k      (params.top_k));
+                    samplers.push_back(llama_sampler_init_top_k(params.top_k));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_P:
-                    samplers.push_back(llama_sampler_init_top_p      (params.top_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_top_p(params.top_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
                     samplers.push_back(llama_sampler_init_top_n_sigma(params.top_n_sigma));
                     break;
                 case COMMON_SAMPLER_TYPE_MIN_P:
-                    samplers.push_back(llama_sampler_init_min_p      (params.min_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_min_p(params.min_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_XTC:
-                    samplers.push_back(llama_sampler_init_xtc        (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    samplers.push_back(llama_sampler_init_xtc(params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
                     break;
                 case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    samplers.push_back(llama_sampler_init_typical    (params.typ_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_typical(params.typ_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    samplers.push_back(llama_sampler_init_temp_ext   (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    samplers.push_back(llama_sampler_init_temp_ext(params.temp, params.dynatemp_range, params.dynatemp_exponent));
                     break;
                 case COMMON_SAMPLER_TYPE_INFILL:
-                    samplers.push_back(llama_sampler_init_infill     (vocab));
+                    samplers.push_back(llama_sampler_init_infill(vocab));
                     break;
                 case COMMON_SAMPLER_TYPE_PENALTIES:
-                    samplers.push_back(llama_sampler_init_penalties  (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    break;
+                case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
+                    // the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects
+                    // a single token, so we will add `dist` at the end of the chain by default,
+                    // unless the user specifically included `adaptive-p`. we set this flag here
+                    // so we know to add the sampler at the very end.
+                    use_adaptive_p = true;
                     break;
                 default:
                     GGML_ASSERT(false && "unknown sampler type");
             }
         }
-
-        samplers.push_back(llama_sampler_init_dist(params.seed));
+        if (use_adaptive_p) {
+            // only if user explicitly included adaptive-p sampler
+            samplers.push_back(llama_sampler_init_adaptive_p(params.adaptive_target, params.adaptive_decay, params.seed));
+        } else {
+            // default: sample from distribution
+            samplers.push_back(llama_sampler_init_dist(params.seed));
+        }
     } else if (params.mirostat == 1) {
         samplers.push_back(llama_sampler_init_temp(params.temp));
         samplers.push_back(llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
@@ -301,10 +329,16 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         llama_sampler_chain_add(chain, smpl);
     }
 
+    if (grmr && params.backend_sampling) {
+        LOG_WRN("%s: backend sampling is not compatible with grammar, disabling\n", __func__);
+
+        params.backend_sampling = false;
+    }
+
     auto * result = new common_sampler {
         /* .params  = */ params,
+        /* .grmr    = */ grmr,
         /* .chain   = */ chain,
-        /* .grammar = */ grammar,
         /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur     = */ {},
         /* .cur_p   = */ {},
@@ -314,47 +348,45 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 }
 
 void common_sampler_free(struct common_sampler * gsmpl) {
-    if (gsmpl) {
-        llama_sampler_free(gsmpl->chain);
-
-        delete gsmpl;
+    if (!gsmpl) {
+        return;
     }
+
+    llama_sampler_free(gsmpl->grmr);
+    llama_sampler_free(gsmpl->chain);
+
+    delete gsmpl;
 }
 
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    if (!gsmpl) {
+        return;
+    }
+
     const auto tm = gsmpl->tm();
 
-    if (gsmpl->grammar) {
-        const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
-
-        for (int i = 0; i < n_smpl; i++) {
-            auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-
-            // the grammar sampler is always the first one
-            if (i == 0) {
-                if (accept_grammar) {
-                    llama_sampler_accept(smpl, token);
-                }
-            } else {
-                llama_sampler_accept(smpl, token);
-            }
-        }
-    } else {
-        llama_sampler_accept(gsmpl->chain, token);
+    if (gsmpl->grmr && accept_grammar) {
+        llama_sampler_accept(gsmpl->grmr, token);
     }
 
+    llama_sampler_accept(gsmpl->chain, token);
+
     gsmpl->prev.push_back(token);
 }
 
 void common_sampler_reset(struct common_sampler * gsmpl) {
+    if (!gsmpl) {
+        return;
+    }
+
     gsmpl->reset();
 }
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
         /* .params  = */ gsmpl->params,
+        /* .grmr    = */ llama_sampler_clone(gsmpl->grmr),
         /* .chain   = */ llama_sampler_clone(gsmpl->chain),
-        /* .grammar = */ gsmpl->grammar,
         /* .prev    = */ gsmpl->prev,
         /* .cur     = */ gsmpl->cur,
         /* .cur_p   = */ gsmpl->cur_p,
@@ -407,10 +439,14 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
 }
 
 struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
+    if (!gsmpl) {
+        return nullptr;
+    }
+
     return gsmpl->chain;
 }
 
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
     llama_synchronize(ctx);
 
     // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
@@ -418,11 +454,61 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 
     llama_token id = LLAMA_TOKEN_NULL;
 
+    auto & grmr  = gsmpl->grmr;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
+    // Check if a backend sampler has already sampled a token in which case we
+    // return that token id directly.
+    {
+        id = llama_get_sampled_token_ith(ctx, idx);
+
+        if (id != LLAMA_TOKEN_NULL) {
+            LOG_DBG("%s: Backend sampler selected token: '%d'. Will not run any CPU samplers\n", __func__, id);
+
+            GGML_ASSERT(!gsmpl->grmr && "using grammar in combination with backend sampling is not supported");
+
+            // TODO: simplify
+            gsmpl->cur.resize(1);
+            gsmpl->cur[0] = { id, 0.0f, 1.0f };
+            cur_p = { gsmpl->cur.data(), gsmpl->cur.size(), 0, true };
+
+            return id;
+        }
+    }
+
     gsmpl->set_logits(ctx, idx);
 
+    if (grammar_first) {
+        llama_sampler_apply(grmr, &cur_p);
+    }
+
+    llama_sampler_apply(chain, &cur_p);
+
+    id = cur_p.data[cur_p.selected].id;
+
+    if (grammar_first) {
+        return id;
+    }
+
+    // check if it the sampled token fits the grammar (grammar-based rejection sampling)
+    {
+        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
+
+        llama_sampler_apply(grmr, &single_token_data_array);
+
+        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+        if (is_valid) {
+            return id;
+        }
+    }
+
+    // resampling:
+    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
+    gsmpl->set_logits(ctx, idx);
+
+    llama_sampler_apply(grmr,  &cur_p);
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
@@ -432,7 +518,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     return id;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
     GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
 
     std::vector<llama_token> result;
@@ -440,7 +526,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
 
     size_t i = 0;
     for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -452,7 +538,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     }
 
     if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -462,13 +548,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     return result;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
     std::vector<int> idxs(draft.size() + 1);
     for (size_t i = 0; i < idxs.size(); ++i) {
         idxs[i] = i;
     }
 
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
 }
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
@@ -553,6 +639,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_XTC:         return 'x';
         case COMMON_SAMPLER_TYPE_INFILL:      return 'i';
         case COMMON_SAMPLER_TYPE_PENALTIES:   return 'e';
+        case COMMON_SAMPLER_TYPE_ADAPTIVE_P:  return 'a';
         default : return '?';
     }
 }
@@ -569,6 +656,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
         case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
         case COMMON_SAMPLER_TYPE_INFILL:      return "infill";
         case COMMON_SAMPLER_TYPE_PENALTIES:   return "penalties";
+        case COMMON_SAMPLER_TYPE_ADAPTIVE_P:  return "adaptive_p";
         default : return "";
     }
 }
@@ -585,6 +673,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "xtc",         COMMON_SAMPLER_TYPE_XTC },
         { "infill",      COMMON_SAMPLER_TYPE_INFILL },
         { "penalties",   COMMON_SAMPLER_TYPE_PENALTIES },
+        { "adaptive_p",  COMMON_SAMPLER_TYPE_ADAPTIVE_P },
     };
 
     // since samplers names are written multiple ways
@@ -600,6 +689,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
         { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
+        { "adaptive-p",  COMMON_SAMPLER_TYPE_ADAPTIVE_P },
     };
 
     std::vector<common_sampler_type> samplers;
@@ -636,6 +726,7 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL),      COMMON_SAMPLER_TYPE_INFILL },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_PENALTIES),   COMMON_SAMPLER_TYPE_PENALTIES },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_ADAPTIVE_P),  COMMON_SAMPLER_TYPE_ADAPTIVE_P },
     };
 
     std::vector<common_sampler_type> samplers;

llama/llama.cpp/common/sampling.h

@@ -36,7 +36,8 @@ struct common_sampler;
 
 // llama_sampler API overloads
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params);
+// note: can mutate params in some cases
+struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
@@ -48,6 +49,7 @@ struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
+// get the underlying llama_sampler_chain
 struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 
 // extended sampling implementation:
@@ -57,7 +59,10 @@ struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
 // - check if the token fits the grammar (if any)
 // - if not: resample by first applying the grammar constraints and then sampling again (slower path)
 //
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx);
+// if grammar_first is true, the grammar is applied before the samplers (slower)
+// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
+//
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
 
 // generalized version of common_sampler_sample
 //
@@ -75,10 +80,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 //
 // returns at least 1 token, up to idxs.size()
 //
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
 
 // assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 

llama/llama.cpp/include/llama-cpp.h

@@ -21,7 +21,9 @@ struct llama_sampler_deleter {
 };
 
 struct llama_adapter_lora_deleter {
-    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
+    void operator()(llama_adapter_lora *) {
+        // llama_adapter_lora_free is deprecated
+    }
 };
 
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;

llama/llama.cpp/include/llama.h

@@ -286,7 +286,7 @@ extern "C" {
         // NULL-terminated list of buffer types to use for tensors that match a pattern
         const struct llama_model_tensor_buft_override * tensor_buft_overrides;
 
-        int32_t n_gpu_layers; // number of layers to store in VRAM
+        int32_t n_gpu_layers; // number of layers to store in VRAM, a negative value means all layers
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
         // the GPU that is used for the entire model when split_mode is LLAMA_SPLIT_MODE_NONE
@@ -309,6 +309,7 @@ extern "C" {
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only;      // only load the vocabulary, no weights
         bool use_mmap;        // use mmap if possible
+        bool use_direct_io;   // use direct io, takes precedence over use_mmap
         bool use_mlock;       // force system to keep model in RAM
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
@@ -316,6 +317,11 @@ extern "C" {
         bool no_alloc;        // only load metadata and simulate memory allocations
     };
 
+    struct llama_sampler_seq_config {
+        llama_seq_id           seq_id;
+        struct llama_sampler * sampler;
+    };
+
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
     //       https://github.com/ggml-org/llama.cpp/pull/7544
     struct llama_context_params {
@@ -364,6 +370,12 @@ extern "C" {
         bool kv_unified;  // use a unified buffer across the input sequences when computing the attention
                           // try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
                           // ref: https://github.com/ggml-org/llama.cpp/pull/14363
+
+        // [EXPERIMENTAL]
+        // backend sampler chain configuration (make sure the caller keeps the sampler chains alive)
+        // note: the samplers must be sampler chains (i.e. use llama_sampler_chain_init)
+        struct llama_sampler_seq_config * samplers;
+        size_t                            n_samplers;
     };
 
     // model quantization parameters
@@ -467,16 +479,23 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
+    enum llama_params_fit_status {
+        LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
+        LLAMA_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
+        LLAMA_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occured, e.g. because no model could be found at the specified path
+    };
+
     // fits mparams and cparams to free device memory (assumes system memory is unlimited)
-    // returns true if the parameters could be successfully modified to fit device memory
-    // this function is NOT thread safe because it modifies the global llama logger state
-    LLAMA_API bool llama_params_fit(
+    //   - returns true if the parameters could be successfully modified to fit device memory
+    //   - this function is NOT thread safe because it modifies the global llama logger state
+    //   - only parameters that have the same value as in llama_default_model_params are modified
+    LLAMA_API enum llama_params_fit_status llama_params_fit(
                                    const char   * path_model,
                     struct llama_model_params   * mparams,
                     struct llama_context_params * cparams,
                                           float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
         struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
-                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                         size_t * margins,               // margins of memory to leave per device in bytes
                                        uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
                             enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
@@ -517,6 +536,7 @@ extern "C" {
     LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_out (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
@@ -600,6 +620,8 @@ extern "C" {
     //
 
     // Load a LoRA adapter from file
+    // The adapter is valid as long as the associated model is not freed
+    // All adapters must be loaded before context creation
     LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init(
             struct llama_model * model,
             const char * path_lora);
@@ -624,7 +646,8 @@ extern "C" {
 
     // Manually free a LoRA adapter
     // NOTE: loaded adapters will be free when the associated model is deleted
-    LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter);
+    LLAMA_API DEPRECATED(void llama_adapter_lora_free(struct llama_adapter_lora * adapter),
+            "adapters are now freed together with the associated model");
 
     // Get the invocation tokens if the current lora is an alora
     LLAMA_API uint64_t            llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter);
@@ -983,6 +1006,32 @@ extern "C" {
     // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 
+    //
+    // backend sampling API [EXPERIMENTAL]
+    // note: use only if the llama_context was created with at least one llama_sampler_seq_config
+    //
+
+    // Get the backend sampled token for the ith token.
+    // Returns LLAMA_TOKEN_NULL if no token was sampled.
+    LLAMA_API llama_token llama_get_sampled_token_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled probabilites for the ith token
+    // The index matches llama_get_sampled_token_ith().
+    // Returns NULL if no probabilites were generated.
+    LLAMA_API float *  llama_get_sampled_probs_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t llama_get_sampled_probs_count_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled logits for the ith token
+    // Returns NULL if no logits were sampled.
+    LLAMA_API float *  llama_get_sampled_logits_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t llama_get_sampled_logits_count_ith(struct llama_context * ctx, int32_t i);
+
+    // Get the backend sampled candidates (token ids) for the ith token
+    // These are needed to map probability/logit indices to vocab token ids.
+    // Returns NULL if no candidates were sampled.
+    LLAMA_API llama_token * llama_get_sampled_candidates_ith      (struct llama_context * ctx, int32_t i);
+    LLAMA_API uint32_t      llama_get_sampled_candidates_count_ith(struct llama_context * ctx, int32_t i);
+
     //
     // Vocab
     //
@@ -1154,11 +1203,16 @@ extern "C" {
     //
     //    llama_sampler_free(smpl);
     //
-    // TODO: In the future, llama_sampler will be utilized to offload the sampling to the backends (e.g. GPU).
-    //
 
     typedef void * llama_sampler_context_t;
 
+    struct llama_sampler_data {
+        struct ggml_tensor * logits;
+        struct ggml_tensor * probs;
+        struct ggml_tensor * sampled;
+        struct ggml_tensor * candidates;
+    };
+
     // user code can implement the interface below in order to create custom llama_sampler
     struct llama_sampler_i {
         const char *           (*name)  (const struct llama_sampler * smpl);                                 // can be NULL
@@ -1168,17 +1222,44 @@ extern "C" {
         struct llama_sampler * (*clone) (const struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
         void                   (*free)  (      struct llama_sampler * smpl);                                 // can be NULL if ctx is NULL
 
-        // TODO: API for internal libllama usage for appending the sampling to an existing ggml_cgraph
-        //void (*apply_ggml) (struct llama_sampler * smpl, ...);
+        // [EXPERIMENTAL]
+        // backend sampling interface:
+
+        // return true if the backend supports all ops needed by the sampler
+        // note: call once per sampler
+        bool (*backend_init)(struct llama_sampler * smpl, ggml_backend_buffer_type_t buft);
+
+        // call after .backend_apply()
+        void (*backend_accept)(
+                struct llama_sampler * smpl,
+                struct ggml_context  * ctx,
+                struct ggml_cgraph   * gf,
+                struct ggml_tensor   * selected_token);
+
+        // call after .backend_init()
+        void (*backend_apply)(
+                struct llama_sampler      * smpl,
+                struct ggml_context       * ctx,
+                struct ggml_cgraph        * gf,
+                struct llama_sampler_data * data);
+
+        // called before graph execution to set inputs for the current ubatch
+        void (*backend_set_input)(struct llama_sampler * smpl);
     };
 
     struct llama_sampler {
-        const struct llama_sampler_i * iface;
-        llama_sampler_context_t        ctx;
+        struct llama_sampler_i * iface;
+
+        llama_sampler_context_t ctx;
     };
 
+    // [EXPERIMENTAL]
+    // attach a sampler to the context
+    // note: prefer initializing the context with llama_context_params.samplers when possible
+    LLAMA_API bool llama_set_sampler(struct llama_context * ctx, llama_seq_id seq_id, struct llama_sampler * smpl);
+
     // mirror of llama_sampler_i:
-    LLAMA_API struct llama_sampler * llama_sampler_init  (const struct llama_sampler_i * iface, llama_sampler_context_t ctx);
+    LLAMA_API struct llama_sampler * llama_sampler_init  (      struct llama_sampler_i * iface, llama_sampler_context_t ctx);
     LLAMA_API const char *           llama_sampler_name  (const struct llama_sampler * smpl);
     LLAMA_API void                   llama_sampler_accept(      struct llama_sampler * smpl, llama_token token);
     LLAMA_API void                   llama_sampler_apply (      struct llama_sampler * smpl, llama_token_data_array * cur_p);
@@ -1194,7 +1275,15 @@ extern "C" {
 
     // important: takes ownership of the sampler object and will free it when llama_sampler_free is called
     LLAMA_API void                   llama_sampler_chain_add(      struct llama_sampler * chain, struct llama_sampler * smpl);
-    LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i);
+
+    // return NULL if:
+    //   - the sampler is NULL
+    //   - the sampler is not a llama_sampler_chain
+    //   - the index is out of bounds, unless i == -1
+    //   - if i == -1, returns the chain itself (can be used to check if the sampler is a chain)
+    LLAMA_API struct llama_sampler * llama_sampler_chain_get(      struct llama_sampler * chain, int32_t i);
+
+    // the total number of samplers in the chain
     LLAMA_API int                    llama_sampler_chain_n  (const struct llama_sampler * chain);
 
     // after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed
@@ -1203,7 +1292,9 @@ extern "C" {
     // available samplers:
 
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
+
+    /// seed == LLAMA_DEFAULT_SEED to use a random seed.
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist(uint32_t seed);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     /// Setting k <= 0 makes this a noop
@@ -1304,6 +1395,33 @@ extern "C" {
                           const char ** seq_breakers,
                               size_t    num_breakers);
 
+    /// adaptive-p: select tokens near a configurable target probability over time.
+    ///
+    /// the adaptive-p sampler transforms the token probability distribution to favor tokens
+    /// that fall near a user-configurable probability target.
+    ///
+    /// internally, the sampler maintains an exponential moving average of the *ORIGINAL*
+    /// probabilities of selected tokens at each sampling step. it uses this EMA to compute an
+    /// adapted target probability at each sampling step, thus maintaining the desired target
+    /// probability over time.
+    ///
+    /// adaptive-p selects a token ID rather than just mutating candidates, so it must be last
+    /// in the sampler chain (like mirostat, dist, greedy).
+    ///
+    /// only mild truncation before this sampler is recommended. we suggest applying min-p
+    /// before adaptive-p as the only other active sampler in the chain.
+    ///
+    /// @param target select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
+    /// @param decay  EMA decay for adaptation; history ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99)
+    /// @param seed   RNG seed
+    ///
+    /// ref: https://github.com/ggml-org/llama.cpp/pull/17927
+    ///
+    LLAMA_API struct llama_sampler * llama_sampler_init_adaptive_p(
+                               float   target,
+                               float   decay,
+                            uint32_t   seed);
+
     LLAMA_API struct llama_sampler * llama_sampler_init_logit_bias(
                              int32_t   n_vocab,
                              int32_t   n_logit_bias,

llama/llama.cpp/src/llama-adapter.cpp

@@ -411,6 +411,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
         }
     }
 
+    // register adapter with model
+    model.loras.insert(&adapter);
+
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
@@ -468,8 +471,8 @@ int32_t llama_adapter_meta_val_str_by_index(const llama_adapter_lora * adapter,
     return snprintf(buf, buf_size, "%s", it->second.c_str());
 }
 
-void llama_adapter_lora_free(llama_adapter_lora * adapter) {
-    delete adapter;
+void llama_adapter_lora_free(llama_adapter_lora *) {
+    // deprecated: adapters are freed by llama_model's destructor
 }
 
 uint64_t llama_adapter_get_alora_n_invocation_tokens(const struct llama_adapter_lora * adapter) {

llama/llama.cpp/src/llama-adapter.h

@@ -77,6 +77,10 @@ struct llama_adapter_lora {
     ~llama_adapter_lora() = default;
 
     llama_adapter_lora_weight * get_weight(ggml_tensor * w);
+
+    uint32_t get_n_nodes() const {
+        return ab_map.size() * 6u; // a, b, scale, add, 2 x mul_mat
+    }
 };
 
 using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

llama/llama.cpp/src/llama-arch.cpp

@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_STARCODER,        "starcoder"        },
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
+    { LLM_ARCH_MODERN_BERT,      "modern-bert"      },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_NEO_BERT,         "neo-bert"         },
@@ -41,6 +42,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHIMOE,           "phimoe"           },
     { LLM_ARCH_PLAMO,            "plamo"            },
     { LLM_ARCH_PLAMO2,           "plamo2"           },
+    { LLM_ARCH_PLAMO3,           "plamo3"           },
     { LLM_ARCH_CODESHELL,        "codeshell"        },
     { LLM_ARCH_ORION,            "orion"            },
     { LLM_ARCH_INTERNLM2,        "internlm2"        },
@@ -79,6 +81,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
+    { LLM_ARCH_EXAONE_MOE,       "exaone-moe"       },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
     { LLM_ARCH_RWKV7,            "rwkv7"            },
@@ -115,6 +118,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_MIMO2,            "mimo2"           },
+    { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
+    { LLM_ARCH_MAINCODER,        "maincoder"        },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -148,6 +154,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_VOCAB_SIZE,                        "%s.vocab_size"                        },
     { LLM_KV_CONTEXT_LENGTH,                    "%s.context_length"                    },
     { LLM_KV_EMBEDDING_LENGTH,                  "%s.embedding_length"                  },
+    { LLM_KV_EMBEDDING_LENGTH_OUT,              "%s.embedding_length_out"              },
     { LLM_KV_FEATURES_LENGTH,                   "%s.features_length"                   },
     { LLM_KV_BLOCK_COUNT,                       "%s.block_count"                       },
     { LLM_KV_LEADING_DENSE_BLOCK_COUNT,         "%s.leading_dense_block_count"         },
@@ -205,6 +212,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_GATE_LORA_RANK,               "%s.attention.gate_lora_rank"               },
     { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
     { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
+    { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,       "%s.attention.sliding_window_pattern"       },
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
@@ -216,6 +224,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ROPE_DIMENSION_COUNT,          "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,       "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_FREQ_BASE_SWA,            "%s.rope.freq_base_swa"                   },
     { LLM_KV_ROPE_SCALE_LINEAR,             "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,             "%s.rope.scaling.type"                    },
     { LLM_KV_ROPE_SCALING_FACTOR,           "%s.rope.scaling.factor"                  },
@@ -500,6 +509,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_DECI:
         case LLM_ARCH_MISTRAL3:
+        case LLM_ARCH_LLAMA_EMBED:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
                 LLM_TENSOR_OUTPUT_NORM,
@@ -781,6 +791,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_CLS,
                 LLM_TENSOR_CLS_OUT,
             };
+        case LLM_ARCH_MODERN_BERT:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_TOKEN_EMBD_NORM,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_CLS,
+                LLM_TENSOR_CLS_OUT,
+            };
         case LLM_ARCH_JINA_BERT_V2:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -930,6 +954,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_ATTN_K_NORM,
                 LLM_TENSOR_ATTN_V,
                 LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_ATTN_GATE,
                 LLM_TENSOR_FFN_NORM,
                 LLM_TENSOR_FFN_GATE_INP,
                 LLM_TENSOR_FFN_GATE_EXPS,
@@ -1060,6 +1086,22 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_ATTN_POST_NORM,
                 LLM_TENSOR_FFN_POST_NORM,
             };
+        case LLM_ARCH_PLAMO3:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_POST_NORM,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_POST_NORM,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+            };
         case LLM_ARCH_CODESHELL:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -1690,6 +1732,38 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_UP,
                 LLM_TENSOR_FFN_POST_NORM,
             };
+        case LLM_ARCH_EXAONE_MOE:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ROPE_FREQS,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_GATE_INP,
+                LLM_TENSOR_FFN_GATE_EXPS,
+                LLM_TENSOR_FFN_DOWN_EXPS,
+                LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_SHEXP,
+                LLM_TENSOR_FFN_UP_SHEXP,
+                LLM_TENSOR_FFN_DOWN_SHEXP,
+                LLM_TENSOR_FFN_EXP_PROBS_B,
+                LLM_TENSOR_NEXTN_EH_PROJ,
+                LLM_TENSOR_NEXTN_EMBED_TOKENS,
+                LLM_TENSOR_NEXTN_ENORM,
+                LLM_TENSOR_NEXTN_HNORM,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
+            };
         case LLM_ARCH_RWKV6:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -2040,6 +2114,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_TOKEN_EMBD,
                 LLM_TENSOR_OUTPUT_NORM_LFM2,
                 LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_DENSE_2_OUT,
             };
         case LLM_ARCH_LFM2MOE:
             return {
@@ -2058,7 +2133,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_SHORTCONV_INPROJ,
                 LLM_TENSOR_SHORTCONV_OUTPROJ,
                 LLM_TENSOR_TOKEN_EMBD,
-                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT_NORM_LFM2,
                 LLM_TENSOR_FFN_GATE_INP,
                 LLM_TENSOR_FFN_GATE_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
@@ -2174,11 +2249,49 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_VISEXP_FFN_DOWN,
                 LLM_TENSOR_VISEXP_FFN_UP,
             };
+        case LLM_ARCH_MIMO2:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_SINKS,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_GATE_INP,
+                LLM_TENSOR_FFN_GATE_EXPS,
+                LLM_TENSOR_FFN_DOWN_EXPS,
+                LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_EXP_PROBS_B,
+            };
         case LLM_ARCH_GPTJ:
         case LLM_ARCH_UNKNOWN:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
             };
+        case LLM_ARCH_MAINCODER:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+            };
         case LLM_ARCH_SOLAR:
             return {
                 LLM_TENSOR_TOKEN_EMBD,

llama/llama.cpp/src/llama-arch.h

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_STARCODER,
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
+    LLM_ARCH_MODERN_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_NEO_BERT,
@@ -45,6 +46,7 @@ enum llm_arch {
     LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
     LLM_ARCH_PLAMO2,
+    LLM_ARCH_PLAMO3,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
@@ -83,6 +85,7 @@ enum llm_arch {
     LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
+    LLM_ARCH_EXAONE_MOE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,
@@ -119,6 +122,9 @@ enum llm_arch {
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_MISTRAL3,
+    LLM_ARCH_MIMO2,
+    LLM_ARCH_LLAMA_EMBED,
+    LLM_ARCH_MAINCODER,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -152,6 +158,7 @@ enum llm_kv {
     LLM_KV_VOCAB_SIZE,
     LLM_KV_CONTEXT_LENGTH,
     LLM_KV_EMBEDDING_LENGTH,
+    LLM_KV_EMBEDDING_LENGTH_OUT,
     LLM_KV_FEATURES_LENGTH,
     LLM_KV_BLOCK_COUNT,
     LLM_KV_LEADING_DENSE_BLOCK_COUNT,
@@ -209,6 +216,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
+    LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
@@ -220,6 +228,7 @@ enum llm_kv {
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_FREQ_BASE_SWA,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,
     LLM_KV_ROPE_SCALING_FACTOR,

llama/llama.cpp/src/llama-chat.cpp

@@ -57,6 +57,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "exaone4",           LLM_CHAT_TEMPLATE_EXAONE_4          },
+    { "exaone-moe",        LLM_CHAT_TEMPLATE_EXAONE_MOE        },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
@@ -74,6 +75,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
     { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
+    { "solar-open",        LLM_CHAT_TEMPLATE_SOLAR_OPEN        },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -136,6 +138,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains("[gMASK]<sop>")) {
         return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        if (tmpl_contains("<|tool_declare|>")) {
+            return LLM_CHAT_TEMPLATE_EXAONE_MOE;
+        }
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
         return LLM_CHAT_TEMPLATE_GLMEDGE;
@@ -216,6 +221,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GROK_2;
     } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
         return LLM_CHAT_TEMPLATE_PANGU_EMBED;
+    } else if (tmpl_contains("<|begin|>") && tmpl_contains("<|end|>") && tmpl_contains("<|content|>")) {
+        return LLM_CHAT_TEMPLATE_SOLAR_OPEN;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -573,6 +580,22 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_MOE) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "user") {
+                ss << "<|user|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "assistant") {
+                ss << "<|assistant|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "tool") {
+                ss << "<|tool|>\n" << trim(message->content) << "<|endofturn|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
         for (size_t i = 0; i < chat.size(); i++) {
@@ -845,6 +868,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[unused9]助手：";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SOLAR_OPEN) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|begin|>" << role << "<|content|>" << message->content << "<|end|>";
+        }
+        if (add_ass) {
+            ss << "<|begin|>assistant";
+        }
     } else {
         // template not supported
         return -1;

llama/llama.cpp/src/llama-chat.h

@@ -36,6 +36,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_EXAONE_4,
+    LLM_CHAT_TEMPLATE_EXAONE_MOE,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
@@ -54,6 +55,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
     LLM_CHAT_TEMPLATE_PANGU_EMBED,
+    LLM_CHAT_TEMPLATE_SOLAR_OPEN,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

llama/llama.cpp/src/llama-context.h

@@ -40,6 +40,14 @@ struct llama_context {
 
     ~llama_context();
 
+    // reserve a new backend scheduler (if needed)
+    // for example, when:
+    //   - changing loras
+    //   - changing samplers
+    //   - changing attention type
+    //   - etc.
+    void sched_reserve();
+
     void synchronize();
 
     const llama_model   & get_model()   const;
@@ -70,6 +78,18 @@ struct llama_context {
     float * get_embeddings_ith(int32_t i);
     float * get_embeddings_seq(llama_seq_id seq_id);
 
+    llama_token * get_sampled_tokens() const;
+    llama_token   get_sampled_token_ith(int32_t idx);
+
+    float * get_sampled_logits_ith(int32_t idx);
+    size_t  get_sampled_logits_count(int32_t idx);
+
+    float * get_sampled_probs_ith(int32_t idx);
+    size_t  get_sampled_probs_count(int32_t idx);
+
+    const llama_token * get_sampled_candidates_ith(int32_t idx);
+    size_t get_sampled_candidates_count(int32_t idx);
+
     void attach_threadpool(
             ggml_threadpool_t threadpool,
             ggml_threadpool_t threadpool_batch);
@@ -192,10 +212,13 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    uint32_t output_reserve(int32_t n_outputs);
+    uint32_t output_reserve(int32_t n_outputs, const llama_batch & batch);
 
     void output_reorder();
 
+    // map the output row index `i` to batch index
+    int64_t output_resolve_row(int32_t i) const;
+
     //
     // graph
     //
@@ -213,6 +236,8 @@ public:
     ggml_cgraph * graph_reserve(
         uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false, size_t * sizes = nullptr);
 
+    bool set_sampler(llama_seq_id seq_id, llama_sampler * sampler);
+
 private:
     llm_graph_params graph_params(
                         llm_graph_result * res,
@@ -252,6 +277,31 @@ private:
     size_t  embd_size = 0; // capacity (of floats) for embeddings
     float * embd      = nullptr;
 
+    // TODO: simplify
+    struct sampling_info {
+        std::map<llama_seq_id, llama_sampler *> samplers;
+
+        float       * logits      = nullptr;
+        size_t        logits_size = 0;
+
+        llama_token * sampled      = nullptr;
+        size_t        sampled_size = 0;
+
+        float       * probs        = nullptr;
+        size_t        probs_size   = 0;
+
+        llama_token * candidates   = nullptr;
+        size_t        candidates_size = 0;
+
+        std::vector<uint32_t> logits_count;
+        std::vector<uint32_t> probs_count;
+        std::vector<uint32_t> candidates_count;
+
+        std::vector<llama_token> token_ids_full_vocab;
+    };
+
+    sampling_info sampling;
+
     // sequence embeddings output (map of [n_embd] vectors)
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
@@ -272,6 +322,8 @@ private:
 
     ggml_backend_sched_ptr sched;
 
+    bool sched_need_reserve = true;
+
     ggml_backend_t backend_cpu = nullptr;
     std::vector<ggml_backend_ptr> backends;
 

llama/llama.cpp/src/llama-cparams.h

@@ -30,10 +30,12 @@ struct llama_cparams {
     bool causal_attn;
     bool offload_kqv;
     bool flash_attn;
+    bool auto_fa;
     bool no_perf;
     bool warmup;
     bool op_offload;
     bool kv_unified;
+    bool pipeline_parallel;
 
     enum llama_pooling_type pooling_type;
 

llama/llama.cpp/src/llama-grammar.cpp

@@ -369,6 +369,44 @@ static void print_rule(
     fprintf(file, "\n");
 }
 
+//
+// Regex utilities
+//
+
+size_t llama_grammar_trigger_pattern::find(const std::string & input) const {
+    auto find_start_pos = [](const std::smatch & match) {
+        // get from the first matched capturing group to the end of the string
+        size_t start = std::string::npos;
+        for (auto i = 1u; i < match.size(); i++) {
+            if (match.length(i) > 0) {
+                start = match.position(i);
+                break;
+            }
+        }
+        if (start == std::string::npos) {
+            start = match.position(0);
+        }
+        return start;
+    };
+
+    if (!pattern.empty() && pattern.front() == '^' && pattern.back() == '$') {
+        // match against the entire input
+        std::smatch match;
+        if (std::regex_match(input, match, regex)) {
+            return find_start_pos(match);
+        }
+    }
+
+    // search anywhere
+    std::smatch match;
+    if (std::regex_search(input, match, regex)) {
+        return find_start_pos(match);
+    }
+
+    return std::string::npos;
+}
+
+
 //
 // implementation
 //
@@ -1321,21 +1359,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
             grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
             grammar.trigger_buffer += piece;
 
-            std::smatch match;
             for (const auto & trigger_pattern : grammar.trigger_patterns) {
-                if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) {
+                auto start = trigger_pattern.find(grammar.trigger_buffer);
+                if (start != std::string::npos) {
                     grammar.awaiting_trigger = false;
-                    // get from the first matched capturing group to the end of the string
-                    size_t start = std::string::npos;
-                    for (auto i = 1u; i < match.size(); i++) {
-                        if (match.length(i) > 0) {
-                            start = match.position(i);
-                            break;
-                        }
-                    }
-                    if (start == std::string::npos) {
-                        start = match.position(0);
-                    }
 
                     // replay tokens that overlap with [start, end)
                     for (const auto & [tok, tok_pos] : grammar.trigger_buffer_positions) {