wip?

Co-authored-by: Patrick Devine <patrick@infrahq.com>

.github/workflows/release.yaml

@@ -111,13 +111,13 @@ jobs:
           - os: windows
             arch: amd64
             preset: 'CUDA 12'
-            install: https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_551.61_windows.exe
-            cuda-version: '12.4'
+            install: https://developer.download.nvidia.com/compute/cuda/12.8.0/local_installers/cuda_12.8.0_571.96_windows.exe
+            cuda-version: '12.8'
           - os: windows
             arch: amd64
             preset: 'ROCm 6'
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
-            rocm-version: '6.1'
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
+            rocm-version: '6.2'
     runs-on: ${{ matrix.arch == 'arm64' && format('{0}-{1}', matrix.os, matrix.arch) || matrix.os }}
     environment: release
     env:
@@ -160,6 +160,10 @@ jobs:
           echo "$hipPath\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
           echo "CC=$hipPath\bin\clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
           echo "CXX=$hipPath\bin\clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+      - if: matrix.preset == 'CPU'
+        run: |
+          echo "CC=clang.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "CXX=clang++.exe" | Out-File -FilePath $env:GITHUB_ENV -Append
       - if: ${{ !cancelled() && steps.cache-install.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@v4
         with:

.github/workflows/test.yaml

@@ -78,10 +78,10 @@ jobs:
         include:
           - preset: CPU
           - preset: CUDA
-            install: https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_522.06_windows.exe
-            flags: '-DCMAKE_CUDA_ARCHITECTURES=87'
+            install: https://developer.download.nvidia.com/compute/cuda/11.3.1/local_installers/cuda_11.3.1_465.89_win10.exe
+            flags: '-DCMAKE_CUDA_ARCHITECTURES=80'
           - preset: ROCm
-            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe
+            install: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q4-WinSvr2022-For-HIP.exe
             flags: '-DAMDGPU_TARGETS=gfx1010'
     runs-on: windows
     steps:
@@ -102,7 +102,7 @@ jobs:
           $ErrorActionPreference = "Stop"
           if ("${{ steps.cache-install.outputs.cache-hit }}" -ne 'true') {
             Invoke-WebRequest -Uri "${{ matrix.install }}" -OutFile "install.exe"
-            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.8", "nvcc_11.8", "cublas_11.8", "cublas_dev_11.8")) -NoNewWindow -Wait
+            Start-Process -FilePath .\install.exe -ArgumentList (@("-s", "cudart_11.3", "nvcc_11.3", "cublas_11.3", "cublas_dev_11.3")) -NoNewWindow -Wait
           }
 
           $cudaPath = (Resolve-Path "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\*").path
@@ -140,6 +140,13 @@ jobs:
         env:
           CMAKE_GENERATOR: Ninja
 
+  go_mod_tidy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: check that 'go mod tidy' is clean
+        run: go mod tidy --diff || (echo "Please run 'go mod tidy'." && exit 1)
+
   test:
     strategy:
       matrix:
@@ -147,15 +154,82 @@ jobs:
     runs-on: ${{ matrix.os }}
     env:
       CGO_ENABLED: '1'
+      GOEXPERIMENT: 'synctest'
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-go@v5
+      - name: checkout
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # 4.2.2
+
+      - name: cache restore
+        uses: actions/cache/restore@1bd1e32a3bdc45362d1e726936510720a7c30a57 # v4.2.0
         with:
+          # Note: unlike the other setups, this is only grabbing the mod download
+          # cache, rather than the whole mod directory, as the download cache
+          # contains zips that can be unpacked in parallel faster than they can be
+          # fetched and extracted by tar
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod/cache
+            ~\AppData\Local\go-build
+          # NOTE: The -3- here should be incremented when the scheme of data to be
+          # cached changes (e.g. path above changes).
+          key: ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}-${{ github.run_id }}
+          restore-keys: |
+            ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}
+            ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-
+
+      - name: Setup Go
+        uses: actions/setup-go@v5
+        with:
+          # The caching strategy of setup-go is less than ideal, and wastes
+          # time by not saving artifacts due to small failures like the linter
+          # complaining, etc. This means subsequent have to rebuild their world
+          # again until all checks pass. For instance, if you mispell a word,
+          # you're punished until you fix it. This is more hostile than
+          # helpful.
+          cache: false
+
           go-version-file: go.mod
+
+      # It is tempting to run this in a platform independent way, but the past
+      # shows this codebase will see introductions of platform specific code
+      # generation, and so we need to check this per platform to ensure we
+      # don't abuse go generate on specific platforms.
+      - name: check that 'go generate' is clean
+        if: always()
+        run: |
+          go generate ./...
+          git diff --name-only --exit-code || (echo "Please run 'go generate ./...'." && exit 1)
+
+      - name: go test
+        if: always()
+        run: go test -count=1 -benchtime=1x ./...
+
+      # TODO(bmizerany): replace this heavy tool with just the
+      # tools/checks/binaries we want and then make them all run in parallel
+      # across jobs, not on a single tiny vm on Github Actions.
       - uses: golangci/golangci-lint-action@v6
         with:
           args: --timeout 10m0s -v
-      - run: go test ./...
+
+      - name: cache save
+        # Always save the cache, even if the job fails. The artifacts produced
+        # during the building of test binaries are not all for naught. They can
+        # be used to speed up subsequent runs.
+        if: always()
+
+        uses: actions/cache/save@1bd1e32a3bdc45362d1e726936510720a7c30a57 # v4.2.0
+        with:
+          # Note: unlike the other setups, this is only grabbing the mod download
+          # cache, rather than the whole mod directory, as the download cache
+          # contains zips that can be unpacked in parallel faster than they can be
+          # fetched and extracted by tar
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod/cache
+            ~\AppData\Local\go-build
+          # NOTE: The -3- here should be incremented when the scheme of data to be
+          # cached changes (e.g. path above changes).
+          key: ${{ github.job }}-${{ runner.os }}-${{ matrix.goarch }}-${{ matrix.buildflags }}-go-3-${{ hashFiles('**/go.sum') }}-${{ github.run_id }}
 
   patches:
     runs-on: ubuntu-latest
@@ -163,5 +237,5 @@ jobs:
       - uses: actions/checkout@v4
       - name: Verify patches apply cleanly and do not change files
         run: |
-          make -f Makefile.sync clean sync
-          git diff --compact-summary --exit-code
+          make -f Makefile.sync clean checkout apply-patches sync
+          git diff --compact-summary --exit-code

.gitignore

@@ -5,7 +5,6 @@
 .swp
 dist
 build
-ollama
 .cache
 *.exe
 .idea
@@ -14,3 +13,4 @@ test_data
 __debug_bin*
 llama/build
 llama/vendor
+/ollama

.golangci.yaml

@@ -6,8 +6,6 @@ linters:
     - bidichk
     - bodyclose
     - containedctx
-    - contextcheck
-    - errcheck
     - gocheckcompilerdirectives
     - gofmt
     - gofumpt
@@ -23,10 +21,11 @@ linters:
     - staticcheck
     - tenv
     - unconvert
-    - unused
-    - usestdlibvars
     - wastedassign
     - whitespace
+  disable:
+    - usestdlibvars
+    - errcheck
 linters-settings:
   staticcheck:
     checks:
@@ -39,5 +38,4 @@ severity:
         - gofmt
         - goimports
         - intrange
-        - usestdlibvars
       severity: info

CMakeLists.txt

@@ -23,6 +23,8 @@ set(GGML_SCHED_MAX_COPIES 4)
 set(GGML_LLAMAFILE ON)
 set(GGML_CUDA_PEER_MAX_BATCH_SIZE 128)
 set(GGML_CUDA_GRAPHS ON)
+set(GGML_CUDA_FA ON)
+set(GGML_CUDA_COMPRESSION_MODE default)
 
 if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
     OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
@@ -85,9 +87,9 @@ if(CMAKE_CUDA_COMPILER)
     )
 endif()
 
-set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a):xnack[+-]$"
+set(WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX "^gfx(906|908|90a|1200|1201):xnack[+-]$"
     CACHE STRING
-    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a):xnack[+-]$\"."
+    "Regular expression describing AMDGPU_TARGETS not supported on Windows. Override to force building these targets. Default \"^gfx(906|908|90a|1200|1201):xnack[+-]$\"."
 )
 
 check_language(HIP)
@@ -96,7 +98,7 @@ if(CMAKE_HIP_COMPILER)
 
     find_package(hip REQUIRED)
     if(NOT AMDGPU_TARGETS)
-        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012])$")
+        list(FILTER AMDGPU_TARGETS INCLUDE REGEX "^gfx(900|94[012]|101[02]|1030|110[012]|120[01])$")
     elseif(WIN32 AND WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX)
         list(FILTER AMDGPU_TARGETS EXCLUDE REGEX ${WINDOWS_AMDGPU_TARGETS_EXCLUDE_REGEX})
     endif()
@@ -105,9 +107,11 @@ if(CMAKE_HIP_COMPILER)
         add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/ml/backend/ggml/ggml/src/ggml-hip)
 
         if (WIN32)
-            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY=1)
+            target_compile_definitions(ggml-hip PRIVATE GGML_CUDA_NO_PEER_COPY)
         endif()
 
+        target_compile_definitions(ggml-hip PRIVATE GGML_HIP_NO_VMM)
+
         set(OLLAMA_HIP_INSTALL_DIR ${OLLAMA_INSTALL_DIR}/rocm)
         install(TARGETS ggml-hip
             RUNTIME_DEPENDENCIES

CMakePresets.json

@@ -21,14 +21,16 @@
       "name": "CUDA 11",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;62;70;72;75;80;86"
+        "CMAKE_CUDA_ARCHITECTURES": "50;52;53;60;61;70;75;80;86",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
       }
     },
     {
       "name": "CUDA 12",
       "inherits": [ "CUDA" ],
       "cacheVariables": {
-        "CMAKE_CUDA_ARCHITECTURES": "60;61;62;70;72;75;80;86;87;89;90;90a"
+        "CMAKE_CUDA_ARCHITECTURES": "50;60;61;70;75;80;86;87;89;90;90a;120",
+        "CMAKE_CUDA_FLAGS": "-Wno-deprecated-gpu-targets"
       }
     },
     {
@@ -56,7 +58,7 @@
       "name": "ROCm 6",
       "inherits": [ "ROCm" ],
       "cacheVariables": {
-        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
+        "AMDGPU_TARGETS": "gfx900;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102;gfx1151;gfx1200;gfx1201;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-"
       }
     }
   ],

CONTRIBUTING.md

@@ -6,8 +6,6 @@ Thank you for your interest in contributing to Ollama! Here are a few guidelines
 
 See the [development documentation](./docs/development.md) for instructions on how to build and run Ollama locally.
 
-## Pull requests
-
 ### Ideal issues
 
 * [Bugs](https://github.com/ollama/ollama/issues?q=is%3Aissue+is%3Aopen+label%3Abug): issues where Ollama stops working or where it results in an unexpected error.
@@ -26,11 +24,64 @@ See the [development documentation](./docs/development.md) for instructions on h
 * Changes that add significant friction to the user experience
 * Changes that create a large future maintenance burden for maintainers and contributors
 
-### Best practices
+## Proposing a (non-trivial) change
 
-* Commit messages: please leave both a title and a description in your commit messages. The title should be a short summary of the changes, with a leading word that explains the section of the code being changed (e.g. `api: fix parsing of prompt field`) . In the description, leave a short 2-3 sentences that explain more about the change and its impact.
-* Tests: please add test coverage to changes where possible.
-* Minimize dependencies: avoid adding new dependencies unless absolutely necessary.
+> By "non-trivial", we mean a change that is not a bug fix or small
+> documentation update. If you are unsure, please ask us on our [Discord
+> server](https://discord.gg/ollama).
+
+Before opening a non-trivial Pull Request, please open an issue to discuss the change and
+get feedback from the maintainers. This helps us understand the context of the
+change and how it fits into Ollama's roadmap and prevents us from duplicating
+work or you from spending time on a change that we may not be able to accept.
+
+Tips for proposals:
+
+* Explain the problem you are trying to solve, not what you are trying to do.
+* Explain why the change is important.
+* Explain how the change will be used.
+* Explain how the change will be tested.
+
+Additionally, for bonus points: Provide draft documentation you would expect to
+see if the change were accepted.
+
+## Pull requests
+
+**Commit messages**
+
+The title should look like:
+
+    <package>: <short description>
+
+The package is the most affected Go package. If the change does not affect Go
+code, then use the directory name instead. Changes to a single well-known
+file in the root directory may use the file name.
+
+The short description should start with a lowercase letter and be a
+continuation of the sentence:
+
+      "This changes Ollama to..."
+
+Examples:
+
+      llm/backend/mlx: support the llama architecture
+      CONTRIBUTING: provide clairity on good commit messages, and bad
+
+Bad Examples:
+
+      feat: add more emoji
+      fix: was not using famous web framework
+      chore: generify code
+
+**Tests**
+
+Please include tests. Strive to test behavior, not implementation.
+
+**New dependencies**
+
+Dependencies should be added sparingly. If you are adding a new dependency,
+please explain why it is necessary and what other ways you attempted that
+did not work without it.
 
 ## Need help?
 

Dockerfile

@@ -2,22 +2,24 @@
 
 ARG FLAVOR=${TARGETARCH}
 
-ARG ROCMVERSION=6.1.2
+ARG ROCMVERSION=6.3.3
 ARG JETPACK5VERSION=r35.4.1
-ARG JETPACK6VERSION=r36.2.0
+ARG JETPACK6VERSION=r36.4.0
 ARG CMAKEVERSION=3.31.2
 
-FROM --platform=linux/amd64 rocm/dev-centos-7:${ROCMVERSION}-complete AS base-amd64
-RUN sed -i -e 's/mirror.centos.org/vault.centos.org/g' -e 's/^#.*baseurl=http/baseurl=http/g' -e 's/^mirrorlist=http/#mirrorlist=http/g' /etc/yum.repos.d/*.repo \
-    && yum install -y yum-utils devtoolset-10-gcc devtoolset-10-gcc-c++ \
-    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo \
-    && curl -s -L https://github.com/ccache/ccache/releases/download/v4.10.2/ccache-4.10.2-linux-x86_64.tar.xz | tar -Jx -C /usr/local/bin --strip-components 1
-ENV PATH=/opt/rh/devtoolset-10/root/usr/bin:/opt/rh/devtoolset-11/root/usr/bin:$PATH
+# CUDA v11 requires gcc v10.  v10.3 has regressions, so the rockylinux 8.5 AppStream has the latest compatible version
+FROM --platform=linux/amd64 rocm/dev-almalinux-8:${ROCMVERSION}-complete AS base-amd64
+RUN yum install -y yum-utils \
+    && yum-config-manager --add-repo https://dl.rockylinux.org/vault/rocky/8.5/AppStream/\$basearch/os/ \
+    && rpm --import https://dl.rockylinux.org/pub/rocky/RPM-GPG-KEY-Rocky-8 \
+    && dnf install -y yum-utils ccache gcc-toolset-10-gcc-10.2.1-8.2.el8 gcc-toolset-10-gcc-c++-10.2.1-8.2.el8 gcc-toolset-10-binutils-2.35-11.el8 \
+    && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo
+ENV PATH=/opt/rh/gcc-toolset-10/root/usr/bin:$PATH
 
-FROM --platform=linux/arm64 rockylinux:8 AS base-arm64
+FROM --platform=linux/arm64 almalinux:8 AS base-arm64
 # install epel-release for ccache
 RUN yum install -y yum-utils epel-release \
-    && yum install -y clang ccache \
+    && dnf install -y clang ccache \
     && yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel8/sbsa/cuda-rhel8.repo
 ENV CC=clang CXX=clang++
 
@@ -29,9 +31,8 @@ COPY ml/backend/ggml/ggml ml/backend/ggml/ggml
 ENV LDFLAGS=-s
 
 FROM base AS cpu
-# amd64 uses gcc which requires devtoolset-11 for AVX extensions while arm64 uses clang
-RUN if [ "$(uname -m)" = "x86_64" ]; then yum install -y devtoolset-11-gcc devtoolset-11-gcc-c++; fi
-ENV PATH=/opt/rh/devtoolset-11/root/usr/bin:$PATH
+RUN dnf install -y gcc-toolset-11-gcc gcc-toolset-11-gcc-c++
+ENV PATH=/opt/rh/gcc-toolset-11/root/usr/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CPU' \
         && cmake --build --parallel --preset 'CPU' \
@@ -39,7 +40,7 @@ RUN --mount=type=cache,target=/root/.ccache \
 
 FROM base AS cuda-11
 ARG CUDA11VERSION=11.3
-RUN yum install -y cuda-toolkit-${CUDA11VERSION//./-}
+RUN dnf install -y cuda-toolkit-${CUDA11VERSION//./-}
 ENV PATH=/usr/local/cuda-11/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CUDA 11' \
@@ -47,8 +48,8 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --install build --component CUDA --strip --parallel 8
 
 FROM base AS cuda-12
-ARG CUDA12VERSION=12.4
-RUN yum install -y cuda-toolkit-${CUDA12VERSION//./-}
+ARG CUDA12VERSION=12.8
+RUN dnf install -y cuda-toolkit-${CUDA12VERSION//./-}
 ENV PATH=/usr/local/cuda-12/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'CUDA 12' \
@@ -56,6 +57,7 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --install build --component CUDA --strip --parallel 8
 
 FROM base AS rocm-6
+ENV PATH=/opt/rocm/hcc/bin:/opt/rocm/hip/bin:/opt/rocm/bin:/opt/rocm/hcc/bin:$PATH
 RUN --mount=type=cache,target=/root/.ccache \
     cmake --preset 'ROCm 6' \
         && cmake --build --parallel --preset 'ROCm 6' \
@@ -84,10 +86,11 @@ RUN --mount=type=cache,target=/root/.ccache \
         && cmake --install build --component CUDA --strip --parallel 8
 
 FROM base AS build
-ARG GOVERSION=1.23.4
-RUN curl -fsSL https://golang.org/dl/go${GOVERSION}.linux-$(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
 WORKDIR /go/src/github.com/ollama/ollama
+COPY go.mod go.sum .
+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
 COPY . .
 ARG GOFLAGS="'-ldflags=-w -s'"
 ENV CGO_ENABLED=1
@@ -101,10 +104,10 @@ COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 FROM --platform=linux/arm64 scratch AS arm64
 COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
 COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
-COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 lib/ollama/cuda_jetpack6
+COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
+COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6
 
-FROM --platform=linux/arm64 scratch AS rocm
+FROM scratch AS rocm
 COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm
 
 FROM ${FLAVOR} AS archive

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggerganov/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=46e3556e01b824e52395fb050b29804b6cff2a7c
+FETCH_HEAD=2016f07bd106c73699ecbaace80f55db5ed95dac
 
 .PHONY: help
 help:
@@ -15,18 +15,18 @@ help:
 	@echo "    make -f $(lastword $(MAKEFILE_LIST)) clean sync"
 
 .PHONY: sync
-sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml apply-patches
+sync: llama/build-info.cpp llama/llama.cpp ml/backend/ggml/ggml
 
 .PHONY: llama/build-info.cpp
 llama/build-info.cpp: llama/build-info.cpp.in
 	sed -e 's|@FETCH_HEAD@|$(FETCH_HEAD)|' $< > $@
 
 .PHONY: llama/llama.cpp
-llama/llama.cpp: llama/vendor/ apply-patches
+llama/llama.cpp: llama/vendor/
 	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
 
-.PHONY: ml/backend/ggml/ggml apply-patches
-ml/backend/ggml/ggml: llama/vendor/ggml/ apply-patches
+.PHONY: ml/backend/ggml/ggml
+ml/backend/ggml/ggml: llama/vendor/ggml/
 	rsync -arvzc -f "merge $@/.rsync-filter" $< $@
 
 PATCHES=$(wildcard llama/patches/*.patch)

README.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <a href="https://ollama.com" />
+  <a href="https://ollama.com">
     <img alt="ollama" height="200px" src="https://github.com/ollama/ollama/assets/3325447/0d0b44e2-8f4a-4e99-9b52-a5c1c741c8f7">
   </a>
 </div>
@@ -54,6 +54,11 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
+| Gemma 3            | 1B         | 815MB | `ollama run gemma3:1b`           |
+| Gemma 3            | 4B         | 3.3GB | `ollama run gemma3`              |
+| Gemma 3            | 12B        | 8.1GB | `ollama run gemma3:12b`          |
+| Gemma 3            | 27B        | 17GB  | `ollama run gemma3:27b`          |
+| QwQ                | 32B        | 20GB  | `ollama run qwq`                 |
 | DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
 | DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
 | Llama 3.3          | 70B        | 43GB  | `ollama run llama3.3`            |
@@ -64,10 +69,7 @@ Here are some example models that can be downloaded:
 | Llama 3.1          | 8B         | 4.7GB | `ollama run llama3.1`            |
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
 | Phi 4              | 14B        | 9.1GB | `ollama run phi4`                |
-| Phi 3 Mini         | 3.8B       | 2.3GB | `ollama run phi3`                |
-| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`           |
-| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`              |
-| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`          |
+| Phi 4 Mini         | 3.8B       | 2.5GB | `ollama run phi4-mini`           |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`             |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`           |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`         |
@@ -75,7 +77,7 @@ 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`               |
-| Solar              | 10.7B      | 6.1GB | `ollama run solar`               |
+| Granite-3.2         | 8B         | 4.9GB | `ollama run granite3.2`          |
 
 > [!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.
@@ -275,6 +277,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 ### Web & Desktop
 
 - [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)
 - [Hollama](https://github.com/fmaclen/hollama)
 - [Lollms-Webui](https://github.com/ParisNeo/lollms-webui)
@@ -282,12 +285,13 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Bionic GPT](https://github.com/bionic-gpt/bionic-gpt)
 - [HTML UI](https://github.com/rtcfirefly/ollama-ui)
 - [Saddle](https://github.com/jikkuatwork/saddle)
+- [TagSpaces](https://www.tagspaces.org) (A platform for file based apps, [utilizing Ollama](https://docs.tagspaces.org/ai/) for the generation of tags and descriptions)
 - [Chatbot UI](https://github.com/ivanfioravanti/chatbot-ollama)
 - [Chatbot UI v2](https://github.com/mckaywrigley/chatbot-ui)
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
-- [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
+- [big-AGI](https://github.com/enricoros/big-AGI)
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)
@@ -321,6 +325,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [RWKV-Runner](https://github.com/josStorer/RWKV-Runner) (RWKV offline LLM deployment tool, also usable as a client for ChatGPT and Ollama)
 - [Ollama Grid Search](https://github.com/dezoito/ollama-grid-search) (app to evaluate and compare models)
 - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama)
+- [Casibase](https://casibase.org) (An open source AI knowledge base and dialogue system combining the latest RAG, SSO, ollama support and multiple large language models.)
 - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS)
 - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama)
 - [Shinkai Desktop](https://github.com/dcSpark/shinkai-apps) (Two click install Local AI using Ollama + Files + RAG)
@@ -343,7 +348,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [PartCAD](https://github.com/openvmp/partcad/) (CAD model generation with OpenSCAD and CadQuery)
 - [Ollama4j Web UI](https://github.com/ollama4j/ollama4j-web-ui) - Java-based Web UI for Ollama built with Vaadin, Spring Boot and Ollama4j
 - [PyOllaMx](https://github.com/kspviswa/pyOllaMx) - macOS application capable of chatting with both Ollama and Apple MLX models.
-- [Claude Dev](https://github.com/saoudrizwan/claude-dev) - VSCode extension for multi-file/whole-repo coding
+- [Cline](https://github.com/cline/cline) - Formerly known as Claude Dev is a VSCode extension for multi-file/whole-repo coding
 - [Cherry Studio](https://github.com/kangfenmao/cherry-studio) (Desktop client with Ollama support)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
 - [Archyve](https://github.com/nickthecook/archyve) (RAG-enabling document library)
@@ -384,6 +389,16 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [OpenDeepResearcher-via-searxng](https://github.com/benhaotang/OpenDeepResearcher-via-searxng) (A Deep Research equivent endpoint with Ollama support for running locally)
 - [AntSK](https://github.com/AIDotNet/AntSK) (Out-of-the-box & Adaptable RAG Chatbot)
 - [MaxKB](https://github.com/1Panel-dev/MaxKB/) (Ready-to-use & flexible RAG Chatbot)
+- [yla](https://github.com/danielekp/yla) (Web interface to freely interact with your customized models)
+- [LangBot](https://github.com/RockChinQ/LangBot) (LLM-based instant messaging bots platform, with Agents, RAG features, supports multiple platforms)
+- [1Panel](https://github.com/1Panel-dev/1Panel/) (Web-based Linux Server Management Tool)
+- [AstrBot](https://github.com/Soulter/AstrBot/) (User-friendly LLM-based multi-platform chatbot with a WebUI, supporting RAG, LLM agents, and plugins integration)
+- [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
+- [Ellama](https://github.com/zeozeozeo/ellama) (Friendly native app to chat with an Ollama instance)
+- [screenpipe](https://github.com/mediar-ai/screenpipe) Build agents powered by your screen history
+- [Ollamb](https://github.com/hengkysteen/ollamb) (Simple yet rich in features, cross-platform built with Flutter and designed for Ollama. Try the [web demo](https://hengkysteen.github.io/demo/ollamb/).)
+- [Writeopia](https://github.com/Writeopia/Writeopia) (Text editor with integration with Ollama)
+- [AppFlowy](https://github.com/AppFlowy-IO/AppFlowy) (AI collaborative workspace with Ollama, cross-platform and self-hostable)
 
 ### Cloud
 
@@ -423,10 +438,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [SwollamaCLI](https://github.com/marcusziade/Swollama) bundled with the Swollama Swift package. [Demo](https://github.com/marcusziade/Swollama?tab=readme-ov-file#cli-usage)
 - [aichat](https://github.com/sigoden/aichat) All-in-one LLM CLI tool featuring Shell Assistant, Chat-REPL, RAG, AI tools & agents, with access to OpenAI, Claude, Gemini, Ollama, Groq, and more.
 - [PowershAI](https://github.com/rrg92/powershai) PowerShell module that brings AI to terminal on Windows, including support for Ollama
+- [DeepShell](https://github.com/Abyss-c0re/deepshell) Your self-hosted AI assistant. Interactive Shell, Files and Folders analysis.
 - [orbiton](https://github.com/xyproto/orbiton) Configuration-free text editor and IDE with support for tab completion with Ollama.
+- [orca-cli](https://github.com/molbal/orca-cli) Ollama Registry CLI Application - Browse, pull and download models from Ollama Registry in your terminal.
+- [GGUF-to-Ollama](https://github.com/jonathanhecl/gguf-to-ollama) - Importing GGUF to Ollama made easy (multiplatform)
 
 ### Apple Vision Pro
 
+- [SwiftChat](https://github.com/aws-samples/swift-chat) (Cross-platform AI chat app supporting Apple Vision Pro via "Designed for iPad")
 - [Enchanted](https://github.com/AugustDev/enchanted)
 
 ### Database
@@ -500,13 +519,18 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [LlmTornado](https://github.com/lofcz/llmtornado) (C# library providing a unified interface for major FOSS & Commercial inference APIs)
 - [Ollama for Zig](https://github.com/dravenk/ollama-zig)
 - [Abso](https://github.com/lunary-ai/abso) (OpenAI-compatible TypeScript SDK for any LLM provider)
+- [Nichey](https://github.com/goodreasonai/nichey) is a Python package for generating custom wikis for your research topic
+- [Ollama for D](https://github.com/kassane/ollama-d)
 
 ### Mobile
 
+- [SwiftChat](https://github.com/aws-samples/swift-chat) (Lightning-fast Cross-platform AI chat app with native UI for Android, iOS and iPad)
 - [Enchanted](https://github.com/AugustDev/enchanted)
 - [Maid](https://github.com/Mobile-Artificial-Intelligence/maid)
 - [Ollama App](https://github.com/JHubi1/ollama-app) (Modern and easy-to-use multi-platform client for Ollama)
 - [ConfiChat](https://github.com/1runeberg/confichat) (Lightweight, standalone, multi-platform, and privacy focused LLM chat interface with optional encryption)
+- [Ollama Android Chat](https://github.com/sunshine0523/OllamaServer) (No need for Termux, start the Ollama service with one click on an Android device)
+- [Reins](https://github.com/ibrahimcetin/reins) (Easily tweak parameters, customize system prompts per chat, and enhance your AI experiments with reasoning model support.)
 
 ### Extensions & Plugins
 
@@ -552,12 +576,14 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [TextLLaMA](https://github.com/adarshM84/TextLLaMA) A Chrome Extension that helps you write emails, correct grammar, and translate into any language
 - [Simple-Discord-AI](https://github.com/zyphixor/simple-discord-ai)
 - [LLM Telegram Bot](https://github.com/innightwolfsleep/llm_telegram_bot) (telegram bot, primary for RP. Oobabooga-like buttons, [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) API integration e.t.c)
+- [mcp-llm](https://github.com/sammcj/mcp-llm) (MCP Server to allow LLMs to call other LLMs)
 
 ### Supported backends
 
 - [llama.cpp](https://github.com/ggerganov/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 intergration 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.
 - [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production.

api/client.go

@@ -10,7 +10,7 @@
 // repository].
 //
 // [the API documentation]: https://github.com/ollama/ollama/blob/main/docs/api.md
-// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/examples
+// [in the GitHub repository]: https://github.com/ollama/ollama/tree/main/api/examples
 package api
 
 import (
@@ -132,7 +132,7 @@ func (c *Client) do(ctx context.Context, method, path string, reqData, respData
 const maxBufferSize = 512 * format.KiloByte
 
 func (c *Client) stream(ctx context.Context, method, path string, data any, fn func([]byte) error) error {
-	var buf *bytes.Buffer
+	var buf io.Reader
 	if data != nil {
 		bts, err := json.Marshal(data)
 		if err != nil {

api/client_test.go

@@ -1,6 +1,13 @@
 package api
 
 import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"strings"
 	"testing"
 )
 
@@ -43,3 +50,206 @@ func TestClientFromEnvironment(t *testing.T) {
 		})
 	}
 }
+
+// testError represents an internal error type with status code and message
+// this is used since the error response from the server is not a standard error struct
+type testError struct {
+	message    string
+	statusCode int
+}
+
+func (e testError) Error() string {
+	return e.message
+}
+
+func TestClientStream(t *testing.T) {
+	testCases := []struct {
+		name      string
+		responses []any
+		wantErr   string
+	}{
+		{
+			name: "immediate error response",
+			responses: []any{
+				testError{
+					message:    "test error message",
+					statusCode: http.StatusBadRequest,
+				},
+			},
+			wantErr: "test error message",
+		},
+		{
+			name: "error after successful chunks, ok response",
+			responses: []any{
+				ChatResponse{Message: Message{Content: "partial response 1"}},
+				ChatResponse{Message: Message{Content: "partial response 2"}},
+				testError{
+					message:    "mid-stream error",
+					statusCode: http.StatusOK,
+				},
+			},
+			wantErr: "mid-stream error",
+		},
+		{
+			name: "successful stream completion",
+			responses: []any{
+				ChatResponse{Message: Message{Content: "chunk 1"}},
+				ChatResponse{Message: Message{Content: "chunk 2"}},
+				ChatResponse{
+					Message:    Message{Content: "final chunk"},
+					Done:       true,
+					DoneReason: "stop",
+				},
+			},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				flusher, ok := w.(http.Flusher)
+				if !ok {
+					t.Fatal("expected http.Flusher")
+				}
+
+				w.Header().Set("Content-Type", "application/x-ndjson")
+
+				for _, resp := range tc.responses {
+					if errResp, ok := resp.(testError); ok {
+						w.WriteHeader(errResp.statusCode)
+						err := json.NewEncoder(w).Encode(map[string]string{
+							"error": errResp.message,
+						})
+						if err != nil {
+							t.Fatal("failed to encode error response:", err)
+						}
+						return
+					}
+
+					if err := json.NewEncoder(w).Encode(resp); err != nil {
+						t.Fatalf("failed to encode response: %v", err)
+					}
+					flusher.Flush()
+				}
+			}))
+			defer ts.Close()
+
+			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
+
+			var receivedChunks []ChatResponse
+			err := client.stream(context.Background(), http.MethodPost, "/v1/chat", nil, func(chunk []byte) error {
+				var resp ChatResponse
+				if err := json.Unmarshal(chunk, &resp); err != nil {
+					return fmt.Errorf("failed to unmarshal chunk: %w", err)
+				}
+				receivedChunks = append(receivedChunks, resp)
+				return nil
+			})
+
+			if tc.wantErr != "" {
+				if err == nil {
+					t.Fatal("expected error but got nil")
+				}
+				if !strings.Contains(err.Error(), tc.wantErr) {
+					t.Errorf("expected error containing %q, got %v", tc.wantErr, err)
+				}
+				return
+			}
+			if err != nil {
+				t.Errorf("unexpected error: %v", err)
+			}
+		})
+	}
+}
+
+func TestClientDo(t *testing.T) {
+	testCases := []struct {
+		name     string
+		response any
+		wantErr  string
+	}{
+		{
+			name: "immediate error response",
+			response: testError{
+				message:    "test error message",
+				statusCode: http.StatusBadRequest,
+			},
+			wantErr: "test error message",
+		},
+		{
+			name: "server error response",
+			response: testError{
+				message:    "internal error",
+				statusCode: http.StatusInternalServerError,
+			},
+			wantErr: "internal error",
+		},
+		{
+			name: "successful response",
+			response: struct {
+				ID      string `json:"id"`
+				Success bool   `json:"success"`
+			}{
+				ID:      "msg_123",
+				Success: true,
+			},
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				if errResp, ok := tc.response.(testError); ok {
+					w.WriteHeader(errResp.statusCode)
+					err := json.NewEncoder(w).Encode(map[string]string{
+						"error": errResp.message,
+					})
+					if err != nil {
+						t.Fatal("failed to encode error response:", err)
+					}
+					return
+				}
+
+				w.Header().Set("Content-Type", "application/json")
+				if err := json.NewEncoder(w).Encode(tc.response); err != nil {
+					t.Fatalf("failed to encode response: %v", err)
+				}
+			}))
+			defer ts.Close()
+
+			client := NewClient(&url.URL{Scheme: "http", Host: ts.Listener.Addr().String()}, http.DefaultClient)
+
+			var resp struct {
+				ID      string `json:"id"`
+				Success bool   `json:"success"`
+			}
+			err := client.do(context.Background(), http.MethodPost, "/v1/messages", nil, &resp)
+
+			if tc.wantErr != "" {
+				if err == nil {
+					t.Fatalf("got nil, want error %q", tc.wantErr)
+				}
+				if err.Error() != tc.wantErr {
+					t.Errorf("error message mismatch: got %q, want %q", err.Error(), tc.wantErr)
+				}
+				return
+			}
+
+			if err != nil {
+				t.Fatalf("got error %q, want nil", err)
+			}
+
+			if expectedResp, ok := tc.response.(struct {
+				ID      string `json:"id"`
+				Success bool   `json:"success"`
+			}); ok {
+				if resp.ID != expectedResp.ID {
+					t.Errorf("response ID mismatch: got %q, want %q", resp.ID, expectedResp.ID)
+				}
+				if resp.Success != expectedResp.Success {
+					t.Errorf("response Success mismatch: got %v, want %v", resp.Success, expectedResp.Success)
+				}
+			}
+		})
+	}
+}

api/types.go

@@ -10,6 +10,9 @@ import (
 	"strconv"
 	"strings"
 	"time"
+
+	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/types/model"
 )
 
 // StatusError is an error with an HTTP status code and message.
@@ -73,13 +76,13 @@ type GenerateRequest struct {
 	// this request.
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
-	// Images is an optional list of base64-encoded images accompanying this
+	// Images is an optional list of raw image bytes accompanying this
 	// request, for multimodal models.
 	Images []ImageData `json:"images,omitempty"`
 
 	// Options lists model-specific options. For example, temperature can be
 	// set through this field, if the model supports it.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -104,7 +107,7 @@ type ChatRequest struct {
 	Tools `json:"tools,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 type Tools []Tool
@@ -160,19 +163,65 @@ func (t *ToolCallFunctionArguments) String() string {
 
 type Tool struct {
 	Type     string       `json:"type"`
+	Items    any          `json:"items,omitempty"`
 	Function ToolFunction `json:"function"`
 }
 
+// PropertyType can be either a string or an array of strings
+type PropertyType []string
+
+// UnmarshalJSON implements the json.Unmarshaler interface
+func (pt *PropertyType) UnmarshalJSON(data []byte) error {
+	// Try to unmarshal as a string first
+	var s string
+	if err := json.Unmarshal(data, &s); err == nil {
+		*pt = []string{s}
+		return nil
+	}
+
+	// If that fails, try to unmarshal as an array of strings
+	var a []string
+	if err := json.Unmarshal(data, &a); err != nil {
+		return err
+	}
+	*pt = a
+	return nil
+}
+
+// MarshalJSON implements the json.Marshaler interface
+func (pt PropertyType) MarshalJSON() ([]byte, error) {
+	if len(pt) == 1 {
+		// If there's only one type, marshal as a string
+		return json.Marshal(pt[0])
+	}
+	// Otherwise marshal as an array
+	return json.Marshal([]string(pt))
+}
+
+// String returns a string representation of the PropertyType
+func (pt PropertyType) String() string {
+	if len(pt) == 0 {
+		return ""
+	}
+	if len(pt) == 1 {
+		return pt[0]
+	}
+	return fmt.Sprintf("%v", []string(pt))
+}
+
 type ToolFunction struct {
 	Name        string `json:"name"`
 	Description string `json:"description"`
 	Parameters  struct {
 		Type       string   `json:"type"`
+		Defs       any      `json:"$defs,omitempty"`
+		Items      any      `json:"items,omitempty"`
 		Required   []string `json:"required"`
 		Properties map[string]struct {
-			Type        string   `json:"type"`
-			Description string   `json:"description"`
-			Enum        []string `json:"enum,omitempty"`
+			Type        PropertyType `json:"type"`
+			Items       any          `json:"items,omitempty"`
+			Description string       `json:"description"`
+			Enum        []any        `json:"enum,omitempty"`
 		} `json:"properties"`
 	} `json:"parameters"`
 }
@@ -258,7 +307,7 @@ type EmbedRequest struct {
 	Truncate *bool `json:"truncate,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // EmbedResponse is the response from [Client.Embed].
@@ -284,7 +333,7 @@ type EmbeddingRequest struct {
 	KeepAlive *Duration `json:"keep_alive,omitempty"`
 
 	// Options lists model-specific options.
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 }
 
 // EmbeddingResponse is the response from [Client.Embeddings].
@@ -330,7 +379,7 @@ type ShowRequest struct {
 	Template string `json:"template"`
 	Verbose  bool   `json:"verbose"`
 
-	Options map[string]interface{} `json:"options"`
+	Options map[string]any `json:"options"`
 
 	// Deprecated: set the model name with Model instead
 	Name string `json:"name"`
@@ -338,16 +387,18 @@ type ShowRequest struct {
 
 // ShowResponse is the response returned from [Client.Show].
 type ShowResponse struct {
-	License       string         `json:"license,omitempty"`
-	Modelfile     string         `json:"modelfile,omitempty"`
-	Parameters    string         `json:"parameters,omitempty"`
-	Template      string         `json:"template,omitempty"`
-	System        string         `json:"system,omitempty"`
-	Details       ModelDetails   `json:"details,omitempty"`
-	Messages      []Message      `json:"messages,omitempty"`
-	ModelInfo     map[string]any `json:"model_info,omitempty"`
-	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
-	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
+	License       string             `json:"license,omitempty"`
+	Modelfile     string             `json:"modelfile,omitempty"`
+	Parameters    string             `json:"parameters,omitempty"`
+	Template      string             `json:"template,omitempty"`
+	System        string             `json:"system,omitempty"`
+	Details       ModelDetails       `json:"details,omitempty"`
+	Messages      []Message          `json:"messages,omitempty"`
+	ModelInfo     map[string]any     `json:"model_info,omitempty"`
+	ProjectorInfo map[string]any     `json:"projector_info,omitempty"`
+	Tensors       []Tensor           `json:"tensors,omitempty"`
+	Capabilities  []model.Capability `json:"capabilities,omitempty"`
+	ModifiedAt    time.Time          `json:"modified_at,omitempty"`
 }
 
 // CopyRequest is the request passed to [Client.Copy].
@@ -359,9 +410,9 @@ type CopyRequest struct {
 // PullRequest is the request passed to [Client.Pull].
 type PullRequest struct {
 	Model    string `json:"model"`
-	Insecure bool   `json:"insecure,omitempty"`
-	Username string `json:"username"`
-	Password string `json:"password"`
+	Insecure bool   `json:"insecure,omitempty"` // Deprecated: ignored
+	Username string `json:"username"`           // Deprecated: ignored
+	Password string `json:"password"`           // Deprecated: ignored
 	Stream   *bool  `json:"stream,omitempty"`
 
 	// Deprecated: set the model name with Model instead
@@ -465,6 +516,13 @@ type ModelDetails struct {
 	QuantizationLevel string   `json:"quantization_level"`
 }
 
+// Tensor describes the metadata for a given tensor.
+type Tensor struct {
+	Name  string   `json:"name"`
+	Type  string   `json:"type"`
+	Shape []uint64 `json:"shape"`
+}
+
 func (m *Metrics) Summary() {
 	if m.TotalDuration > 0 {
 		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)
@@ -493,7 +551,7 @@ func (m *Metrics) Summary() {
 	}
 }
 
-func (opts *Options) FromMap(m map[string]interface{}) error {
+func (opts *Options) FromMap(m map[string]any) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct
 
@@ -550,12 +608,12 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 				}
 				field.SetString(val)
 			case reflect.Slice:
-				// JSON unmarshals to []interface{}, not []string
-				val, ok := val.([]interface{})
+				// JSON unmarshals to []any, not []string
+				val, ok := val.([]any)
 				if !ok {
 					return fmt.Errorf("option %q must be of type array", key)
 				}
-				// convert []interface{} to []string
+				// convert []any to []string
 				slice := make([]string, len(val))
 				for i, item := range val {
 					str, ok := item.(string)
@@ -609,7 +667,7 @@ func DefaultOptions() Options {
 
 		Runner: Runner{
 			// options set when the model is loaded
-			NumCtx:    2048,
+			NumCtx:    int(envconfig.ContextLength()),
 			NumBatch:  512,
 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
 			NumThread: 0,  // let the runtime decide
@@ -662,7 +720,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 }
 
 // FormatParams converts specified parameter options to their correct types
-func FormatParams(params map[string][]string) (map[string]interface{}, error) {
+func FormatParams(params map[string][]string) (map[string]any, error) {
 	opts := Options{}
 	valueOpts := reflect.ValueOf(&opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts)           // types of the fields in the options struct
@@ -676,7 +734,7 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		}
 	}
 
-	out := make(map[string]interface{})
+	out := make(map[string]any)
 	// iterate params and set values based on json struct tags
 	for key, vals := range params {
 		if opt, ok := jsonOpts[key]; !ok {

api/types_test.go

@@ -134,7 +134,7 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
-			var oMap map[string]interface{}
+			var oMap map[string]any
 			err := json.Unmarshal([]byte(test.req), &oMap)
 			require.NoError(t, err)
 			opts := DefaultOptions()
@@ -231,3 +231,144 @@ func TestMessage_UnmarshalJSON(t *testing.T) {
 		}
 	}
 }
+
+func TestToolFunction_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   string
+		wantErr string
+	}{
+		{
+			name: "valid enum with same types",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": ["a", "b", "c"]
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+		{
+			name: "empty enum array",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": []
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var tf ToolFunction
+			err := json.Unmarshal([]byte(tt.input), &tf)
+
+			if tt.wantErr != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
+func TestPropertyType_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected PropertyType
+	}{
+		{
+			name:     "string type",
+			input:    `"string"`,
+			expected: PropertyType{"string"},
+		},
+		{
+			name:     "array of types",
+			input:    `["string", "number"]`,
+			expected: PropertyType{"string", "number"},
+		},
+		{
+			name:     "array with single type",
+			input:    `["string"]`,
+			expected: PropertyType{"string"},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var pt PropertyType
+			if err := json.Unmarshal([]byte(test.input), &pt); err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if len(pt) != len(test.expected) {
+				t.Errorf("Length mismatch: got %v, expected %v", len(pt), len(test.expected))
+			}
+
+			for i, v := range pt {
+				if v != test.expected[i] {
+					t.Errorf("Value mismatch at index %d: got %v, expected %v", i, v, test.expected[i])
+				}
+			}
+		})
+	}
+}
+
+func TestPropertyType_MarshalJSON(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    PropertyType
+		expected string
+	}{
+		{
+			name:     "single type",
+			input:    PropertyType{"string"},
+			expected: `"string"`,
+		},
+		{
+			name:     "multiple types",
+			input:    PropertyType{"string", "number"},
+			expected: `["string","number"]`,
+		},
+		{
+			name:     "empty type",
+			input:    PropertyType{},
+			expected: `[]`,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			data, err := json.Marshal(test.input)
+			if err != nil {
+				t.Errorf("Unexpected error: %v", err)
+			}
+
+			if string(data) != test.expected {
+				t.Errorf("Marshaled data mismatch: got %v, expected %v", string(data), test.expected)
+			}
+		})
+	}
+}

benchmark/server_benchmark_test.go

@@ -0,0 +1,178 @@
+package benchmark
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// Command line flags
+var modelFlag string
+
+func init() {
+	flag.StringVar(&modelFlag, "m", "", "Name of the model to benchmark")
+	flag.Lookup("m").DefValue = "model"
+}
+
+// modelName returns the model name from flags, failing the test if not set
+func modelName(b *testing.B) string {
+	if modelFlag == "" {
+		b.Fatal("Error: -m flag is required for benchmark tests")
+	}
+	return modelFlag
+}
+
+type TestCase struct {
+	name      string
+	prompt    string
+	maxTokens int
+}
+
+// runGenerateBenchmark contains the common generate and metrics logic
+func runGenerateBenchmark(b *testing.B, ctx context.Context, client *api.Client, req *api.GenerateRequest) {
+	start := time.Now()
+	var ttft time.Duration
+	var metrics api.Metrics
+
+	err := client.Generate(ctx, req, func(resp api.GenerateResponse) error {
+		if ttft == 0 && resp.Response != "" {
+			ttft = time.Since(start)
+		}
+		if resp.Done {
+			metrics = resp.Metrics
+		}
+		return nil
+	})
+
+	// Report custom metrics as part of the benchmark results
+	b.ReportMetric(float64(ttft.Milliseconds()), "ttft_ms")
+	b.ReportMetric(float64(metrics.LoadDuration.Milliseconds()), "load_ms")
+
+	// Token throughput metrics
+	promptThroughput := float64(metrics.PromptEvalCount) / metrics.PromptEvalDuration.Seconds()
+	genThroughput := float64(metrics.EvalCount) / metrics.EvalDuration.Seconds()
+	b.ReportMetric(promptThroughput, "prompt_tok/s")
+	b.ReportMetric(genThroughput, "gen_tok/s")
+
+	// Token counts
+	b.ReportMetric(float64(metrics.PromptEvalCount), "prompt_tokens")
+	b.ReportMetric(float64(metrics.EvalCount), "gen_tokens")
+	if err != nil {
+		b.Fatal(err)
+	}
+}
+
+// BenchmarkColdStart runs benchmarks with model loading from cold state
+func BenchmarkColdStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/cold/%s", m, tt.name), func(b *testing.B) {
+			ctx := context.Background()
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				b.StopTimer()
+				// Ensure model is unloaded before each iteration
+				unload(client, m, b)
+				b.StartTimer()
+
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// BenchmarkWarmStart runs benchmarks with pre-loaded model
+func BenchmarkWarmStart(b *testing.B) {
+	client := setup(b)
+	tests := []TestCase{
+		{"short_prompt", "Write a long story", 100},
+		{"medium_prompt", "Write a detailed economic analysis", 500},
+		{"long_prompt", "Write a comprehensive AI research paper", 1000},
+	}
+	m := modelName(b)
+
+	for _, tt := range tests {
+		b.Run(fmt.Sprintf("%s/warm/%s", m, tt.name), func(b *testing.B) {
+			ctx := context.Background()
+
+			// Pre-warm the model
+			warmup(client, m, tt.prompt, b)
+
+			// Set number of tokens as our throughput metric
+			b.SetBytes(int64(tt.maxTokens))
+
+			for b.Loop() {
+				req := &api.GenerateRequest{
+					Model:   m,
+					Prompt:  tt.prompt,
+					Options: map[string]any{"num_predict": tt.maxTokens, "temperature": 0.1},
+				}
+
+				runGenerateBenchmark(b, ctx, client, req)
+			}
+		})
+	}
+}
+
+// setup verifies server and model availability
+func setup(b *testing.B) *api.Client {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		b.Fatal(err)
+	}
+	if _, err := client.Show(context.Background(), &api.ShowRequest{Model: modelName(b)}); err != nil {
+		b.Fatalf("Model unavailable: %v", err)
+	}
+
+	return client
+}
+
+// warmup ensures the model is loaded and warmed up
+func warmup(client *api.Client, model string, prompt string, b *testing.B) {
+	for range 3 {
+		err := client.Generate(
+			context.Background(),
+			&api.GenerateRequest{
+				Model:   model,
+				Prompt:  prompt,
+				Options: map[string]any{"num_predict": 50, "temperature": 0.1},
+			},
+			func(api.GenerateResponse) error { return nil },
+		)
+		if err != nil {
+			b.Logf("Error during model warm-up: %v", err)
+		}
+	}
+}
+
+// unload forces model unloading using KeepAlive: 0 parameter
+func unload(client *api.Client, model string, b *testing.B) {
+	req := &api.GenerateRequest{
+		Model:     model,
+		KeepAlive: &api.Duration{Duration: 0},
+	}
+	if err := client.Generate(context.Background(), req, func(api.GenerateResponse) error { return nil }); err != nil {
+		b.Logf("Unload error: %v", err)
+	}
+	time.Sleep(1 * time.Second)
+}

cmd/cmd.go

@@ -18,6 +18,8 @@ import (
 	"os/signal"
 	"path/filepath"
 	"runtime"
+	"slices"
+	"sort"
 	"strconv"
 	"strings"
 	"sync/atomic"
@@ -34,7 +36,6 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
-	"github.com/ollama/ollama/llama"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/progress"
 	"github.com/ollama/ollama/runner"
@@ -256,6 +257,7 @@ func StopHandler(cmd *cobra.Command, args []string) error {
 		if strings.Contains(err.Error(), "not found") {
 			return fmt.Errorf("couldn't find model \"%s\" to stop", args[0])
 		}
+		return err
 	}
 	return nil
 }
@@ -266,7 +268,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	opts := runOptions{
 		Model:    args[0],
 		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]interface{}{},
+		Options:  map[string]any{},
 	}
 
 	format, err := cmd.Flags().GetString("format")
@@ -338,10 +340,21 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}
 
-	// TODO(jessegross): We should either find another way to know if this is
-	// a vision model or remove the logic. Also consider that other modalities will
-	// need different behavior anyways.
-	opts.MultiModal = len(info.ProjectorInfo) != 0 || envconfig.NewEngine()
+	opts.MultiModal = slices.Contains(info.Capabilities, model.CapabilityVision)
+
+	// TODO: remove the projector info and vision info checks below,
+	// these are left in for backwards compatibility with older servers
+	// that don't have the capabilities field in the model info
+	if len(info.ProjectorInfo) != 0 {
+		opts.MultiModal = true
+	}
+	for k := range info.ModelInfo {
+		if strings.Contains(k, ".vision.") {
+			opts.MultiModal = true
+			break
+		}
+	}
+
 	opts.ParentModel = info.Details.ParentModel
 
 	if interactive {
@@ -562,8 +575,9 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 	parameters, errParams := cmd.Flags().GetBool("parameters")
 	system, errSystem := cmd.Flags().GetBool("system")
 	template, errTemplate := cmd.Flags().GetBool("template")
+	verbose, errVerbose := cmd.Flags().GetBool("verbose")
 
-	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate} {
+	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate, errVerbose} {
 		if boolErr != nil {
 			return errors.New("error retrieving flags")
 		}
@@ -601,7 +615,7 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
 	}
 
-	req := api.ShowRequest{Name: args[0]}
+	req := api.ShowRequest{Name: args[0], Verbose: verbose}
 	resp, err := client.Show(cmd.Context(), &req)
 	if err != nil {
 		return err
@@ -624,10 +638,10 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	return showInfo(resp, os.Stdout)
+	return showInfo(resp, verbose, os.Stdout)
 }
 
-func showInfo(resp *api.ShowResponse, w io.Writer) error {
+func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 	tableRender := func(header string, rows func() [][]string) {
 		fmt.Fprintln(w, " ", header)
 		table := tablewriter.NewWriter(w)
@@ -661,6 +675,15 @@ func showInfo(resp *api.ShowResponse, w io.Writer) error {
 		return
 	})
 
+	if len(resp.Capabilities) > 0 {
+		tableRender("Capabilities", func() (rows [][]string) {
+			for _, capability := range resp.Capabilities {
+				rows = append(rows, []string{"", capability.String()})
+			}
+			return
+		})
+	}
+
 	if resp.ProjectorInfo != nil {
 		tableRender("Projector", func() (rows [][]string) {
 			arch := resp.ProjectorInfo["general.architecture"].(string)
@@ -684,6 +707,47 @@ func showInfo(resp *api.ShowResponse, w io.Writer) error {
 		})
 	}
 
+	if resp.ModelInfo != nil && verbose {
+		tableRender("Metadata", func() (rows [][]string) {
+			keys := make([]string, 0, len(resp.ModelInfo))
+			for k := range resp.ModelInfo {
+				keys = append(keys, k)
+			}
+			sort.Strings(keys)
+
+			for _, k := range keys {
+				var v string
+				switch vData := resp.ModelInfo[k].(type) {
+				case bool:
+					v = fmt.Sprintf("%t", vData)
+				case string:
+					v = vData
+				case float64:
+					v = fmt.Sprintf("%g", vData)
+				case []any:
+					n := 3
+					if len(vData) < n {
+						n = len(vData)
+					}
+					v = fmt.Sprintf("%v", vData[:n])
+				default:
+					v = fmt.Sprintf("%T", vData)
+				}
+				rows = append(rows, []string{"", k, v})
+			}
+			return
+		})
+	}
+
+	if len(resp.Tensors) > 0 && verbose {
+		tableRender("Tensors", func() (rows [][]string) {
+			for _, t := range resp.Tensors {
+				rows = append(rows, []string{"", t.Name, t.Type, fmt.Sprint(t.Shape)})
+			}
+			return
+		})
+	}
+
 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
 		for scanner.Scan() && (len(rows) < n || n < 0) {
@@ -744,13 +808,38 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
+			if resp.Completed == 0 {
+				// This is the initial status update for the
+				// layer, which the server sends before
+				// beginning the download, for clients to
+				// compute total size and prepare for
+				// downloads, if needed.
+				//
+				// Skipping this here to avoid showing a 0%
+				// progress bar, which *should* clue the user
+				// into the fact that many things are being
+				// downloaded and that the current active
+				// download is not that last. However, in rare
+				// cases it seems to be triggering to some, and
+				// it isn't worth explaining, so just ignore
+				// and regress to the old UI that keeps giving
+				// you the "But wait, there is more!" after
+				// each "100% done" bar, which is "better."
+				return nil
+			}
+
 			if spinner != nil {
 				spinner.Stop()
 			}
 
 			bar, ok := bars[resp.Digest]
 			if !ok {
-				bar = progress.NewBar(fmt.Sprintf("pulling %s...", resp.Digest[7:19]), resp.Total, resp.Completed)
+				name, isDigest := strings.CutPrefix(resp.Digest, "sha256:")
+				name = strings.TrimSpace(name)
+				if isDigest {
+					name = name[:min(12, len(name))]
+				}
+				bar = progress.NewBar(fmt.Sprintf("pulling %s:", name), resp.Total, resp.Completed)
 				bars[resp.Digest] = bar
 				p.Add(resp.Digest, bar)
 			}
@@ -770,11 +859,7 @@ func PullHandler(cmd *cobra.Command, args []string) error {
 	}
 
 	request := api.PullRequest{Name: args[0], Insecure: insecure}
-	if err := client.Pull(cmd.Context(), &request, fn); err != nil {
-		return err
-	}
-
-	return nil
+	return client.Pull(cmd.Context(), &request, fn)
 }
 
 type generateContextKey string
@@ -788,7 +873,7 @@ type runOptions struct {
 	Format      string
 	System      string
 	Images      []api.ImageData
-	Options     map[string]interface{}
+	Options     map[string]any
 	MultiModal  bool
 	KeepAlive   *api.Duration
 }
@@ -1190,6 +1275,7 @@ func NewCLI() *cobra.Command {
 	showCmd.Flags().Bool("parameters", false, "Show parameters of a model")
 	showCmd.Flags().Bool("template", false, "Show template of a model")
 	showCmd.Flags().Bool("system", false, "Show system message of a model")
+	showCmd.Flags().BoolP("verbose", "v", false, "Show detailed model information")
 
 	runCmd := &cobra.Command{
 		Use:     "run MODEL [PROMPT]",
@@ -1274,7 +1360,6 @@ func NewCLI() *cobra.Command {
 
 	runnerCmd := &cobra.Command{
 		Use:    "runner",
-		Short:  llama.PrintSystemInfo(),
 		Hidden: true,
 		RunE: func(cmd *cobra.Command, args []string) error {
 			return runner.Execute(os.Args[1:])
@@ -1317,12 +1402,12 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
 				envVars["OLLAMA_SCHED_SPREAD"],
-				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
+				envVars["OLLAMA_CONTEXT_LENGTH"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

cmd/cmd_test.go

@@ -16,6 +16,7 @@ import (
 	"github.com/spf13/cobra"
 
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
 )
 
 func TestShowInfo(t *testing.T) {
@@ -27,7 +28,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -57,7 +58,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -68,6 +69,60 @@ func TestShowInfo(t *testing.T) {
     embedding length    0       
     quantization        FP16    
 
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("verbose model", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "8B",
+				QuantizationLevel: "FP16",
+			},
+			Parameters: `
+			stop up`,
+			ModelInfo: map[string]any{
+				"general.architecture":    "test",
+				"general.parameter_count": float64(8_000_000_000),
+				"some.true_bool":          true,
+				"some.false_bool":         false,
+				"test.context_length":     float64(1000),
+				"test.embedding_length":   float64(11434),
+			},
+			Tensors: []api.Tensor{
+				{Name: "blk.0.attn_k.weight", Type: "BF16", Shape: []uint64{42, 3117}},
+				{Name: "blk.0.attn_q.weight", Type: "FP16", Shape: []uint64{3117, 42}},
+			},
+		}, true, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture        test     
+    parameters          8B       
+    context length      1000     
+    embedding length    11434    
+    quantization        FP16     
+
+  Parameters
+    stop    up    
+
+  Metadata
+    general.architecture       test     
+    general.parameter_count    8e+09    
+    some.false_bool            false    
+    some.true_bool             true     
+    test.context_length        1000     
+    test.embedding_length      11434    
+
+  Tensors
+    blk.0.attn_k.weight    BF16    [42 3117]    
+    blk.0.attn_q.weight    FP16    [3117 42]    
+
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
@@ -89,7 +144,7 @@ func TestShowInfo(t *testing.T) {
 			stop you
 			stop up
 			temperature 99`,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -126,7 +181,7 @@ func TestShowInfo(t *testing.T) {
 				"clip.vision.embedding_length": float64(0),
 				"clip.vision.projection_dim":   float64(0),
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -159,7 +214,7 @@ func TestShowInfo(t *testing.T) {
 Ahoy, matey!
 Weigh anchor!
 			`,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -188,7 +243,7 @@ Weigh anchor!
 				QuantizationLevel: "FP16",
 			},
 			License: license,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -206,6 +261,34 @@ Weigh anchor!
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
 		}
 	})
+
+	t.Run("capabilities", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "7B",
+				QuantizationLevel: "FP16",
+			},
+			Capabilities: []model.Capability{model.CapabilityVision, model.CapabilityTools},
+		}, false, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := "  Model\n" +
+			"    architecture    test    \n" +
+			"    parameters      7B      \n" +
+			"    quantization    FP16    \n" +
+			"\n" +
+			"  Capabilities\n" +
+			"    vision    \n" +
+			"    tools     \n" +
+			"\n"
+
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
 }
 
 func TestDeleteHandler(t *testing.T) {
@@ -332,6 +415,7 @@ func TestGetModelfileName(t *testing.T) {
 				if err != nil {
 					t.Fatalf("temp modelfile creation failed: %v", err)
 				}
+				defer tempFile.Close()
 
 				expectedFilename = tempFile.Name()
 				err = cmd.Flags().Set("file", expectedFilename)
@@ -707,3 +791,132 @@ func TestCreateHandler(t *testing.T) {
 		})
 	}
 }
+
+func TestNewCreateRequest(t *testing.T) {
+	tests := []struct {
+		name     string
+		from     string
+		opts     runOptions
+		expected *api.CreateRequest
+	}{
+		{
+			"basic test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "",
+				Prompt:      "You are a fun AI agent",
+				Messages:    []api.Message{},
+				WordWrap:    true,
+			},
+			&api.CreateRequest{
+				From:  "mymodel",
+				Model: "newmodel",
+			},
+		},
+		{
+			"parent model test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "parentmodel",
+				Messages:    []api.Message{},
+				WordWrap:    true,
+			},
+			&api.CreateRequest{
+				From:  "parentmodel",
+				Model: "newmodel",
+			},
+		},
+		{
+			"parent model as filepath test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "/some/file/like/etc/passwd",
+				Messages:    []api.Message{},
+				WordWrap:    true,
+			},
+			&api.CreateRequest{
+				From:  "mymodel",
+				Model: "newmodel",
+			},
+		},
+		{
+			"parent model as windows filepath test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "D:\\some\\file\\like\\etc\\passwd",
+				Messages:    []api.Message{},
+				WordWrap:    true,
+			},
+			&api.CreateRequest{
+				From:  "mymodel",
+				Model: "newmodel",
+			},
+		},
+		{
+			"options test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "parentmodel",
+				Options: map[string]any{
+					"temperature": 1.0,
+				},
+			},
+			&api.CreateRequest{
+				From:  "parentmodel",
+				Model: "newmodel",
+				Parameters: map[string]any{
+					"temperature": 1.0,
+				},
+			},
+		},
+		{
+			"messages test",
+			"newmodel",
+			runOptions{
+				Model:       "mymodel",
+				ParentModel: "parentmodel",
+				System:      "You are a fun AI agent",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "hello there!",
+					},
+					{
+						Role:    "assistant",
+						Content: "hello to you!",
+					},
+				},
+				WordWrap: true,
+			},
+			&api.CreateRequest{
+				From:   "parentmodel",
+				Model:  "newmodel",
+				System: "You are a fun AI agent",
+				Messages: []api.Message{
+					{
+						Role:    "user",
+						Content: "hello there!",
+					},
+					{
+						Role:    "assistant",
+						Content: "hello to you!",
+					},
+				},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			actual := NewCreateRequest(tt.from, tt.opts)
+			if !cmp.Equal(actual, tt.expected) {
+				t.Errorf("expected output %#v, got %#v", tt.expected, actual)
+			}
+		})
+	}
+}

cmd/interactive.go

@@ -18,6 +18,7 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/readline"
 	"github.com/ollama/ollama/types/errtypes"
+	"github.com/ollama/ollama/types/model"
 )
 
 type MultilineState int
@@ -195,6 +196,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
+				if strings.Contains(err.Error(), "not found") {
+					fmt.Printf("error: %v\n", err)
+					continue
+				}
 				return err
 			}
 			continue
@@ -343,7 +348,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					_ = showInfo(resp, os.Stderr)
+					_ = showInfo(resp, false, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")
@@ -455,9 +460,16 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 }
 
 func NewCreateRequest(name string, opts runOptions) *api.CreateRequest {
+	parentModel := opts.ParentModel
+
+	modelName := model.ParseName(parentModel)
+	if !modelName.IsValid() {
+		parentModel = ""
+	}
+
 	req := &api.CreateRequest{
-		Name: name,
-		From: cmp.Or(opts.ParentModel, opts.Model),
+		Model: name,
+		From:  cmp.Or(parentModel, opts.Model),
 	}
 
 	if opts.System != "" {
@@ -491,6 +503,7 @@ func normalizeFilePath(fp string) string {
 		"\\\\", "\\", // Escaped backslash
 		"\\*", "*", // Escaped asterisk
 		"\\?", "?", // Escaped question mark
+		"\\~", "~", // Escaped tilde
 	).Replace(fp)
 }
 

convert/convert.go

@@ -7,14 +7,20 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
+	"slices"
 	"strings"
 
 	"github.com/ollama/ollama/fs/ggml"
 )
 
 type ModelParameters struct {
-	Architectures []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+	Architectures []string       `json:"architectures"`
+	VocabSize     uint32         `json:"vocab_size"`
+	TextModel     TextParameters `json:"text_config"`
+}
+
+type TextParameters struct {
+	VocabSize uint32 `json:"vocab_size"`
 }
 
 type AdapterParameters struct {
@@ -79,14 +85,6 @@ func (ModelParameters) specialTokenTypes() []string {
 	}
 }
 
-func (ModelParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
-}
-
-func (AdapterParameters) writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
-	return ggml.WriteGGUF(ws, kv, ts)
-}
-
 type ModelConverter interface {
 	// KV maps parameters to LLM key-values
 	KV(*Tokenizer) ggml.KV
@@ -98,8 +96,6 @@ type ModelConverter interface {
 
 	// specialTokenTypes returns any special token types the model uses
 	specialTokenTypes() []string
-	// writeFile writes the model to the provided io.WriteSeeker
-	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 type moreParser interface {
@@ -114,8 +110,6 @@ type AdapterConverter interface {
 	// Replacements returns a list of string pairs to replace in tensor names.
 	// See [strings.Replacer](https://pkg.go.dev/strings#Replacer) for details
 	Replacements() []string
-
-	writeFile(io.WriteSeeker, ggml.KV, []ggml.Tensor) error
 }
 
 func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
@@ -153,7 +147,7 @@ func ConvertAdapter(fsys fs.FS, ws io.WriteSeeker, baseKV ggml.KV) error {
 		return err
 	}
 
-	return conv.writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
+	return writeFile(ws, conv.KV(baseKV), conv.Tensors(ts))
 }
 
 // Convert writes an Ollama compatible model to the provided io.WriteSeeker based on configurations
@@ -177,14 +171,20 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 
 	var conv ModelConverter
 	switch p.Architectures[0] {
-	case "LlamaForCausalLM", "MistralForCausalLM":
+	case "LlamaForCausalLM":
 		conv = &llamaModel{}
+	case "Llama4ForConditionalGeneration":
+		conv = &llama4Model{}
+	case "Mistral3ForConditionalGeneration":
+		conv = &mistral3Model{}
 	case "MixtralForCausalLM":
 		conv = &mixtralModel{}
 	case "GemmaForCausalLM":
 		conv = &gemmaModel{}
 	case "Gemma2ForCausalLM":
 		conv = &gemma2Model{}
+	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
+		conv = &gemma3Model{Architecture: p.Architectures[0]}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
@@ -194,7 +194,7 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	case "CohereForCausalLM":
 		conv = &commandrModel{}
 	default:
-		return errors.New("unsupported architecture")
+		return fmt.Errorf("unsupported architecture %q", p.Architectures[0])
 	}
 
 	if err := json.Unmarshal(bts, conv); err != nil {
@@ -213,7 +213,14 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	}
 
 	vocabSize := int(p.VocabSize)
+	if vocabSize == 0 {
+		tVocabSize := int(p.TextModel.VocabSize)
+		vocabSize = tVocabSize
+	}
+
 	switch {
+	case vocabSize == 0:
+		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
 		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {
@@ -232,5 +239,13 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		return err
 	}
 
-	return conv.writeFile(ws, conv.KV(t), conv.Tensors(ts))
+	return writeFile(ws, conv.KV(t), conv.Tensors(ts))
+}
+
+func writeFile(ws io.WriteSeeker, kv ggml.KV, ts []ggml.Tensor) error {
+	for i := range ts {
+		ts[i].Shape = slices.Clone(ts[i].Shape)
+		slices.Reverse(ts[i].Shape)
+	}
+	return ggml.WriteGGUF(ws, kv, ts)
 }

convert/convert_gemma.go

@@ -45,7 +45,7 @@ func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
 	var out []ggml.Tensor
 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "_norm.weight") {
+		if !strings.HasPrefix(t.Name(), "v.") && strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 

convert/convert_gemma3.go

@@ -0,0 +1,142 @@
+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type gemma3Model struct {
+	gemmaModel
+	Architecture string
+	TextModel    struct {
+		HeadDim          uint32 `json:"head_dim"`
+		HiddenSize       uint32 `json:"hidden_size"`
+		HiddenLayers     uint32 `json:"num_hidden_layers"`
+		IntermediateSize uint32 `json:"intermediate_size"`
+		SlidingWindow    uint32 `json:"sliding_window"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumAttentionHeads uint32  `json:"num_attention_heads"` // attention.head_count 16
+		LayerNormEpsilon  float32 `json:"layer_norm_eps"`      // attention.layer_norm_epsilon 1e-05
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`   // block_count 32
+		HiddenSize        uint32  `json:"hidden_size"`         // embedding_length 1280
+		IntermediateSize  uint32  `json:"intermediate_size"`   // feed_forward_length 5120
+		ImageSize         uint32  `json:"image_size"`          // image_size 560
+		NumChannels       uint32  `json:"num_channels"`        // num_channels 3
+		PatchSize         uint32  `json:"patch_size"`          // patch_size 14
+	} `json:"vision_config"`
+	MaxPositionEmbeddings    uint32  `json:"max_position_embeddings"`
+	NumAttentionHeads        uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads         uint32  `json:"num_key_value_heads"`
+	RMSNormEPS               float32 `json:"rms_norm_eps"`
+	HeadDim                  uint32  `json:"head_dim"`
+	FinalLogitSoftcap        float32 `json:"final_logit_softcapping"`
+	RopeLocalTheta           float32 `json:"rope_local_base_freq"`
+	RopeGlobalTheta          float32 `json:"rope_global_base_freq"`
+	SlidingWindow            uint32  `json:"sliding_window"`
+	MultiModalTokensPerImage uint32  `json:"mm_tokens_per_image"`
+}
+
+const (
+	gemma4BLayerCount  = 34
+	gemma12BLayerCount = 48
+	gemma27BLayerCount = 62
+)
+
+func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma3"
+
+	numBlocks := cmp.Or(p.HiddenLayers, p.TextModel.HiddenLayers)
+	kv["gemma3.block_count"] = numBlocks
+
+	var (
+		numHeads   uint32
+		numKVHeads uint32
+	)
+
+	switch numBlocks {
+	case gemma4BLayerCount:
+		numHeads = 8
+		numKVHeads = 4
+	case gemma12BLayerCount:
+		numHeads = 16
+		numKVHeads = 8
+	case gemma27BLayerCount:
+		numHeads = 32
+		numKVHeads = 16
+	default:
+		numHeads = p.NumAttentionHeads
+		numKVHeads = p.NumKeyValueHeads
+	}
+
+	kv["gemma3.attention.head_count"] = numHeads
+	kv["gemma3.attention.head_count_kv"] = numKVHeads
+
+	switch p.Architecture {
+	case "Gemma3ForCausalLM":
+		kv["gemma3.context_length"] = p.MaxPositionEmbeddings
+		kv["gemma3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+		kv["gemma3.attention.key_length"] = p.HeadDim
+		kv["gemma3.attention.value_length"] = p.HeadDim
+		kv["gemma3.attention.sliding_window"] = p.SlidingWindow
+		kv["gemma3.final_logit_softcapping"] = cmp.Or(p.FinalLogitSoftcap, 30)
+		kv["gemma3.rope.local.freq_base"] = cmp.Or(p.RopeLocalTheta, 10000.0)
+		kv["gemma3.rope.global.freq_base"] = cmp.Or(p.RopeGlobalTheta, 1000000.0)
+		kv["gemma3.embedding_length"] = p.HiddenSize
+		kv["gemma3.feed_forward_length"] = p.IntermediateSize
+	default:
+		kv["gemma3.context_length"] = cmp.Or(p.MaxPositionEmbeddings, 131072)
+		kv["gemma3.embedding_length"] = p.TextModel.HiddenSize
+		kv["gemma3.feed_forward_length"] = p.TextModel.IntermediateSize
+		kv["gemma3.attention.sliding_window"] = p.TextModel.SlidingWindow
+		kv["gemma3.vision.block_count"] = p.VisionModel.NumHiddenLayers
+		kv["gemma3.vision.embedding_length"] = p.VisionModel.HiddenSize
+		kv["gemma3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+		kv["gemma3.vision.image_size"] = p.VisionModel.ImageSize
+		kv["gemma3.vision.patch_size"] = p.VisionModel.PatchSize
+		kv["gemma3.vision.num_channels"] = cmp.Or(p.VisionModel.NumChannels, 3)
+		kv["gemma3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+		kv["gemma3.vision.attention.layer_norm_epsilon"] = cmp.Or(p.VisionModel.LayerNormEpsilon, 1e-6)
+		kv["gemma3.attention.key_length"] = cmp.Or(p.TextModel.HeadDim, 256)
+		kv["gemma3.attention.value_length"] = cmp.Or(p.TextModel.HeadDim, 256)
+	}
+
+	if p.MultiModalTokensPerImage > 0 {
+		kv["gemma3.mm.tokens_per_image"] = p.MultiModalTokensPerImage
+	}
+
+	return kv
+}
+
+func (p *gemma3Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"vision_tower.vision_model.embeddings", "v",
+		"vision_tower.vision_model", "v",
+		"vision_model.vision_model.embeddings", "v",
+		"vision_model.vision_model", "v",
+		"language_model.", "",
+		"model.layers", "blk",
+		"encoder.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.out_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"input_projection_weight", "input_projection.weight",
+		"multi_modal_projector", "mm",
+	}
+}

convert/convert_llama.go

@@ -28,12 +28,12 @@ type llamaModel struct {
 	NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
 	RopeTheta             float32 `json:"rope_theta"`
 	RopeScaling           struct {
-		Type                            string  `json:"type"`
-		RopeType                        string  `json:"rope_type"`
-		Factor                          float32 `json:"factor"`
-		LowFrequencyFactor              float32 `json:"low_freq_factor"`
-		HighFrequencyFactor             float32 `json:"high_freq_factor"`
-		OriginalMaxPositionalEmbeddings uint32  `json:"original_max_positional_embeddings"`
+		Type                          string  `json:"type"`
+		RopeType                      string  `json:"rope_type"`
+		Factor                        float32 `json:"factor"`
+		LowFrequencyFactor            float32 `json:"low_freq_factor"`
+		HighFrequencyFactor           float32 `json:"high_freq_factor"`
+		OriginalMaxPositionEmbeddings uint32  `json:"original_max_position_embeddings"`
 
 		factors ropeFactor
 	} `json:"rope_scaling"`
@@ -42,6 +42,8 @@ type llamaModel struct {
 	LayerNormEpsilon float32 `json:"layer_norm_epsilon"`
 	NormEpsilon      float32 `json:"norm_epsilon"`
 	HeadDim          uint32  `json:"head_dim"`
+
+	skipRepack bool
 }
 
 var _ ModelConverter = (*llamaModel)(nil)
@@ -70,6 +72,10 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 		kv["llama.rope.dimension_count"] = p.HiddenSize / headCount
 	}
 
+	if p.HeadDim > 0 {
+		kv["llama.attention.head_dim"] = p.HeadDim
+	}
+
 	if p.RopeTheta > 0 {
 		kv["llama.rope.freq_base"] = p.RopeTheta
 	}
@@ -84,7 +90,7 @@ func (p *llamaModel) KV(t *Tokenizer) ggml.KV {
 			factorLow := cmp.Or(p.RopeScaling.LowFrequencyFactor, 1.0)
 			factorHigh := cmp.Or(p.RopeScaling.HighFrequencyFactor, 4.0)
 
-			original := cmp.Or(p.RopeScaling.OriginalMaxPositionalEmbeddings, 8192)
+			original := cmp.Or(p.RopeScaling.OriginalMaxPositionEmbeddings, 8192)
 			lambdaLow := float32(original) / factorLow
 			lambdaHigh := float32(original) / factorHigh
 
@@ -133,9 +139,10 @@ func (p *llamaModel) Tensors(ts []Tensor) []ggml.Tensor {
 	}
 
 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "attn_q.weight") ||
-			strings.HasSuffix(t.Name(), "attn_k.weight") {
-			t.SetRepacker(p.repack)
+		if strings.HasSuffix(t.Name(), "attn_q.weight") || strings.HasSuffix(t.Name(), "attn_k.weight") {
+			if !p.skipRepack {
+				t.SetRepacker(p.repack)
+			}
 		}
 
 		out = append(out, ggml.Tensor{

convert/convert_llama4.go

@@ -0,0 +1,169 @@
+package convert
+
+import (
+	"slices"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type llama4Model struct {
+	ModelParameters
+	TextModel struct {
+		llamaModel
+		NumExpertsPerToken     uint32 `json:"num_experts_per_tok"`
+		NumLocalExperts        uint32 `json:"num_local_experts"`
+		InterleaveMOELayerStep uint32 `json:"interleave_moe_layer_step"`
+		UseQKNorm              bool   `json:"use_qk_norm"`
+		IntermediateSizeMLP    uint32 `json:"intermediate_size_mlp"`
+		AttentionChunkSize     uint32 `json:"attention_chunk_size"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+		HiddenSize        uint32  `json:"hidden_size"`
+		IntermediateSize  uint32  `json:"intermediate_size"`
+		NumAttentionHeads uint32  `json:"num_attention_heads"`
+		ImageSize         uint32  `json:"image_size"`
+		PatchSize         uint32  `json:"patch_size"`
+		RopeTheta         float32 `json:"rope_theta"`
+		NormEpsilon       float32 `json:"norm_eps"`
+		PixelShuffleRatio float32 `json:"pixel_shuffle_ratio"`
+	} `json:"vision_config"`
+}
+
+// KV implements ModelConverter.
+func (p *llama4Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "llama4"
+
+	for k, v := range p.TextModel.KV(t) {
+		if strings.HasPrefix(k, "llama.") {
+			kv[strings.ReplaceAll(k, "llama.", "llama4.")] = v
+		}
+	}
+
+	kv["llama4.feed_forward_length"] = p.TextModel.IntermediateSizeMLP
+	kv["llama4.expert_feed_forward_length"] = p.TextModel.IntermediateSize
+
+	kv["llama4.expert_count"] = p.TextModel.NumLocalExperts
+	kv["llama4.expert_used_count"] = p.TextModel.NumExpertsPerToken
+	kv["llama4.interleave_moe_layer_step"] = p.TextModel.InterleaveMOELayerStep
+	kv["llama4.use_qk_norm"] = p.TextModel.UseQKNorm
+	kv["llama4.attention.chunk_size"] = p.TextModel.AttentionChunkSize
+
+	kv["llama4.vision.block_count"] = p.VisionModel.NumHiddenLayers
+	kv["llama4.vision.embedding_length"] = p.VisionModel.HiddenSize
+	kv["llama4.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+	kv["llama4.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+	kv["llama4.vision.image_size"] = p.VisionModel.ImageSize
+	kv["llama4.vision.patch_size"] = p.VisionModel.PatchSize
+	kv["llama4.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+	kv["llama4.vision.layer_norm_epsilon"] = p.VisionModel.NormEpsilon
+	kv["llama4.vision.pixel_shuffle_ratio"] = p.VisionModel.PixelShuffleRatio
+	return kv
+}
+
+// Replacements implements ModelConverter.
+func (p *llama4Model) Replacements() []string {
+	return append(
+		p.TextModel.Replacements(),
+		"language_model.", "",
+		"vision_model", "v",
+		"multi_modal_projector", "mm",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.", "ffn_",
+		"shared_expert.down_proj", "down_shexp",
+		"shared_expert.gate_proj", "gate_shexp",
+		"shared_expert.up_proj", "up_shexp",
+		"experts.down_proj", "down_exps.weight",
+		"experts.gate_up_proj", "gate_up_exps.weight",
+		"router", "gate_inp",
+		"patch_embedding.linear", "patch_embedding",
+	)
+}
+
+// Tensors implements ModelConverter.
+func (p *llama4Model) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
+
+	var textTensors []Tensor
+	for _, t := range ts {
+		if strings.HasPrefix(t.Name(), "v.") || strings.HasPrefix(t.Name(), "mm.") {
+			out = append(out, ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    t.Shape(),
+				WriterTo: t,
+			})
+		} else if strings.Contains(t.Name(), "ffn_gate_up_exps") {
+			// gate and up projectors are fused
+			// dims[1], dims[2] must be swapped
+			// [experts, hidden_size, intermediate_size * 2] --> [experts, intermediate_size, hidden_size]
+			halfDim := int(t.Shape()[2]) / 2
+
+			newShape := slices.Clone(t.Shape())
+			newShape[1], newShape[2] = newShape[2]/2, newShape[1]
+			for i, name := range []string{"ffn_gate_exps", "ffn_up_exps"} {
+				// clone tensor since we need separate repackers
+				tt := t.Clone()
+				tt.SetRepacker(p.repack(nil, nil, tensor.S(i*halfDim, (i+1)*halfDim)))
+				out = append(out, ggml.Tensor{
+					Name:     strings.ReplaceAll(tt.Name(), "ffn_gate_up_exps", name),
+					Kind:     tt.Kind(),
+					Shape:    newShape,
+					WriterTo: tt,
+				})
+			}
+		} else if strings.Contains(t.Name(), "ffn_down_exps") {
+			// dims[1], dims[2] must be swapped
+			// [experts, intermediate_size, hidden_size] --> [experts, hidden_size, intermediate_size]
+			t.SetRepacker(p.repack())
+			newShape := slices.Clone(t.Shape())
+			newShape[1], newShape[2] = newShape[2], newShape[1]
+			out = append(out, ggml.Tensor{
+				Name:     t.Name(),
+				Kind:     t.Kind(),
+				Shape:    newShape,
+				WriterTo: t,
+			})
+		} else {
+			textTensors = append(textTensors, t)
+		}
+	}
+
+	p.TextModel.skipRepack = true
+	out = append(out, p.TextModel.Tensors(textTensors)...)
+	return out
+}
+
+func (p *llama4Model) repack(slice ...tensor.Slice) Repacker {
+	return func(name string, data []float32, shape []uint64) ([]float32, error) {
+		dims := make([]int, len(shape))
+		for i, dim := range shape {
+			dims[i] = int(dim)
+		}
+
+		var t tensor.Tensor = tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+		t, err := t.Slice(slice...)
+		if err != nil {
+			return nil, err
+		}
+
+		if err := t.T(0, 2, 1); err != nil {
+			return nil, err
+		}
+
+		t = tensor.Materialize(t)
+		// flatten tensor so it can be return as a vector
+		if err := t.Reshape(t.Shape().TotalSize()); err != nil {
+			return nil, err
+		}
+
+		return native.VectorF32(t.(*tensor.Dense))
+	}
+}

convert/convert_mistral.go

@@ -0,0 +1,190 @@
+package convert
+
+import (
+	"cmp"
+	"fmt"
+	"strings"
+
+	"github.com/pdevine/tensor"
+	"github.com/pdevine/tensor/native"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type mistral3Model struct {
+	ModelParameters
+	ImageTokenIndex    uint32 `json:"image_token_index"`
+	SpatialMergeSize   uint32 `json:"spatial_merge_size"`
+	VisionFeatureLayer int32  `json:"vision_feature_layer"`
+	TextModel          struct {
+		NumHiddenLayers       uint32  `json:"num_hidden_layers"`
+		MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
+		HiddenSize            uint32  `json:"hidden_size"`
+		IntermediateSize      uint32  `json:"intermediate_size"`
+		NumAttentionHeads     uint32  `json:"num_attention_heads"`
+		NumKeyValueHeads      uint32  `json:"num_key_value_heads"`
+		RopeTheta             float32 `json:"rope_theta"`
+		RMSNormEPS            float32 `json:"rms_norm_eps"`
+		HeadDim               uint32  `json:"head_dim"`
+		SlidingWindow         *uint32 `json:"sliding_window"`
+		HiddenAct             string  `json:"hidden_act"`
+		VocabSize             uint32  `json:"vocab_size"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumAttentionHeads uint32  `json:"num_attention_heads"`
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`
+		HiddenSize        uint32  `json:"hidden_size"`
+		IntermediateSize  uint32  `json:"intermediate_size"`
+		ImageSize         uint32  `json:"image_size"`
+		NumChannels       uint32  `json:"num_channels"`
+		PatchSize         uint32  `json:"patch_size"`
+		HeadDim           uint32  `json:"head_dim"`
+		HiddenAct         string  `json:"hidden_act"`
+		RopeTheta         float32 `json:"rope_theta"`
+	} `json:"vision_config"`
+	MultiModalProjectorBias bool   `json:"multimodal_projector_bias"`
+	ProjectorHiddenAct      string `json:"projector_hidden_act"`
+}
+
+func (p *mistral3Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "mistral3"
+	kv["mistral3.vocab_size"] = p.TextModel.VocabSize
+
+	// Text configuration
+	kv["mistral3.block_count"] = p.TextModel.NumHiddenLayers
+	kv["mistral3.context_length"] = p.TextModel.MaxPositionEmbeddings
+	kv["mistral3.embedding_length"] = p.TextModel.HiddenSize
+	kv["mistral3.feed_forward_length"] = p.TextModel.IntermediateSize
+	kv["mistral3.attention.head_count"] = p.TextModel.NumAttentionHeads
+	kv["mistral3.attention.head_count_kv"] = p.TextModel.NumKeyValueHeads
+	kv["mistral3.attention.layer_norm_rms_epsilon"] = p.TextModel.RMSNormEPS
+	kv["mistral3.attention.key_length"] = p.TextModel.HeadDim
+	kv["mistral3.attention.value_length"] = p.TextModel.HeadDim
+	kv["mistral3.rope.dimension_count"] = p.TextModel.HiddenSize / p.TextModel.NumHiddenLayers
+	kv["mistral3.rope.freq_base"] = p.TextModel.RopeTheta
+
+	// Vision configuration
+	kv["mistral3.vision.block_count"] = p.VisionModel.NumHiddenLayers
+	kv["mistral3.vision.embedding_length"] = p.VisionModel.HiddenSize
+	kv["mistral3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+	kv["mistral3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+	kv["mistral3.vision.attention.key_length"] = p.VisionModel.HeadDim
+	kv["mistral3.vision.image_size"] = p.VisionModel.ImageSize
+	kv["mistral3.vision.patch_size"] = p.VisionModel.PatchSize
+	kv["mistral3.vision.num_channels"] = p.VisionModel.NumChannels
+	// kv["mistral3.vision.attention.layer_norm_epsilon"] = 1e-05 // Default value
+	kv["mistral3.vision.rope.freq_base"] = p.VisionModel.RopeTheta
+
+	// Multimodal configuration
+	kv["mistral3.image_token_index"] = p.ImageTokenIndex
+	kv["mistral3.spatial_merge_size"] = p.SpatialMergeSize
+
+	kv["mistral3.mm.projector_bias"] = p.MultiModalProjectorBias
+
+	if p.ProjectorHiddenAct != "" {
+		kv["mistral3.mm.projector_hidden_act"] = p.ProjectorHiddenAct
+	}
+
+	return kv
+}
+
+func (p *mistral3Model) Tensors(ts []Tensor) []ggml.Tensor {
+	var out []ggml.Tensor
+
+	for _, t := range ts {
+		if !strings.HasPrefix(t.Name(), "v.") {
+			if strings.HasSuffix(t.Name(), ".attn_q.weight") ||
+				strings.HasSuffix(t.Name(), ".attn_k.weight") {
+				t.SetRepacker(p.repack)
+			}
+		}
+
+		out = append(out, ggml.Tensor{
+			Name:     t.Name(),
+			Kind:     t.Kind(),
+			Shape:    t.Shape(),
+			WriterTo: t,
+		})
+	}
+
+	return out
+}
+
+func (p *mistral3Model) Replacements() []string {
+	return []string{
+		"language_model.model.norm", "output_norm",
+		"language_model.model.", "",
+		"language_model.", "",
+		"layers", "blk",
+		"transformer.layers", "blk",
+		"vision_tower", "v",
+		"ln_pre", "encoder_norm",
+		"input_layernorm", "attn_norm",
+		"post_attention_layernorm", "ffn_norm",
+		"embed_tokens", "token_embd",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"mlp.down_proj", "ffn_down",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.up_proj", "ffn_up",
+		"attention.q_proj", "attn_q",
+		"attention.k_proj", "attn_k",
+		"attention.v_proj", "attn_v",
+		"attention.o_proj", "attn_output",
+		"attention_norm", "attn_norm",
+		"feed_forward.gate_proj", "ffn_gate",
+		"feed_forward.down_proj", "ffn_down",
+		"feed_forward.up_proj", "ffn_up",
+		"multi_modal_projector", "mm",
+		"ffn_norm", "ffn_norm",
+		"lm_head", "output",
+	}
+}
+
+func (p *mistral3Model) repack(name string, data []float32, shape []uint64) ([]float32, error) {
+	var dims []int
+	for _, dim := range shape {
+		dims = append(dims, int(dim))
+	}
+
+	var heads uint32
+	if strings.HasSuffix(name, ".attn_q.weight") {
+		heads = p.TextModel.NumAttentionHeads
+	} else if strings.HasSuffix(name, ".attn_k.weight") {
+		heads = cmp.Or(p.TextModel.NumKeyValueHeads, p.TextModel.NumAttentionHeads)
+	} else {
+		return nil, fmt.Errorf("unknown tensor for repack: %s", name)
+	}
+
+	n := tensor.New(tensor.WithShape(dims...), tensor.WithBacking(data))
+	if err := n.Reshape(append([]int{int(heads), 2, dims[0] / int(heads) / 2}, dims[1:]...)...); err != nil {
+		return nil, err
+	}
+
+	if err := n.T(0, 2, 1, 3); err != nil {
+		return nil, err
+	}
+
+	if err := n.Reshape(dims...); err != nil {
+		return nil, err
+	}
+
+	if err := n.Transpose(); err != nil {
+		return nil, err
+	}
+
+	ts, err := native.SelectF32(n, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	var f32s []float32
+	for _, t := range ts {
+		f32s = append(f32s, t...)
+	}
+
+	return f32s, nil
+}

convert/convert_phi3.go

@@ -118,6 +118,5 @@ func (p *phi3Model) Replacements() []string {
 type ropeFactor []float32
 
 func (r ropeFactor) WriteTo(w io.Writer) (int64, error) {
-	err := binary.Write(w, binary.LittleEndian, r)
-	return 0, err
+	return 0, binary.Write(w, binary.LittleEndian, r)
 }

convert/convert_test.go

@@ -11,7 +11,6 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
-	"math"
 	"os"
 	"path/filepath"
 	"slices"
@@ -48,7 +47,7 @@ func convertFull(t *testing.T, fsys fs.FS) (*os.File, ggml.KV, ggml.Tensors) {
 	}
 	t.Cleanup(func() { r.Close() })
 
-	m, _, err := ggml.Decode(r, math.MaxInt)
+	m, _, err := ggml.Decode(r, -1)
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -332,7 +331,7 @@ func TestConvertAdapter(t *testing.T) {
 			}
 			defer r.Close()
 
-			m, _, err := ggml.Decode(r, math.MaxInt)
+			m, _, err := ggml.Decode(r, -1)
 			if err != nil {
 				t.Fatal(err)
 			}

convert/reader.go

@@ -11,14 +11,15 @@ type Tensor interface {
 	Name() string
 	Shape() []uint64
 	Kind() uint32
-	SetRepacker(repacker)
+	SetRepacker(Repacker)
 	WriteTo(io.Writer) (int64, error)
+	Clone() Tensor
 }
 
 type tensorBase struct {
-	name  string
-	shape []uint64
-	repacker
+	name     string
+	shape    []uint64
+	repacker Repacker
 }
 
 func (t tensorBase) Name() string {
@@ -36,7 +37,8 @@ const (
 
 func (t tensorBase) Kind() uint32 {
 	if strings.HasSuffix(t.name, ".ffn_gate_inp.weight") ||
-		t.name == "token_types.weight" {
+		t.name == "token_types.weight" ||
+		t.name == "v.positional_embedding_vlm" {
 		// these tensors are always F32
 		return 0
 	}
@@ -51,21 +53,18 @@ func (t tensorBase) Kind() uint32 {
 	}
 }
 
-func (t *tensorBase) SetRepacker(fn repacker) {
+func (t *tensorBase) SetRepacker(fn Repacker) {
 	t.repacker = fn
 }
 
-type repacker func(string, []float32, []uint64) ([]float32, error)
+type Repacker func(string, []float32, []uint64) ([]float32, error)
 
 func parseTensors(fsys fs.FS, replacer *strings.Replacer) ([]Tensor, error) {
 	patterns := []struct {
 		Pattern string
 		Func    func(fs.FS, *strings.Replacer, ...string) ([]Tensor, error)
 	}{
-		{"model-*-of-*.safetensors", parseSafetensors},
-		{"model.safetensors", parseSafetensors},
-		{"adapters.safetensors", parseSafetensors},
-		{"adapter_model.safetensors", parseSafetensors},
+		{"*.safetensors", parseSafetensors},
 		{"pytorch_model-*-of-*.bin", parseTorch},
 		{"pytorch_model.bin", parseTorch},
 		{"consolidated.*.pth", parseTorch},

convert/reader_safetensors.go

@@ -94,6 +94,21 @@ type safetensor struct {
 	*tensorBase
 }
 
+func (st safetensor) Clone() Tensor {
+	return &safetensor{
+		fs:     st.fs,
+		path:   st.path,
+		dtype:  st.dtype,
+		offset: st.offset,
+		size:   st.size,
+		tensorBase: &tensorBase{
+			name:     st.name,
+			repacker: st.repacker,
+			shape:    slices.Clone(st.shape),
+		},
+	}
+}
+
 func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 	f, err := st.fs.Open(st.path)
 	if err != nil {

convert/reader_torch.go

@@ -43,6 +43,17 @@ type torch struct {
 	*tensorBase
 }
 
+func (t torch) Clone() Tensor {
+	return torch{
+		storage: t.storage,
+		tensorBase: &tensorBase{
+			name:     t.name,
+			shape:    t.shape,
+			repacker: t.repacker,
+		},
+	}
+}
+
 func (pt torch) WriteTo(w io.Writer) (int64, error) {
 	return 0, nil
 }

convert/sentencepiece/sentencepiece_model.pb.go

@@ -1360,7 +1360,7 @@ func file_sentencepiece_model_proto_rawDescGZIP() []byte {
 
 var file_sentencepiece_model_proto_enumTypes = make([]protoimpl.EnumInfo, 2)
 var file_sentencepiece_model_proto_msgTypes = make([]protoimpl.MessageInfo, 6)
-var file_sentencepiece_model_proto_goTypes = []interface{}{
+var file_sentencepiece_model_proto_goTypes = []any{
 	(TrainerSpec_ModelType)(0),         // 0: sentencepiece.TrainerSpec.ModelType
 	(ModelProto_SentencePiece_Type)(0), // 1: sentencepiece.ModelProto.SentencePiece.Type
 	(*TrainerSpec)(nil),                // 2: sentencepiece.TrainerSpec
@@ -1392,7 +1392,7 @@ func file_sentencepiece_model_proto_init() {
 		return
 	}
 	if !protoimpl.UnsafeEnabled {
-		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[0].Exporter = func(v any, i int) any {
 			switch v := v.(*TrainerSpec); i {
 			case 0:
 				return &v.state
@@ -1406,7 +1406,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[1].Exporter = func(v any, i int) any {
 			switch v := v.(*NormalizerSpec); i {
 			case 0:
 				return &v.state
@@ -1420,7 +1420,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[2].Exporter = func(v any, i int) any {
 			switch v := v.(*SelfTestData); i {
 			case 0:
 				return &v.state
@@ -1434,7 +1434,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[3].Exporter = func(v any, i int) any {
 			switch v := v.(*ModelProto); i {
 			case 0:
 				return &v.state
@@ -1448,7 +1448,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[4].Exporter = func(v any, i int) any {
 			switch v := v.(*SelfTestData_Sample); i {
 			case 0:
 				return &v.state
@@ -1460,7 +1460,7 @@ func file_sentencepiece_model_proto_init() {
 				return nil
 			}
 		}
-		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v interface{}, i int) interface{} {
+		file_sentencepiece_model_proto_msgTypes[5].Exporter = func(v any, i int) any {
 			switch v := v.(*ModelProto_SentencePiece); i {
 			case 0:
 				return &v.state

convert/tokenizer_spm.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"log/slog"
 	"os"
+	"reflect"
 	"slices"
 
 	"google.golang.org/protobuf/proto"
@@ -15,6 +17,8 @@ import (
 )
 
 func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	slog.Debug("using spm vocabulary")
+
 	ast, err := parseAdditionalSpecialTokens(fsys)
 	if err != nil {
 		return nil, err
@@ -43,10 +47,19 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			v.Types = append(v.Types, int32(t))
 		default:
 			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
-			if slices.Contains(ast, piece.GetPiece()) {
+
+			// temporary fix to handle gemma3 broken configs
+			if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>"}, piece.GetPiece()) {
 				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
 			}
 
+			for _, t := range ast {
+				if t.Content == piece.GetPiece() {
+					tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+					break
+				}
+			}
+
 			v.Types = append(v.Types, tt)
 		}
 	}
@@ -78,10 +91,16 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 		return cmp.Compare(i.id, j.id)
 	})
 
-	n := len(v.Tokens)
-	for i, t := range ts {
-		if t.id != i+n {
-			return nil, fmt.Errorf("invalid token id: %d", t.id)
+	for _, t := range ts {
+		if t.id < len(v.Tokens) {
+			if v.Tokens[t.id] == t.content {
+				slog.Warn("tokenizer", "duplicate token", t.content, "id", t.id)
+				continue
+			}
+			return nil, fmt.Errorf("token mismatch: %s != %s at pos [%d]", t.content, v.Tokens[t.id], t.id)
+		}
+		if t.id != len(v.Tokens) {
+			return nil, fmt.Errorf("invalid token id: [%d] as pos [%d]", t.id, len(v.Tokens))
 		}
 
 		v.Tokens = append(v.Tokens, t.content)
@@ -92,7 +111,15 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 	return &v, nil
 }
 
-func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
+type specialToken struct {
+	Content    string `json:"content"`
+	Lstrip     bool   `json:"lstrip"`
+	Normalized bool   `json:"normalized"`
+	Rstrip     bool   `json:"rstrip"`
+	SingleWord bool   `json:"single_word"`
+}
+
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]specialToken, error) {
 	f, err := fsys.Open("special_tokens_map.json")
 	if errors.Is(err, os.ErrNotExist) {
 		return nil, nil
@@ -102,12 +129,43 @@ func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
 	defer f.Close()
 
 	var m struct {
-		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
+		AdditionalSpecialTokens any `json:"additional_special_tokens"`
 	}
 
 	if err := json.NewDecoder(f).Decode(&m); err != nil {
 		return nil, err
 	}
 
-	return m.AdditionalSpecialTokens, nil
+	var ast []specialToken
+
+	switch st := m.AdditionalSpecialTokens.(type) {
+	case []string:
+		for _, s := range st {
+			ast = append(ast, specialToken{Content: s})
+		}
+	case []any:
+		for _, s := range st {
+			// marshal and unmarshal the object to get the special token
+			tMap := s.(map[string]any)
+			data, err := json.Marshal(tMap)
+			if err != nil {
+				return nil, err
+			}
+
+			var token specialToken
+			err = json.Unmarshal(data, &token)
+			if err != nil {
+				return nil, err
+			}
+
+			ast = append(ast, token)
+		}
+
+	default:
+		slog.Warn("special token", "unknown token", reflect.TypeOf(st))
+	}
+
+	slog.Debug("spm tokenizer", "additional tokens", ast)
+
+	return ast, nil
 }

discover/cpu_common.go

@@ -12,7 +12,7 @@ func IsNUMA() bool {
 		// numa support in llama.cpp is linux only
 		return false
 	}
-	ids := map[string]interface{}{}
+	ids := map[string]any{}
 	packageIds, _ := filepath.Glob("/sys/devices/system/cpu/cpu*/topology/physical_package_id")
 	for _, packageId := range packageIds {
 		id, err := os.ReadFile(packageId)

discover/cuda_common.go

@@ -57,7 +57,8 @@ func cudaVariant(gpuInfo CudaGPUInfo) string {
 		}
 	}
 
-	if gpuInfo.computeMajor < 6 || gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
+	// driver 12.0 has problems with the cuda v12 library, so run v11 on those older drivers
+	if gpuInfo.DriverMajor < 12 || (gpuInfo.DriverMajor == 12 && gpuInfo.DriverMinor == 0) {
 		return "v11"
 	}
 	return "v12"

discover/gpu_linux.go

@@ -111,6 +111,7 @@ func GetCPUDetails() ([]CPU, error) {
 	if err != nil {
 		return nil, err
 	}
+	defer file.Close()
 	return linuxCPUDetails(file)
 }
 
@@ -168,13 +169,11 @@ func linuxCPUDetails(file io.Reader) ([]CPU, error) {
 	for id, s := range socketByID {
 		s.CoreCount = len(coreBySocket[id])
 		s.ThreadCount = 0
-		for _, tc := range threadsByCoreBySocket[id] {
-			s.ThreadCount += tc
-		}
 
 		// This only works if HT is enabled, consider a more reliable model, maybe cache size comparisons?
 		efficiencyCoreCount := 0
 		for _, threads := range threadsByCoreBySocket[id] {
+			s.ThreadCount += threads
 			if threads == 1 {
 				efficiencyCoreCount++
 			}

docs/api.md

@@ -173,7 +173,7 @@ curl http://localhost:11434/api/generate -d '{
 
 ##### Response
 
-```json
+```json5
 {
   "model": "codellama:code",
   "created_at": "2024-07-22T20:47:51.147561Z",
@@ -558,6 +558,10 @@ Final response:
 {
   "model": "llama3.2",
   "created_at": "2023-08-04T19:22:45.499127Z",
+  "message": {
+    "role": "assistant",
+    "content": ""
+  },
   "done": true,
   "total_duration": 4883583458,
   "load_duration": 1334875,
@@ -1213,13 +1217,13 @@ Show information about a model including details, modelfile, template, parameter
 
 ```shell
 curl http://localhost:11434/api/show -d '{
-  "model": "llama3.2"
+  "model": "llava"
 }'
 ```
 
 #### Response
 
-```json
+```json5
 {
   "modelfile": "# Modelfile generated by \"ollama show\"\n# To build a new Modelfile based on this one, replace the FROM line with:\n# FROM llava:latest\n\nFROM /Users/matt/.ollama/models/blobs/sha256:200765e1283640ffbd013184bf496e261032fa75b99498a9613be4e94d63ad52\nTEMPLATE \"\"\"{{ .System }}\nUSER: {{ .Prompt }}\nASSISTANT: \"\"\"\nPARAMETER num_ctx 4096\nPARAMETER stop \"\u003c/s\u003e\"\nPARAMETER stop \"USER:\"\nPARAMETER stop \"ASSISTANT:\"",
   "parameters": "num_keep                       24\nstop                           \"<|start_header_id|>\"\nstop                           \"<|end_header_id|>\"\nstop                           \"<|eot_id|>\"",
@@ -1256,7 +1260,11 @@ curl http://localhost:11434/api/show -d '{
     "tokenizer.ggml.pre": "llama-bpe",
     "tokenizer.ggml.token_type": [],        // populates if `verbose=true`
     "tokenizer.ggml.tokens": []             // populates if `verbose=true`
-  }
+  },
+  "capabilities": [
+    "completion",
+    "vision"
+  ],
 }
 ```
 

docs/benchmark.md

@@ -0,0 +1,59 @@
+# Benchmark
+
+Go benchmark tests that measure end-to-end performance of a running Ollama server. Run these tests to evaluate model inference performance on your hardware and measure the impact of code changes.
+
+## When to use
+
+Run these benchmarks when:
+- Making changes to the model inference engine
+- Modifying model loading/unloading logic
+- Changing prompt processing or token generation code
+- Implementing a new model architecture
+- Testing performance across different hardware setups
+
+## Prerequisites
+- Ollama server running locally with `ollama serve` on `127.0.0.1:11434`
+## Usage and Examples
+
+>[!NOTE]
+>All commands must be run from the root directory of the Ollama project.
+
+Basic syntax:
+```bash
+go test -bench=. ./benchmark/... -m $MODEL_NAME
+```
+
+Required flags:
+- `-bench=.`: Run all benchmarks
+- `-m`: Model name to benchmark
+
+Optional flags:
+- `-count N`: Number of times to run the benchmark (useful for statistical analysis)
+- `-timeout T`: Maximum time for the benchmark to run (e.g. "10m" for 10 minutes)
+
+Common usage patterns:
+
+Single benchmark run with a model specified:
+```bash
+go test -bench=. ./benchmark/... -m llama3.3
+```
+
+## Output metrics
+
+The benchmark reports several key metrics:
+
+- `gen_tok/s`: Generated tokens per second
+- `prompt_tok/s`: Prompt processing tokens per second
+- `ttft_ms`: Time to first token in milliseconds
+- `load_ms`: Model load time in milliseconds
+- `gen_tokens`: Total tokens generated
+- `prompt_tokens`: Total prompt tokens processed
+
+Each benchmark runs two scenarios:
+- Cold start: Model is loaded from disk for each test
+- Warm start: Model is pre-loaded in memory
+
+Three prompt lengths are tested for each scenario:
+- Short prompt (100 tokens)
+- Medium prompt (500 tokens)
+- Long prompt (1000 tokens)

docs/development.md

@@ -41,20 +41,11 @@ Install prerequisites:
 - [CMake](https://cmake.org/download/)
 - [Visual Studio 2022](https://visualstudio.microsoft.com/downloads/) including the Native Desktop Workload
 - (Optional) AMD GPU support
-    - [ROCm](https://rocm.github.io/install.html)
+    - [ROCm](https://rocm.docs.amd.com/en/latest/)
     - [Ninja](https://github.com/ninja-build/ninja/releases)
 - (Optional) NVIDIA GPU support
     - [CUDA SDK](https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network)
 
-> [!IMPORTANT]
-> Ensure prerequisites are in `PATH` before running CMake.
-
-> [!IMPORTANT]
-> ROCm is not compatible with Visual Studio CMake generators. Use `-GNinja` when configuring the project.
-
-> [!IMPORTANT]
-> CUDA is only compatible with Visual Studio CMake generators.
-
 Then, configure and build the project:
 
 ```shell
@@ -62,6 +53,14 @@ cmake -B build
 cmake --build build --config Release
 ```
 
+> [!IMPORTANT]
+> Building for ROCm requires additional flags:
+> ```
+> cmake -B build -G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++
+> cmake --build build --config Release
+> ```
+
+
 Lastly, run Ollama:
 
 ```shell
@@ -70,7 +69,7 @@ go run . serve
 
 ## Windows (ARM)
 
-Windows ARM does not support additional acceleration libraries at this time.
+Windows ARM does not support additional acceleration libraries at this time.  Do not use cmake, simply `go run` or `go build`.
 
 ## Linux
 
@@ -119,6 +118,35 @@ To run tests, use `go test`:
 go test ./...
 ```
 
+> NOTE: In rare cirumstances, you may nedd to change a package using the new
+> "synctest" package in go1.24.
+>
+> If you do not have the "synctest" package enabled, you will not see build or
+> test failures resulting from your change(s), if any, locally, but CI will
+> break.
+>
+> If you see failures in CI, you can either keep pushing changes to see if the
+> CI build passes, or you can enable the "synctest" package locally to see the
+> failures before pushing.
+>
+> To enable the "synctest" package for testing, run the following command:
+>
+> ```shell
+> GOEXPERIMENT=synctest go test ./...
+> ```
+>
+> If you wish to enable synctest for all go commands, you can set the
+> `GOEXPERIMENT` environment variable in your shell profile or by using:
+>
+> ```shell
+> go env -w GOEXPERIMENT=synctest
+> ```
+>
+> Which will enable the "synctest" package for all go commands without needing
+> to set it for all shell sessions.
+>
+> The synctest package is not required for production builds.
+
 ## Library detection
 
 Ollama looks for acceleration libraries in the following paths relative to the `ollama` executable:
@@ -128,4 +156,4 @@ Ollama looks for acceleration libraries in the following paths relative to the `
 * `.` (macOS)
 * `build/lib/ollama` (for development)
 
-If the libraries are not found, Ollama will not run with any acceleration libraries.
+If the libraries are not found, Ollama will not run with any acceleration libraries.

docs/faq.md

@@ -20,12 +20,18 @@ Please refer to the [GPU docs](./gpu.md).
 
 ## 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, unless you have a single GPU with <= 4 GB of VRAM, in which case it will default to 2048 tokens. 
+
+This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
+
+```shell
+OLLAMA_CONTEXT_LENGTH=8192 ollama serve
+```
 
 To change this when using `ollama run`, use `/set parameter`:
 
 ```shell
-/set parameter num_ctx 4096
+/set parameter num_ctx 8192
 ```
 
 When using the API, specify the `num_ctx` parameter:
@@ -35,7 +41,7 @@ curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
-    "num_ctx": 4096
+    "num_ctx": 8192
   }
 }'
 ```
@@ -187,6 +193,13 @@ cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11
 
 Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
 
+For browser extensions, you'll need to explicitly allow the extension's origin pattern. Set `OLLAMA_ORIGINS` to include `chrome-extension://*`, `moz-extension://*`, and `safari-web-extension://*` if you wish to allow all browser extensions access, or specific extensions as needed:
+
+```
+# Allow all Chrome, Firefox, and Safari extensions
+OLLAMA_ORIGINS=chrome-extension://*,moz-extension://*,safari-web-extension://* ollama serve
+```
+
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## Where are models stored?

docs/linux.md

@@ -75,7 +75,7 @@ RestartSec=3
 Environment="PATH=$PATH"
 
 [Install]
-WantedBy=default.target
+WantedBy=multi-user.target
 ```
 
 Then start the service:

docs/template.md

@@ -12,7 +12,7 @@ A basic Go template consists of three main parts:
 
 Here's an example of a simple chat template:
 
-```gotmpl
+```go
 {{- range .Messages }}
 {{ .Role }}: {{ .Content }}
 {{- end }}
@@ -162,6 +162,6 @@ CodeLlama [7B](https://ollama.com/library/codellama:7b-code) and [13B](https://o
 
 Codestral [22B](https://ollama.com/library/codestral:22b) supports fill-in-middle.
 
-```gotmpl
+```go
 [SUFFIX]{{ .Suffix }}[PREFIX] {{ .Prompt }}
 ```

docs/troubleshooting.md

@@ -9,7 +9,7 @@ cat ~/.ollama/logs/server.log
 On **Linux** systems with systemd, the logs can be found with this command:
 
 ```shell
-journalctl -u ollama --no-pager
+journalctl -u ollama --no-pager --follow --pager-end 
 ```
 
 When you run Ollama in a **container**, the logs go to stdout/stderr in the container:
@@ -26,7 +26,6 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
-- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
 
 To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
 
@@ -69,9 +68,9 @@ If you run into problems on Linux and want to install an older version, or you'd
 curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
 ```
 
-## Linux tmp noexec 
+## Linux docker
 
-If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
+If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.
 
 ## NVIDIA GPU Discovery
 
@@ -100,8 +99,6 @@ On linux, AMD GPU access typically requires `video` and/or `render` group member
 
 When running in a container, in some Linux distributions and container runtimes, the ollama process may be unable to access the GPU.  Use `ls -lnd /dev/kfd /dev/dri /dev/dri/*` on the host system to determine the **numeric** group IDs on your system, and pass additional `--group-add ...` arguments to the container so it can access the required devices.   For example, in the following output `crw-rw---- 1 0  44 226,   0 Sep 16 16:55 /dev/dri/card0` the group ID column is `44` 
 
-If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.
-
 If you are experiencing problems getting Ollama to correctly discover or use your GPU for inference, the following may help isolate the failure.
 - `AMD_LOG_LEVEL=3` Enable info log levels in the AMD HIP/ROCm libraries.  This can help show more detailed error codes that can help troubleshoot problems
 - `OLLAMA_DEBUG=1` During GPU discovery additional information will be reported

docs/windows.md

@@ -62,7 +62,6 @@ the explorer window by hitting `<Ctrl>+R` and type in:
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
-- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
 ## Uninstall
 
@@ -81,9 +80,11 @@ help you keep up to date.
 
 If you'd like to install or integrate Ollama as a service, a standalone
 `ollama-windows-amd64.zip` zip file is available containing only the Ollama CLI
-and GPU library dependencies for Nvidia and AMD. This allows for embedding
-Ollama in existing applications, or running it as a system service via `ollama
-serve` with tools such as [NSSM](https://nssm.cc/).
+and GPU library dependencies for Nvidia.  If you have an AMD GPU, also download
+and extract the additional ROCm package `ollama-windows-amd64-rocm.zip` into the
+same directory.  This allows for embedding Ollama in existing applications, or
+running it as a system service via `ollama serve` with tools such as
+[NSSM](https://nssm.cc/). 
 
 > [!NOTE]  
 > If you are upgrading from a prior version, you should remove the old directories first.

envconfig/config.go

@@ -53,8 +53,8 @@ func Host() *url.URL {
 	}
 }
 
-// Origins returns a list of allowed origins. Origins can be configured via the OLLAMA_ORIGINS environment variable.
-func Origins() (origins []string) {
+// AllowedOrigins returns a list of allowed origins. AllowedOrigins can be configured via the OLLAMA_ORIGINS environment variable.
+func AllowedOrigins() (origins []string) {
 	if s := Var("OLLAMA_ORIGINS"); s != "" {
 		origins = strings.Split(s, ",")
 	}
@@ -73,6 +73,7 @@ func Origins() (origins []string) {
 		"file://*",
 		"tauri://*",
 		"vscode-webview://*",
+		"vscode-file://*",
 	)
 
 	return origins
@@ -167,6 +168,8 @@ var (
 	MultiUserCache = Bool("OLLAMA_MULTIUSER_CACHE")
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
+	// ContextLength sets the default context length
+	ContextLength = Int64("OLLAMA_CONTEXT_LENGTH", -1)
 )
 
 func String(s string) func() string {
@@ -224,6 +227,20 @@ func Uint64(key string, defaultValue uint64) func() uint64 {
 	}
 }
 
+func Int64(key string, defaultValue int64) func() int64 {
+	return func() int64 {
+		if s := Var(key); s != "" {
+			if n, err := strconv.ParseInt(s, 10, 64); err != nil {
+				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
+			} else {
+				return n
+			}
+		}
+
+		return defaultValue
+	}
+}
+
 // Set aside VRAM per GPU
 var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
 
@@ -249,9 +266,10 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory(), "Do not preserve readline history"},
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune(), "Do not prune model blobs on startup"},
 		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel(), "Maximum number of parallel requests"},
-		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", Origins(), "A comma separated list of allowed origins"},
+		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default 4096 or 2048 with low VRAM)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational

envconfig/config_test.go

@@ -69,6 +69,7 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
+			"vscode-file://*",
 		}},
 		{"http://10.0.0.1", []string{
 			"http://10.0.0.1",
@@ -88,6 +89,7 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
+			"vscode-file://*",
 		}},
 		{"http://172.16.0.1,https://192.168.0.1", []string{
 			"http://172.16.0.1",
@@ -108,6 +110,7 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
+			"vscode-file://*",
 		}},
 		{"http://totally.safe,http://definitely.legit", []string{
 			"http://totally.safe",
@@ -128,13 +131,14 @@ func TestOrigins(t *testing.T) {
 			"file://*",
 			"tauri://*",
 			"vscode-webview://*",
+			"vscode-file://*",
 		}},
 	}
 	for _, tt := range cases {
 		t.Run(tt.value, func(t *testing.T) {
 			t.Setenv("OLLAMA_ORIGINS", tt.value)
 
-			if diff := cmp.Diff(Origins(), tt.expect); diff != "" {
+			if diff := cmp.Diff(AllowedOrigins(), tt.expect); diff != "" {
 				t.Errorf("%s: mismatch (-want +got):\n%s", tt.value, diff)
 			}
 		})
@@ -272,3 +276,19 @@ func TestVar(t *testing.T) {
 		})
 	}
 }
+
+func TestContextLength(t *testing.T) {
+	cases := map[string]int64{
+		"":     -1,
+		"4096": 4096,
+	}
+
+	for k, v := range cases {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv("OLLAMA_CONTEXT_LENGTH", k)
+			if i := ContextLength(); i != v {
+				t.Errorf("%s: expected %d, got %d", k, v, i)
+			}
+		})
+	}
+}

format/time_test.go

@@ -5,7 +5,7 @@ import (
 	"time"
 )
 
-func assertEqual(t *testing.T, a interface{}, b interface{}) {
+func assertEqual(t *testing.T, a any, b any) {
 	if a != b {
 		t.Errorf("Assert failed, expected %v, got %v", b, a)
 	}

fs/config.go

@@ -0,0 +1,13 @@
+package fs
+
+type Config interface {
+	Architecture() string
+	String(string, ...string) string
+	Uint(string, ...uint32) uint32
+	Float(string, ...float32) float32
+	Bool(string, ...bool) bool
+
+	Strings(string, ...[]string) []string
+	Ints(string, ...[]int32) []int32
+	Floats(string, ...[]float32) []float32
+}

fs/ggml/ggml.go

@@ -33,7 +33,7 @@ func (kv KV) Kind() string {
 }
 
 func (kv KV) ParameterCount() uint64 {
-	return keyValue[uint64](kv, "general.parameter_count")
+	return keyValue(kv, "general.parameter_count", uint64(0))
 }
 
 func (kv KV) FileType() fileType {
@@ -100,27 +100,47 @@ func (kv KV) Float(key string, defaultValue ...float32) float32 {
 	return keyValue(kv, key, append(defaultValue, 0)...)
 }
 
-func (kv KV) Strings(key string, defaultValue ...[]string) []string {
-	r := keyValue(kv, key, &array{})
-	s := make([]string, r.size)
-	for i := range r.size {
-		s[i] = r.values[i].(string)
-	}
+func (kv KV) Bool(key string, defaultValue ...bool) bool {
+	return keyValue(kv, key, append(defaultValue, false)...)
+}
 
-	return s
+func (kv KV) Strings(key string, defaultValue ...[]string) []string {
+	return keyValue(kv, key, &array[string]{values: append(defaultValue, []string(nil))[0]}).values
+}
+
+func (kv KV) Ints(key string, defaultValue ...[]int32) []int32 {
+	return keyValue(kv, key, &array[int32]{values: append(defaultValue, []int32(nil))[0]}).values
 }
 
 func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
-	r := keyValue(kv, key, &array{})
-	s := make([]uint32, r.size)
-	for i := range r.size {
-		s[i] = uint32(r.values[i].(int32))
-	}
-
-	return s
+	return keyValue(kv, key, &array[uint32]{values: append(defaultValue, []uint32(nil))[0]}).values
 }
 
-func keyValue[T string | uint32 | uint64 | float32 | *array](kv KV, key string, defaultValue ...T) T {
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	return keyValue(kv, key, &array[float32]{values: append(defaultValue, []float32(nil))[0]}).values
+}
+
+func (kv KV) OllamaEngineRequired() bool {
+	return slices.Contains([]string{
+		"gemma3",
+		"mistral3",
+		"llama4",
+	}, kv.Architecture())
+}
+
+type valueTypes interface {
+	uint8 | int8 | uint16 | int16 |
+		uint32 | int32 | uint64 | int64 |
+		string | float32 | float64 | bool
+}
+
+type arrayValueTypes interface {
+	*array[uint8] | *array[int8] | *array[uint16] | *array[int16] |
+		*array[uint32] | *array[int32] | *array[uint64] | *array[int64] |
+		*array[string] | *array[float32] | *array[float64] | *array[bool]
+}
+
+func keyValue[T valueTypes | arrayValueTypes](kv KV, key string, defaultValue ...T) T {
 	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
 		key = kv.Architecture() + "." + key
 	}
@@ -207,11 +227,26 @@ func (t Tensor) block() (n int) {
 
 func (t Tensor) blockSize() uint64 {
 	switch t.Kind {
-	case 0, 1, 24, 25, 26, 27, 28, 30: // F32, F16, I8, I16, I32, I64, F64, BF16
+	case
+		0,  // F32
+		1,  // F16
+		24, // I8
+		25, // I16
+		26, // I32
+		27, // I64
+		28, // F64
+		30: // BF16
 		return 1
-	case 2, 3, 4, 5, 6, 7, 8, 9, 20: // Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, Q8_1, IQ4_NL
+	case
+		2,  // Q4_0
+		3,  // Q4_1
+		6,  // Q5_0
+		7,  // Q5_1
+		8,  // Q8_0
+		9,  // Q8_1
+		20: // IQ4_NL
 		return 32
-	default: // All others
+	default:
 		return 256
 	}
 }
@@ -235,7 +270,7 @@ func (t Tensor) typeSize() uint64 {
 	case 8: // Q8_0
 		return 2 + blockSize
 	case 9: // Q8_1
-		return 4 + 4 + blockSize
+		return 2 + 2 + blockSize
 	case 10: // Q2_K
 		return blockSize/16 + blockSize/4 + 2 + 2
 	case 11: // Q3_K
@@ -247,7 +282,7 @@ func (t Tensor) typeSize() uint64 {
 	case 14: // Q6_K
 		return blockSize/2 + blockSize/4 + blockSize/16 + 2
 	case 15: // Q8_K
-		return 2 + blockSize + 2*blockSize/16
+		return 4 + blockSize + 2*blockSize/16
 	case 16: // IQ2_XXS
 		return 2 + 2*blockSize/8
 	case 17: // IQ2_XS
@@ -276,6 +311,8 @@ func (t Tensor) typeSize() uint64 {
 		return 8
 	case 29: // IQ1_M
 		return blockSize/8 + blockSize/16 + blockSize/32
+	case 30: // BF16
+		return 2
 	default:
 		return 0
 	}
@@ -293,6 +330,10 @@ func (t Tensor) Size() uint64 {
 	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
+func (t Tensor) Type() string {
+	return fileType(t.Kind).String()
+}
+
 type container interface {
 	Name() string
 	Decode(io.ReadSeeker) (model, error)
@@ -334,13 +375,8 @@ func DetectContentType(b []byte) string {
 // Decode decodes a GGML model from the given reader.
 //
 // It collects array values for arrays with a size less than or equal to
-// maxArraySize. If maxArraySize is 0, the default value of 1024 is used. If
-// the maxArraySize is negative, all arrays are collected.
+// maxArraySize. If the maxArraySize is negative, all arrays are collected.
 func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
-	if maxArraySize == 0 {
-		maxArraySize = 1024
-	}
-
 	rs = bufioutil.NewBufferedSeeker(rs, 32<<10)
 
 	var magic uint32
@@ -375,11 +411,11 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, int64, error) {
 	}, offset, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	embedding := f.KV().EmbeddingLength()
 	heads := f.KV().HeadCount()
 	headsKV := f.KV().HeadCountKV()
-	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array).size)
+	vocab := uint64(f.KV()["tokenizer.ggml.tokens"].(*array[string]).size)
 
 	embeddingHeads := f.KV().EmbeddingHeadCount()
 	embeddingHeadsK := f.KV().EmbeddingHeadCountK()
@@ -388,10 +424,13 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	layers := f.Tensors().GroupLayers()
 
 	bytesPerElement := kvCacheBytesPerElement(kvCacheType)
-	kv = uint64(float64(context*f.KV().BlockCount()*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	kv = make([]uint64, f.KV().BlockCount())
+	for i := range kv {
+		kv[i] = uint64(float64(context*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+	}
 
 	switch f.KV().Architecture() {
-	case "llama":
+	case "llama", "llama4":
 		fullOffload = max(
 			4*batch*(1+4*embedding+context*(1+heads)),
 			4*batch*(embedding+vocab),
@@ -405,7 +444,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 
 		if ffnGateExpsWeight, ok := layers["blk.0"]["ffn_gate_exps.weight"]; ok {
 			// mixtral 8x22b
-			ff := uint64(f.KV()["llama.feed_forward_length"].(uint32))
+			ff := uint64(f.KV().Uint("feed_forward_length"))
 			partialOffload = max(
 				3*ffnGateExpsWeight.Size()+4*batch*(2*ff+headsKV+embedding+context+embeddingHeads*headsKV),
 				4*(context*batch*heads+context*embeddingHeads*headsKV+batch*1024+embeddingHeads*headsKV*batch),
@@ -422,16 +461,14 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	case "mllama":
 		var visionTokens, tiles uint64 = 1601, 4
 
-		if crossAttentionLayers, ok := f.KV()["mllama.attention.cross_attention_layers"].(*array); ok {
-			kv = headsKV *
-				(embeddingHeadsK + embeddingHeadsV) * // one for K, one for V
-				(2* // sizeof(float16)
-					(f.KV().BlockCount()-uint64(crossAttentionLayers.size))* // num non-cross attention layers
-					context +
-					4* // sizeof(float32)
-						uint64(crossAttentionLayers.size)* // num cross attention layers
-						visionTokens*
-						tiles)
+		crossAttentionLayers := f.KV().Ints("attention.cross_attention_layers")
+		for i := range kv {
+			if slices.Contains(crossAttentionLayers, int32(i)) {
+				kv[i] = headsKV * (embeddingHeadsK + embeddingHeadsV) *
+					4 * // sizeof(float32)
+					visionTokens *
+					tiles
+			}
 		}
 
 		fullOffload = max(
@@ -455,7 +492,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 			// vocab graph
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
-	case "gemma", "gemma2":
+	case "gemma", "gemma2", "gemma3":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -467,6 +504,20 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 				4*embeddingHeadsK*context*8+
 				embedding*embeddingHeadsK*heads*9/16,
 		)
+
+		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
+		// engine. Gemma3 always uses the Ollama engine.
+		if f.KV().Architecture() == "gemma3" {
+			const gemma3GlobalCacheCount = 6
+			slidingWindow := (uint64(numParallel) * uint64(f.KV().Uint("attention.sliding_window"))) + batch
+			for i := range kv {
+				// Every 6th layer is a global layer, which is the full context size that has already been set. The other
+				// layers are the smaller local (sliding) layers.
+				if (i+1)%gemma3GlobalCacheCount != 0 {
+					kv[i] = uint64(float64(slidingWindow*(embeddingHeadsK+embeddingHeadsV)*headsKV) * bytesPerElement)
+				}
+			}
+		}
 	case "command-r":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
@@ -544,6 +595,59 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 	return
 }
 
+func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
+	if llm.KV().Uint("vision.block_count") == 0 {
+		return
+	}
+
+	for name, layer := range llm.Tensors().GroupLayers() {
+		if name == "v" || strings.HasPrefix(name, "v.") {
+			for _, tensor := range layer {
+				weights += tensor.Size()
+			}
+		}
+	}
+
+	imageSize := uint64(llm.KV().Uint("vision.image_size"))
+	patchSize := uint64(llm.KV().Uint("vision.patch_size"))
+	if patchSize == 0 {
+		slog.Warn("unknown patch size for vision model")
+		return
+	}
+
+	numChannels := uint64(llm.KV().Uint("vision.num_channels"))
+
+	numPatches := (imageSize / patchSize) * (imageSize / patchSize)
+	if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
+		numPatches++
+	}
+
+	headCount := uint64(llm.KV().Uint("vision.attention.head_count"))
+	embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
+
+	switch llm.KV().Architecture() {
+	case "mllama":
+		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
+
+		maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles"))
+
+		graphSize = 4 * (8 +
+			imageSize*imageSize*numChannels*maxNumTiles +
+			embeddingLength*numPatches*maxNumTiles +
+			9*embeddingLength*numPaddedPatches*maxNumTiles +
+			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
+	case "gemma3", "mistral3":
+		graphSize = 4 * (imageSize*imageSize*numChannels +
+			embeddingLength*patchSize +
+			numPatches*numPatches*headCount)
+	case "llama4":
+		// vision graph is computed independently in the same schedule
+		// and is negligible compared to the worst case text graph
+	}
+
+	return weights, graphSize
+}
+
 // SupportsKVCacheType checks if the requested cache type is supported
 func (f GGML) SupportsKVCacheType(cacheType string) bool {
 	return slices.Contains([]string{"f16", "q8_0", "q4_0"}, cacheType)

fs/ggml/ggml_test.go

@@ -2,7 +2,9 @@ package ggml
 
 import (
 	"maps"
+	"math"
 	"slices"
+	"strconv"
 	"strings"
 	"testing"
 
@@ -157,3 +159,113 @@ func TestTensorLayers(t *testing.T) {
 		})
 	}
 }
+
+// ref: https://github.com/ggml-org/llama.cpp/blob/a82c9e7c23ef6db48cebfa194dc9cebbc4ac3552/ggml/src/ggml.c#L572
+func TestTensorTypes(t *testing.T) {
+	cases := []struct {
+		kind      uint32
+		blockSize uint64
+		typeSize  uint64
+	}{
+		{0, 1, 4},
+		{1, 1, 2},
+		{2, 32, 18},
+		{3, 32, 20},
+		{6, 32, 22},
+		{7, 32, 24},
+		{8, 32, 34},
+		{9, 32, 36},
+		{10, 256, 84},
+		{11, 256, 110},
+		{12, 256, 144},
+		{13, 256, 176},
+		{14, 256, 210},
+		{15, 256, 292},
+		{16, 256, 66},
+		{17, 256, 74},
+		{18, 256, 98},
+		{19, 256, 50},
+		{20, 32, 18},
+		{21, 256, 110},
+		{22, 256, 82},
+		{23, 256, 136},
+		{24, 1, 1},
+		{25, 1, 2},
+		{26, 1, 4},
+		{27, 1, 8},
+		{28, 1, 8},
+		{29, 256, 56},
+		{30, 1, 2},
+	}
+
+	for _, tt := range cases {
+		t.Run(strconv.Itoa(int(tt.kind)), func(t *testing.T) {
+			tensor := Tensor{Kind: tt.kind}
+			if tensor.blockSize() != tt.blockSize {
+				t.Errorf("unexpected block size: got=%d want=%d", tensor.blockSize(), tt.blockSize)
+			}
+
+			if tensor.typeSize() != tt.typeSize {
+				t.Errorf("unexpected type size: got=%d want=%d", tensor.typeSize(), tt.typeSize)
+			}
+		})
+	}
+}
+
+func TestKeyValue(t *testing.T) {
+	kv := KV{
+		"general.architecture": "test",
+		"test.strings":         &array[string]{size: 3, values: []string{"a", "b", "c"}},
+		"test.float32s":        &array[float32]{size: 3, values: []float32{1.0, 2.0, 3.0}},
+		"test.int32s":          &array[int32]{size: 3, values: []int32{1, 2, 3}},
+		"test.uint32s":         &array[uint32]{size: 3, values: []uint32{1, 2, 3}},
+	}
+
+	if diff := cmp.Diff(kv.Strings("strings"), []string{"a", "b", "c"}); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Strings("nonexistent.strings"), []string(nil)); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Strings("default.strings", []string{"ollama"}), []string{"ollama"}); diff != "" {
+		t.Errorf("unexpected strings (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("float32s"), []float32{1.0, 2.0, 3.0}); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("nonexistent.float32s"), []float32(nil)); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Floats("default.float32s", []float32{math.MaxFloat32}), []float32{math.MaxFloat32}); diff != "" {
+		t.Errorf("unexpected float32s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("int32s"), []int32{1, 2, 3}); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("nonexistent.int32s"), []int32(nil)); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Ints("default.int32s", []int32{math.MaxInt32}), []int32{math.MaxInt32}); diff != "" {
+		t.Errorf("unexpected int8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("uint32s"), []uint32{1, 2, 3}); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("nonexistent.uint32s"), []uint32(nil)); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+
+	if diff := cmp.Diff(kv.Uints("default.uint32s", []uint32{math.MaxUint32}), []uint32{math.MaxUint32}); diff != "" {
+		t.Errorf("unexpected uint8s (-got +want):\n%s", diff)
+	}
+}

fs/ggml/gguf.go

@@ -36,10 +36,6 @@ type containerGGUF struct {
 	maxArraySize int
 }
 
-func (c *containerGGUF) canCollectArray(size int) bool {
-	return c.maxArraySize < 0 || size <= c.maxArraySize
-}
-
 func (c *containerGGUF) Name() string {
 	return "gguf"
 }
@@ -235,10 +231,7 @@ func (llm *gguf) Decode(rs io.ReadSeeker) error {
 	// patch KV with parameter count
 	llm.kv["general.parameter_count"] = llm.parameters
 
-	alignment, ok := llm.kv["general.alignment"].(uint32)
-	if !ok {
-		alignment = 32
-	}
+	alignment := llm.kv.Uint("general.alignment", 32)
 
 	offset, err := rs.Seek(0, io.SeekCurrent)
 	if err != nil {
@@ -298,6 +291,23 @@ func readGGUFV1String(llm *gguf, r io.Reader) (string, error) {
 	return b.String(), nil
 }
 
+func readGGUFV1StringsData(llm *gguf, r io.Reader, a *array[string]) (any, error) {
+	for i := range a.size {
+		if a.values != nil {
+			e, err := readGGUFV1String(llm, r)
+			if err != nil {
+				return nil, err
+			}
+
+			a.values[i] = e
+		} else {
+			discardGGUFString(llm, r)
+		}
+	}
+
+	return a, nil
+}
+
 func discardGGUFString(llm *gguf, r io.Reader) error {
 	buf := llm.scratch[:8]
 	_, err := io.ReadFull(r, buf)
@@ -355,78 +365,44 @@ func writeGGUFString(w io.Writer, s string) error {
 	return err
 }
 
-type array struct {
-	size   int
-	values []any
-}
-
-func (a *array) MarshalJSON() ([]byte, error) {
-	return json.Marshal(a.values)
-}
-
-func readGGUFV1Array(llm *gguf, r io.Reader) (*array, error) {
-	t, err := readGGUF[uint32](llm, r)
-	if err != nil {
-		return nil, err
-	}
-
-	n, err := readGGUF[uint32](llm, r)
-	if err != nil {
-		return nil, err
-	}
-
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, 0, int(n))
-	}
-
-	for i := range n {
-		var e any
-		switch t {
-		case ggufTypeUint8:
-			e, err = readGGUF[uint8](llm, r)
-		case ggufTypeInt8:
-			e, err = readGGUF[int8](llm, r)
-		case ggufTypeUint16:
-			e, err = readGGUF[uint16](llm, r)
-		case ggufTypeInt16:
-			e, err = readGGUF[int16](llm, r)
-		case ggufTypeUint32:
-			e, err = readGGUF[uint32](llm, r)
-		case ggufTypeInt32:
-			e, err = readGGUF[int32](llm, r)
-		case ggufTypeUint64:
-			e, err = readGGUF[uint64](llm, r)
-		case ggufTypeInt64:
-			e, err = readGGUF[int64](llm, r)
-		case ggufTypeFloat32:
-			e, err = readGGUF[float32](llm, r)
-		case ggufTypeFloat64:
-			e, err = readGGUF[float64](llm, r)
-		case ggufTypeBool:
-			e, err = readGGUF[bool](llm, r)
-		case ggufTypeString:
-			e, err = readGGUFV1String(llm, r)
-		default:
-			return nil, fmt.Errorf("invalid array type: %d", t)
-		}
-		if err != nil {
-			return nil, err
-		}
-
+func readGGUFStringsData(llm *gguf, r io.Reader, a *array[string]) (any, error) {
+	for i := range a.size {
 		if a.values != nil {
+			e, err := readGGUFString(llm, r)
+			if err != nil {
+				return nil, err
+			}
+
 			a.values[i] = e
+		} else {
+			discardGGUFString(llm, r)
 		}
 	}
 
 	return a, nil
 }
 
-func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
-	if llm.Version == 1 {
-		return readGGUFV1Array(llm, r)
-	}
+type array[T any] struct {
+	// size is the actual size of the array
+	size int
 
+	// values is the array of values. this is nil if the array is larger than configured maxSize
+	values []T
+}
+
+func (a *array[T]) MarshalJSON() ([]byte, error) {
+	return json.Marshal(a.values)
+}
+
+func newArray[T any](size, maxSize int) *array[T] {
+	a := array[T]{size: size}
+	if maxSize < 0 || size <= maxSize {
+		a.values = make([]T, size)
+	}
+	return &a
+}
+
+func readGGUFArray(llm *gguf, r io.Reader) (any, error) {
 	t, err := readGGUF[uint32](llm, r)
 	if err != nil {
 		return nil, err
@@ -437,45 +413,55 @@ func readGGUFArray(llm *gguf, r io.Reader) (*array, error) {
 		return nil, err
 	}
 
-	a := &array{size: int(n)}
-	if llm.canCollectArray(int(n)) {
-		a.values = make([]any, int(n))
-	}
-
-	for i := range n {
-		var e any
-		switch t {
-		case ggufTypeUint8:
-			e, err = readGGUF[uint8](llm, r)
-		case ggufTypeInt8:
-			e, err = readGGUF[int8](llm, r)
-		case ggufTypeUint16:
-			e, err = readGGUF[uint16](llm, r)
-		case ggufTypeInt16:
-			e, err = readGGUF[int16](llm, r)
-		case ggufTypeUint32:
-			e, err = readGGUF[uint32](llm, r)
-		case ggufTypeInt32:
-			e, err = readGGUF[int32](llm, r)
-		case ggufTypeUint64:
-			e, err = readGGUF[uint64](llm, r)
-		case ggufTypeInt64:
-			e, err = readGGUF[int64](llm, r)
-		case ggufTypeFloat32:
-			e, err = readGGUF[float32](llm, r)
-		case ggufTypeFloat64:
-			e, err = readGGUF[float64](llm, r)
-		case ggufTypeBool:
-			e, err = readGGUF[bool](llm, r)
-		case ggufTypeString:
-			if a.values != nil {
-				e, err = readGGUFString(llm, r)
-			} else {
-				err = discardGGUFString(llm, r)
-			}
-		default:
-			return nil, fmt.Errorf("invalid array type: %d", t)
+	switch t {
+	case ggufTypeUint8:
+		a := newArray[uint8](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt8:
+		a := newArray[int8](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint16:
+		a := newArray[uint16](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt16:
+		a := newArray[int16](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint32:
+		a := newArray[uint32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt32:
+		a := newArray[int32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeUint64:
+		a := newArray[uint64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeInt64:
+		a := newArray[int64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeFloat32:
+		a := newArray[float32](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeFloat64:
+		a := newArray[float64](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeBool:
+		a := newArray[bool](int(n), llm.maxArraySize)
+		return readGGUFArrayData(llm, r, a)
+	case ggufTypeString:
+		a := newArray[string](int(n), llm.maxArraySize)
+		if llm.Version == 1 {
+			return readGGUFV1StringsData(llm, r, a)
 		}
+
+		return readGGUFStringsData(llm, r, a)
+	default:
+		return nil, fmt.Errorf("invalid array type: %d", t)
+	}
+}
+
+func readGGUFArrayData[T any](llm *gguf, r io.Reader, a *array[T]) (any, error) {
+	for i := range a.size {
+		e, err := readGGUF[T](llm, r)
 		if err != nil {
 			return nil, err
 		}
@@ -506,6 +492,8 @@ func writeGGUFArray[S ~[]E, E any](w io.Writer, t uint32, s S) error {
 }
 
 func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
+	alignment := kv.Uint("general.alignment", 32)
+
 	if err := binary.Write(ws, binary.LittleEndian, []byte("GGUF")); err != nil {
 		return err
 	}
@@ -543,16 +531,15 @@ func WriteGGUF(ws io.WriteSeeker, kv KV, ts []Tensor) error {
 
 	var s uint64
 	for _, t := range ts {
-		t.Offset = s
+		t.Offset = s + uint64(ggufPadding(int64(s), int64(alignment)))
 		if err := ggufWriteTensorInfo(ws, t); err != nil {
 			return err
 		}
 		s += t.Size()
 	}
 
-	var alignment int64 = 32
 	for _, t := range ts {
-		if err := ggufWriteTensor(ws, t, alignment); err != nil {
+		if err := ggufWriteTensor(ws, t, int64(alignment)); err != nil {
 			return err
 		}
 	}
@@ -629,8 +616,8 @@ func ggufWriteTensorInfo(ws io.WriteSeeker, t Tensor) error {
 		return err
 	}
 
-	for i := range len(t.Shape) {
-		if err := binary.Write(ws, binary.LittleEndian, t.Shape[len(t.Shape)-i-1]); err != nil {
+	for _, n := range t.Shape {
+		if err := binary.Write(ws, binary.LittleEndian, n); err != nil {
 			return err
 		}
 	}

go.mod

@@ -1,6 +1,6 @@
 module github.com/ollama/ollama
 
-go 1.23.4
+go 1.24.0
 
 require (
 	github.com/containerd/console v1.0.3
@@ -11,7 +11,7 @@ require (
 	github.com/spf13/cobra v1.7.0
 	github.com/stretchr/testify v1.9.0
 	github.com/x448/float16 v0.8.4
-	golang.org/x/sync v0.10.0
+	golang.org/x/sync v0.11.0
 )
 
 require (
@@ -24,7 +24,7 @@ require (
 	github.com/nlpodyssey/gopickle v0.3.0
 	github.com/pdevine/tensor v0.0.0-20240510204454-f88f4562727c
 	golang.org/x/image v0.22.0
-	gonum.org/v1/gonum v0.15.0
+	golang.org/x/tools v0.30.0
 )
 
 require (
@@ -44,6 +44,7 @@ require (
 	github.com/xtgo/set v1.0.0 // indirect
 	go4.org/unsafe/assume-no-moving-gc v0.0.0-20231121144256-b99613f794b6 // indirect
 	golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1 // indirect
+	gonum.org/v1/gonum v0.15.0 // indirect
 	gorgonia.org/vecf32 v0.9.0 // indirect
 	gorgonia.org/vecf64 v0.9.0 // indirect
 )
@@ -69,12 +70,12 @@ require (
 	github.com/twitchyliquid64/golang-asm v0.15.1 // indirect
 	github.com/ugorji/go/codec v1.2.12 // indirect
 	golang.org/x/arch v0.8.0 // indirect
-	golang.org/x/crypto v0.31.0
-	golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa
-	golang.org/x/net v0.25.0 // indirect
-	golang.org/x/sys v0.28.0
-	golang.org/x/term v0.27.0
-	golang.org/x/text v0.21.0
+	golang.org/x/crypto v0.33.0
+	golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa
+	golang.org/x/net v0.35.0 // indirect
+	golang.org/x/sys v0.30.0
+	golang.org/x/term v0.29.0
+	golang.org/x/text v0.22.0
 	google.golang.org/protobuf v1.34.1
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -214,16 +214,16 @@ golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACk
 golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20191011191535-87dc89f01550/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI=
 golang.org/x/crypto v0.0.0-20200622213623-75b288015ac9/go.mod h1:LzIPMQfyMNhhGPhUkYOs5KpL4U8rLKemX1yGLhDgUto=
-golang.org/x/crypto v0.31.0 h1:ihbySMvVjLAeSH1IbfcRTkD/iNscyz8rGzjF/E5hV6U=
-golang.org/x/crypto v0.31.0/go.mod h1:kDsLvtWBEx7MV9tJOj9bnXsPbxwJQ6csT/x4KIN4Ssk=
+golang.org/x/crypto v0.33.0 h1:IOBPskki6Lysi0lo9qQvbxiQ+FvsCC/YWOecCHAixus=
+golang.org/x/crypto v0.33.0/go.mod h1:bVdXmD7IV/4GdElGPozy6U7lWdRXA4qyRVGJV57uQ5M=
 golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190125153040-c74c464bbbf2/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20190306152737-a1d7652674e8/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA=
 golang.org/x/exp v0.0.0-20191002040644-a1355ae1e2c3/go.mod h1:NOZ3BPKG0ec/BKJQgnvsSFpcKLM5xXVWnvZS97DWHgE=
-golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa h1:FRnLl4eNAQl8hwxVVC17teOw8kdjVDVAiFMtgUdTSRQ=
-golang.org/x/exp v0.0.0-20231110203233-9a3e6036ecaa/go.mod h1:zk2irFbV9DP96SEBUUAy67IdHUaZuSnrz1n472HUCLE=
+golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa h1:t2QcU6V556bFjYgu4L6C+6VrCPyJZ+eyRsABUPs1mz4=
+golang.org/x/exp v0.0.0-20250218142911-aa4b98e5adaa/go.mod h1:BHOTPb3L19zxehTsLoJXVaTktb06DFgmdW6Wb9s8jqk=
 golang.org/x/image v0.0.0-20180708004352-c73c2afc3b81/go.mod h1:ux5Hcp/YLpHSI86hEcLt0YII63i6oz57MZXIpbrjZUs=
 golang.org/x/image v0.0.0-20190227222117-0694c2d4d067/go.mod h1:kZ7UVZpmo3dzQBMxlp+ypCbDeSB+sBbTgSJuh5dn5js=
 golang.org/x/image v0.0.0-20190802002840-cff245a6509b/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=
@@ -257,8 +257,8 @@ golang.org/x/net v0.0.0-20200822124328-c89045814202/go.mod h1:/O7V0waA8r7cgGh81R
 golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU=
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.0.0-20210614182718-04defd469f4e/go.mod h1:9nx3DQGgdP8bBQD5qxJ1jj9UTztislL4KSBs9R2vV5Y=
-golang.org/x/net v0.25.0 h1:d/OCCoBEUq33pjydKrGQhw7IlUPI2Oylr+8qLx49kac=
-golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=
+golang.org/x/net v0.35.0 h1:T5GQRQb2y08kTAByq9L4/bz8cipCdA8FbRTXewonqY8=
+golang.org/x/net v0.35.0/go.mod h1:EglIi67kWsHKlRzzVMUD93VMSWGFOMSZgxFjparz1Qk=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.0.0-20200107190931-bf48bf16ab8d/go.mod h1:gOpvHmFTYa4IltrdGE7lF6nIHvwfUNPOp7c8zoXwtLw=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -268,8 +268,8 @@ golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJ
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20210220032951-036812b2e83c/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
-golang.org/x/sync v0.10.0 h1:3NQrjDixjgGwUOCaF8w2+VYHv0Ve/vGYSbdkTa98gmQ=
-golang.org/x/sync v0.10.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
+golang.org/x/sync v0.11.0 h1:GGz8+XQP4FvTTrjZPzNKTMFtSXH80RAzG+5ghFPgK9w=
+golang.org/x/sync v0.11.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=
 golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=
 golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=
@@ -285,17 +285,17 @@ golang.org/x/sys v0.0.0-20210510120138-977fb7262007/go.mod h1:oPkhp1MJrh7nUepCBc
 golang.org/x/sys v0.0.0-20210630005230-0f9fa26af87c/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
 golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=
-golang.org/x/sys v0.28.0 h1:Fksou7UEQUWlKvIdsqzJmUmCX3cZuD2+P3XyyzwMhlA=
-golang.org/x/sys v0.28.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
+golang.org/x/sys v0.30.0 h1:QjkSwP/36a20jFYWkSue1YwXzLmsV5Gfq7Eiy72C1uc=
+golang.org/x/sys v0.30.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
-golang.org/x/term v0.27.0 h1:WP60Sv1nlK1T6SupCHbXzSaN0b9wUmsPoRS9b61A23Q=
-golang.org/x/term v0.27.0/go.mod h1:iMsnZpn0cago0GOrHO2+Y7u7JPn5AylBrcoWkElMTSM=
+golang.org/x/term v0.29.0 h1:L6pJp37ocefwRRtYPKSWOWzOtWSxVajvz2ldH/xi3iU=
+golang.org/x/term v0.29.0/go.mod h1:6bl4lRlvVuDgSf3179VpIxBF0o10JUpXWOnI7nErv7s=
 golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=
 golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.5/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
 golang.org/x/text v0.3.6/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=
-golang.org/x/text v0.21.0 h1:zyQAAkrwaneQ066sspRyJaG9VNi/YJ1NfzcGB3hZ/qo=
-golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=
+golang.org/x/text v0.22.0 h1:bofq7m3/HAFvbF51jz3Q9wLg3jkvSPuiZu/pD1XwgtM=
+golang.org/x/text v0.22.0/go.mod h1:YRoo4H8PVmsu+E3Ou7cqLVH8oXWIHVoX0jqUWALQhfY=
 golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
 golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=
@@ -309,6 +309,8 @@ golang.org/x/tools v0.0.0-20200130002326-2f3ba24bd6e7/go.mod h1:TB2adYChydJhpapK
 golang.org/x/tools v0.0.0-20200619180055-7c47624df98f/go.mod h1:EkVYQZoAsY45+roYkvgYkIh4xh/qjgUK9TdY2XT94GE=
 golang.org/x/tools v0.0.0-20210106214847-113979e3529a/go.mod h1:emZCQorbCU4vsT4fOWvOPXz4eW1wZW4PmDk9uLelYpA=
 golang.org/x/tools v0.1.4/go.mod h1:o0xws9oXOQQZyjljx8fwUC0k7L1pTE6eaCbjGeHmOkk=
+golang.org/x/tools v0.30.0 h1:BgcpHewrV5AUp2G9MebG4XPFI1E2W41zU1SaqVA9vJY=
+golang.org/x/tools v0.30.0/go.mod h1:c347cR/OJfw5TI+GfX7RUPNMdDRRbjvYTS0jPyvsVtY=
 golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=
 golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

integration/api_test.go

@@ -0,0 +1,412 @@
+//go:build integration
+
+package integration
+
+import (
+	"bytes"
+	"context"
+	"fmt"
+	"math/rand"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestAPIGenerate(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.GenerateRequest{
+		Model:  smol,
+		Prompt: "why is the sky blue? be brief",
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	anyResp := []string{"rayleigh", "scattering"}
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tests := []struct {
+		name   string
+		stream bool
+	}{
+		{
+			name:   "stream",
+			stream: true,
+		},
+		{
+			name:   "no_stream",
+			stream: false,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			stallTimer := time.NewTimer(initialTimeout)
+			var buf bytes.Buffer
+			fn := func(response api.GenerateResponse) error {
+				// Fields that must always be present
+				if response.Model == "" {
+					t.Errorf("response missing model: %#v", response)
+				}
+				if response.Done {
+					// Required fields for final updates:
+					if response.DoneReason == "" && *req.Stream {
+						// TODO - is the lack of done reason on non-stream a bug?
+						t.Errorf("final response missing done_reason: %#v", response)
+					}
+					if response.Metrics.TotalDuration == 0 {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.Metrics.LoadDuration == 0 {
+						t.Errorf("final response missing load_duration: %#v", response)
+					}
+					if response.Metrics.PromptEvalDuration == 0 {
+						t.Errorf("final response missing prompt_eval_duration: %#v", response)
+					}
+					if response.Metrics.EvalCount == 0 {
+						t.Errorf("final response missing eval_count: %#v", response)
+					}
+					if response.Metrics.EvalDuration == 0 {
+						t.Errorf("final response missing eval_duration: %#v", response)
+					}
+					if len(response.Context) == 0 {
+						t.Errorf("final response missing context: %#v", response)
+					}
+
+					// Note: caching can result in no prompt eval count, so this can't be verified reliably
+					// if response.Metrics.PromptEvalCount == 0 {
+					// 	t.Errorf("final response missing prompt_eval_count: %#v", response)
+					// }
+
+				} // else incremental response, nothing to check right now...
+				buf.Write([]byte(response.Response))
+				if !stallTimer.Reset(streamTimeout) {
+					return fmt.Errorf("stall was detected while streaming response, aborting")
+				}
+				return nil
+			}
+
+			done := make(chan int)
+			var genErr error
+			go func() {
+				req.Stream = &test.stream
+				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
+				genErr = client.Generate(ctx, &req, fn)
+				done <- 0
+			}()
+
+			select {
+			case <-stallTimer.C:
+				if buf.Len() == 0 {
+					t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
+				} else {
+					t.Errorf("generate stalled.  Response so far:%s", buf.String())
+				}
+			case <-done:
+				if genErr != nil {
+					t.Fatalf("failed with %s request prompt %s ", req.Model, req.Prompt)
+				}
+				// Verify the response contains the expected data
+				response := buf.String()
+				atLeastOne := false
+				for _, resp := range anyResp {
+					if strings.Contains(strings.ToLower(response), resp) {
+						atLeastOne = true
+						break
+					}
+				}
+				if !atLeastOne {
+					t.Errorf("none of %v found in %s", anyResp, response)
+				}
+			case <-ctx.Done():
+				t.Error("outer test context done while waiting for generate")
+			}
+		})
+	}
+
+	// Validate PS while we're at it...
+	resp, err := client.ListRunning(ctx)
+	if err != nil {
+		t.Fatalf("list models API error: %s", err)
+	}
+	if resp == nil || len(resp.Models) == 0 {
+		t.Fatalf("list models API returned empty list while model should still be loaded")
+	}
+	// Find the model we just loaded and verify some attributes
+	found := false
+	for _, model := range resp.Models {
+		if strings.Contains(model.Name, req.Model) {
+			found = true
+			if model.Model == "" {
+				t.Errorf("model field omitted: %#v", model)
+			}
+			if model.Size == 0 {
+				t.Errorf("size omitted: %#v", model)
+			}
+			if model.Digest == "" {
+				t.Errorf("digest omitted: %#v", model)
+			}
+			verifyModelDetails(t, model.Details)
+			var nilTime time.Time
+			if model.ExpiresAt == nilTime {
+				t.Errorf("expires_at omitted: %#v", model)
+			}
+			// SizeVRAM could be zero.
+		}
+	}
+	if !found {
+		t.Errorf("unable to locate running model: %#v", resp)
+	}
+}
+
+func TestAPIChat(t *testing.T) {
+	initialTimeout := 60 * time.Second
+	streamTimeout := 30 * time.Second
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	// Set up the test data
+	req := api.ChatRequest{
+		Model: smol,
+		Messages: []api.Message{
+			{
+				Role:    "user",
+				Content: "why is the sky blue?  be brief",
+			},
+		},
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+	anyResp := []string{"rayleigh", "scattering"}
+
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	tests := []struct {
+		name   string
+		stream bool
+	}{
+		{
+			name:   "stream",
+			stream: true,
+		},
+		{
+			name:   "no_stream",
+			stream: false,
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			stallTimer := time.NewTimer(initialTimeout)
+			var buf bytes.Buffer
+			fn := func(response api.ChatResponse) error {
+				// Fields that must always be present
+				if response.Model == "" {
+					t.Errorf("response missing model: %#v", response)
+				}
+				if response.Done {
+					// Required fields for final updates:
+					var nilTime time.Time
+					if response.CreatedAt == nilTime {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.DoneReason == "" {
+						t.Errorf("final response missing done_reason: %#v", response)
+					}
+					if response.Metrics.TotalDuration == 0 {
+						t.Errorf("final response missing total_duration: %#v", response)
+					}
+					if response.Metrics.LoadDuration == 0 {
+						t.Errorf("final response missing load_duration: %#v", response)
+					}
+					if response.Metrics.PromptEvalDuration == 0 {
+						t.Errorf("final response missing prompt_eval_duration: %#v", response)
+					}
+					if response.Metrics.EvalCount == 0 {
+						t.Errorf("final response missing eval_count: %#v", response)
+					}
+					if response.Metrics.EvalDuration == 0 {
+						t.Errorf("final response missing eval_duration: %#v", response)
+					}
+
+					if response.Metrics.PromptEvalCount == 0 {
+						t.Errorf("final response missing prompt_eval_count: %#v", response)
+					}
+				} // else incremental response, nothing to check right now...
+				buf.Write([]byte(response.Message.Content))
+				if !stallTimer.Reset(streamTimeout) {
+					return fmt.Errorf("stall was detected while streaming response, aborting")
+				}
+				return nil
+			}
+
+			done := make(chan int)
+			var genErr error
+			go func() {
+				req.Stream = &test.stream
+				req.Options["seed"] = rand.Int() // bust cache for prompt eval results
+				genErr = client.Chat(ctx, &req, fn)
+				done <- 0
+			}()
+
+			select {
+			case <-stallTimer.C:
+				if buf.Len() == 0 {
+					t.Errorf("chat never started.  Timed out after :%s", initialTimeout.String())
+				} else {
+					t.Errorf("chat stalled.  Response so far:%s", buf.String())
+				}
+			case <-done:
+				if genErr != nil {
+					t.Fatalf("failed with %s request prompt %v", req.Model, req.Messages)
+				}
+				// Verify the response contains the expected data
+				response := buf.String()
+				atLeastOne := false
+				for _, resp := range anyResp {
+					if strings.Contains(strings.ToLower(response), resp) {
+						atLeastOne = true
+						break
+					}
+				}
+				if !atLeastOne {
+					t.Errorf("none of %v found in %s", anyResp, response)
+				}
+			case <-ctx.Done():
+				t.Error("outer test context done while waiting for chat")
+			}
+		})
+	}
+}
+
+func TestAPIListModels(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Make sure we have at least one model so an empty list can be considered a failure
+	if err := PullIfMissing(ctx, client, smol); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	resp, err := client.List(ctx)
+	if err != nil {
+		t.Fatalf("unable to list models: %s", err)
+	}
+	if len(resp.Models) == 0 {
+		t.Fatalf("list should not be empty")
+	}
+	model := resp.Models[0]
+	if model.Name == "" {
+		t.Errorf("first model name empty: %#v", model)
+	}
+	var nilTime time.Time
+	if model.ModifiedAt == nilTime {
+		t.Errorf("first model modified_at empty: %#v", model)
+	}
+	if model.Size == 0 {
+		t.Errorf("first model size empty: %#v", model)
+	}
+	if model.Digest == "" {
+		t.Errorf("first model digest empty: %#v", model)
+	}
+	verifyModelDetails(t, model.Details)
+}
+
+func verifyModelDetails(t *testing.T, details api.ModelDetails) {
+	if details.Format == "" {
+		t.Errorf("first model details.format empty: %#v", details)
+	}
+	if details.Family == "" {
+		t.Errorf("first model details.family empty: %#v", details)
+	}
+	if details.ParameterSize == "" {
+		t.Errorf("first model details.parameter_size empty: %#v", details)
+	}
+	if details.QuantizationLevel == "" {
+		t.Errorf("first model details.quantization_level empty: %#v", details)
+	}
+}
+
+func TestAPIShowModel(t *testing.T) {
+	modelName := "llama3.2"
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	if err := PullIfMissing(ctx, client, modelName); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+	resp, err := client.Show(ctx, &api.ShowRequest{Name: modelName})
+	if err != nil {
+		t.Fatalf("unable to show model: %s", err)
+	}
+	if resp.License == "" {
+		t.Errorf("%s missing license: %#v", modelName, resp)
+	}
+	if resp.Modelfile == "" {
+		t.Errorf("%s missing modelfile: %#v", modelName, resp)
+	}
+	if resp.Parameters == "" {
+		t.Errorf("%s missing parameters: %#v", modelName, resp)
+	}
+	if resp.Template == "" {
+		t.Errorf("%s missing template: %#v", modelName, resp)
+	}
+	// llama3 omits system
+	verifyModelDetails(t, resp.Details)
+	// llama3 ommits messages
+	if len(resp.ModelInfo) == 0 {
+		t.Errorf("%s missing model_info: %#v", modelName, resp)
+	}
+	// llama3 omits projectors
+	var nilTime time.Time
+	if resp.ModifiedAt == nilTime {
+		t.Errorf("%s missing modified_at: %#v", modelName, resp)
+	}
+}
+
+func TestAPIEmbeddings(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 1*time.Minute)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	req := api.EmbeddingRequest{
+		Model:  "orca-mini",
+		Prompt: "why is the sky blue?",
+		Options: map[string]interface{}{
+			"temperature": 0,
+			"seed":        123,
+		},
+	}
+
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("pull failed %s", err)
+	}
+
+	resp, err := client.Embeddings(ctx, &req)
+	if err != nil {
+		t.Fatalf("embeddings call failed %s", err)
+	}
+	if len(resp.Embedding) == 0 {
+		t.Errorf("zero length embedding response")
+	}
+}

integration/basic_test.go

@@ -14,15 +14,15 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestOrcaMiniBlueSky(t *testing.T) {
+func TestBlueSky(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -31,6 +31,7 @@ func TestOrcaMiniBlueSky(t *testing.T) {
 }
 
 func TestUnicode(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
 	ctx, cancel := context.WithTimeout(context.Background(), 3*time.Minute)
 	defer cancel()
 	// Set up the test data
@@ -39,7 +40,7 @@ func TestUnicode(t *testing.T) {
 		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
 		Prompt: "天空为什么是蓝色的?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			// Workaround deepseek context shifting bug
@@ -61,7 +62,7 @@ func TestExtendedUnicodeOutput(t *testing.T) {
 		Model:  "gemma2:2b",
 		Prompt: "Output some smily face emoji",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},
@@ -93,10 +94,10 @@ func TestUnicodeModelDir(t *testing.T) {
 	defer cancel()
 
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		Prompt: "why is the sky blue?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 		},

integration/concurrency_test.go

@@ -21,11 +21,11 @@ func TestMultiModelConcurrency(t *testing.T) {
 	var (
 		req = [2]api.GenerateRequest{
 			{
-				Model:     "orca-mini",
+				Model:     "llama3.2:1b",
 				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -34,7 +34,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -67,7 +67,7 @@ func TestMultiModelConcurrency(t *testing.T) {
 	wg.Wait()
 }
 
-func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
+func TestIntegrationConcurrentPredict(t *testing.T) {
 	req, resp := GenerateRequests()
 	reqLimit := len(req)
 	iterLimit := 5
@@ -117,6 +117,9 @@ func TestMultiModelStress(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
+	if maxVram < 2*format.GibiByte {
+		t.Skip("VRAM less than 2G, skipping model stress tests")
+	}
 
 	type model struct {
 		name string
@@ -125,8 +128,8 @@ func TestMultiModelStress(t *testing.T) {
 
 	smallModels := []model{
 		{
-			name: "orca-mini",
-			size: 2992 * format.MebiByte,
+			name: "llama3.2:1b",
+			size: 2876 * format.MebiByte,
 		},
 		{
 			name: "phi",

integration/context_test.go

@@ -23,7 +23,7 @@ func TestLongInputContext(t *testing.T) {
 		Model:  "llama2",
 		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			"num_ctx":     128,
@@ -50,7 +50,7 @@ func TestContextExhaustion(t *testing.T) {
 		Model:  "llama2",
 		Prompt: "Write me a story with a ton of emojis?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"temperature": 0,
 			"seed":        123,
 			"num_ctx":     128,

integration/llm_image_test.go

@@ -12,14 +12,63 @@ import (
 	"github.com/stretchr/testify/require"
 )
 
-func TestIntegrationLlava(t *testing.T) {
+func TestVisionModels(t *testing.T) {
+	skipUnderMinVRAM(t, 6)
+	type testCase struct {
+		model string
+	}
+	testCases := []testCase{
+		{
+			model: "llava:7b",
+		},
+		{
+			model: "llama3.2-vision",
+		},
+		{
+			model: "gemma3",
+		},
+	}
+
+	for _, v := range testCases {
+		t.Run(v.model, func(t *testing.T) {
+			image, err := base64.StdEncoding.DecodeString(imageEncoding)
+			require.NoError(t, err)
+			req := api.GenerateRequest{
+				Model:  v.model,
+				Prompt: "what does the text in this image say?",
+				Stream: &stream,
+				Options: map[string]any{
+					"seed":        42,
+					"temperature": 0.0,
+				},
+				Images: []api.ImageData{
+					image,
+				},
+			}
+			ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
+			defer cancel()
+			client, _, cleanup := InitServerConnection(ctx, t)
+
+			// Note: sometimes it returns "the ollamas" sometimes "the ollams"
+			resp := "the ollam"
+			defer cleanup()
+			require.NoError(t, PullIfMissing(ctx, client, req.Model))
+			// llava models on CPU can be quite slow to start
+			DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
+		})
+	}
+}
+
+func TestIntegrationSplitBatch(t *testing.T) {
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
 	require.NoError(t, err)
 	req := api.GenerateRequest{
-		Model:  "llava:7b",
+		Model: "gemma3:4b",
+		// Fill up a chunk of the batch so the image will partially spill over into the next one
+		System: "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed aliquet, justo in malesuada lobortis, odio ligula volutpat quam, quis faucibus ipsum magna quis sapien. Aliquam in venenatis diam, eu viverra magna. Phasellus imperdiet hendrerit volutpat. Vivamus sem ex, facilisis placerat felis non, dictum elementum est. Phasellus aliquam imperdiet lacus, eget placerat ligula sodales vel. Pellentesque nec auctor mi. Curabitur arcu nisi, faucibus eget nunc id, viverra interdum mi. Curabitur ornare ipsum ex, ac euismod ex aliquam in. Vestibulum id magna at purus accumsan fermentum. Proin scelerisque posuere nunc quis interdum. Maecenas sed mollis nisl. Etiam vitae ipsum interdum, placerat est quis, tincidunt velit. Nullam tempor nibh non lorem volutpat efficitur. Cras laoreet diam imperdiet ipsum auctor bibendum. Suspendisse ultrices urna sed metus sagittis suscipit. Quisque ullamcorper aliquam nibh ut mollis. Aenean dapibus mauris pharetra, venenatis elit ac, hendrerit odio. Cras vestibulum erat tempor, lobortis justo eu, lobortis ipsum. Nam laoreet dapibus sem. Proin vel diam ultrices, elementum ante et, ornare lectus. Proin eu accumsan nisl. Praesent ac ex vitae ipsum vulputate tristique facilisis sit amet lacus. Nullam faucibus magna a pellentesque pretium. Nunc lacinia ullamcorper sollicitudin. Donec vitae accumsan turpis, sed porttitor est. Donec porttitor mi vitae augue faucibus, vel mollis diam tincidunt.",
 		Prompt: "what does the text in this image say?",
 		Stream: &stream,
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},
@@ -39,33 +88,6 @@ func TestIntegrationLlava(t *testing.T) {
 	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
 }
 
-func TestIntegrationMllama(t *testing.T) {
-	image, err := base64.StdEncoding.DecodeString(imageEncoding)
-	require.NoError(t, err)
-	req := api.GenerateRequest{
-		// TODO fix up once we publish the final image
-		Model:  "x/llama3.2-vision",
-		Prompt: "what does the text in this image say?",
-		Stream: &stream,
-		Options: map[string]interface{}{
-			"seed":        42,
-			"temperature": 0.0,
-		},
-		Images: []api.ImageData{
-			image,
-		},
-	}
-
-	resp := "the ollamas"
-	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-	require.NoError(t, PullIfMissing(ctx, client, req.Model))
-	// mllama models on CPU can be quite slow to start,
-	DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
-}
-
 const imageEncoding = `iVBORw0KGgoAAAANSUhEUgAAANIAAAB4CAYAAACHHqzKAAAAAXNSR0IArs4c6QAAAIRlWElmTU0AKgAAAAgABQESAAMAAAABAAEAAAEaAAUAAAABAAAASgEb
 AAUAAAABAAAAUgEoAAMAAAABAAIAAIdpAAQAAAABAAAAWgAAAAAAAABIAAAAAQAAAEgAAAABAAOgAQADAAAAAQABAACgAgAEAAAAAQAAANKgAwAEAAAAAQAA
 AHgAAAAAXdsepgAAAAlwSFlzAAALEwAACxMBAJqcGAAAAVlpVFh0WE1MOmNvbS5hZG9iZS54bXAAAAAAADx4OnhtcG1ldGEgeG1sbnM6eD0iYWRvYmU6bnM6

integration/llm_test.go

@@ -17,30 +17,30 @@ var (
 	stream = false
 	req    = [2]api.GenerateRequest{
 		{
-			Model:  "orca-mini",
+			Model:  smol,
 			Prompt: "why is the ocean blue?",
 			Stream: &stream,
-			Options: map[string]interface{}{
+			Options: map[string]any{
 				"seed":        42,
 				"temperature": 0.0,
 			},
 		}, {
-			Model:  "orca-mini",
+			Model:  smol,
 			Prompt: "what is the origin of the us thanksgiving holiday?",
 			Stream: &stream,
-			Options: map[string]interface{}{
+			Options: map[string]any{
 				"seed":        42,
 				"temperature": 0.0,
 			},
 		},
 	}
 	resp = [2][]string{
-		{"sunlight"},
+		{"sunlight", "scattering", "interact"},
 		{"england", "english", "massachusetts", "pilgrims"},
 	}
 )
 
-func TestIntegrationSimpleOrcaMini(t *testing.T) {
+func TestIntegrationSimple(t *testing.T) {
 	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
 	defer cancel()
 	GenerateTestHelper(ctx, t, req[0], resp[0])

integration/max_queue_test.go

@@ -30,9 +30,9 @@ func TestMaxQueue(t *testing.T) {
 	t.Setenv("OLLAMA_MAX_QUEUE", strconv.Itoa(threadCount))
 
 	req := api.GenerateRequest{
-		Model:  "orca-mini",
+		Model:  smol,
 		Prompt: "write a long historical fiction story about christopher columbus.  use at least 10 facts from his actual journey",
-		Options: map[string]interface{}{
+		Options: map[string]any{
 			"seed":        42,
 			"temperature": 0.0,
 		},
@@ -52,8 +52,8 @@ func TestMaxQueue(t *testing.T) {
 	embedCtx := ctx
 
 	var genwg sync.WaitGroup
+	genwg.Add(1)
 	go func() {
-		genwg.Add(1)
 		defer genwg.Done()
 		slog.Info("Starting generate request")
 		DoGenerate(ctx, t, client, req, resp, 45*time.Second, 5*time.Second)
@@ -61,7 +61,7 @@ func TestMaxQueue(t *testing.T) {
 	}()
 
 	// Give the generate a chance to get started before we start hammering on embed requests
-	time.Sleep(5 * time.Millisecond)
+	time.Sleep(10 * time.Millisecond)
 
 	threadCount += 10 // Add a few extra to ensure we push the queue past its limit
 	busyCount := 0
@@ -71,8 +71,8 @@ func TestMaxQueue(t *testing.T) {
 	counterMu := sync.Mutex{}
 	var embedwg sync.WaitGroup
 	for i := 0; i < threadCount; i++ {
+		embedwg.Add(1)
 		go func(i int) {
-			embedwg.Add(1)
 			defer embedwg.Done()
 			slog.Info("embed started", "id", i)
 			embedReq := api.EmbeddingRequest{

integration/model_arch_test.go

@@ -0,0 +1,195 @@
+//go:build integration && models
+
+package integration
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io/ioutil"
+	"log/slog"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/format"
+)
+
+var (
+	started    = time.Now()
+	chatModels = []string{
+		"granite3-moe:latest",
+		"granite-code:latest",
+		"nemotron-mini:latest",
+		"command-r:latest",
+		"gemma2:latest",
+		"gemma:latest",
+		"internlm2:latest",
+		"phi3.5:latest",
+		"phi3:latest",
+		// "phi:latest", // flaky, sometimes generates no response on first query
+		"stablelm2:latest", // Predictions are off, crashes on small VRAM GPUs
+		"falcon:latest",
+		"falcon2:latest",
+		"minicpm-v:latest",
+		"mistral:latest",
+		"orca-mini:latest",
+		"llama2:latest",
+		"llama3.1:latest",
+		"llama3.2:latest",
+		"llama3.2-vision:latest",
+		"qwen2.5-coder:latest",
+		"qwen:latest",
+		"solar-pro:latest",
+	}
+)
+
+func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
+	deadline, hasDeadline := t.Deadline()
+	if !hasDeadline {
+		return 8 * time.Minute, 10 * time.Minute
+	} else if deadline.Compare(time.Now().Add(2*time.Minute)) <= 0 {
+		t.Skip("too little time")
+		return time.Duration(0), time.Duration(0)
+	}
+	return -time.Since(deadline.Add(-2 * time.Minute)), -time.Since(deadline.Add(-20 * time.Second))
+}
+
+func TestModelsGenerate(t *testing.T) {
+	softTimeout, hardTimeout := getTimeouts(t)
+	slog.Info("Setting timeouts", "soft", softTimeout, "hard", hardTimeout)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// TODO use info API eventually
+	var maxVram uint64
+	var err error
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err = strconv.ParseUint(s, 10, 64)
+		if err != nil {
+			t.Fatalf("invalid  OLLAMA_MAX_VRAM %v", err)
+		}
+	} else {
+		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
+	}
+
+	for _, model := range chatModels {
+		t.Run(model, func(t *testing.T) {
+			if time.Now().Sub(started) > softTimeout {
+				t.Skip("skipping remaining tests to avoid excessive runtime")
+			}
+			if err := PullIfMissing(ctx, client, model); err != nil {
+				t.Fatalf("pull failed %s", err)
+			}
+			if maxVram > 0 {
+				resp, err := client.List(ctx)
+				if err != nil {
+					t.Fatalf("list models failed %v", err)
+				}
+				for _, m := range resp.Models {
+					if m.Name == model && float32(m.Size)*1.2 > float32(maxVram) {
+						t.Skipf("model %s is too large for available VRAM: %s > %s", model, format.HumanBytes(m.Size), format.HumanBytes(int64(maxVram)))
+					}
+				}
+			}
+			// TODO - fiddle with context size
+			req := api.GenerateRequest{
+				Model:  model,
+				Prompt: "why is the sky blue?",
+				Options: map[string]interface{}{
+					"temperature": 0,
+					"seed":        123,
+				},
+			}
+			anyResp := []string{"rayleigh", "scattering", "atmosphere", "nitrogen", "oxygen"}
+			DoGenerate(ctx, t, client, req, anyResp, 120*time.Second, 30*time.Second)
+		})
+	}
+}
+
+func TestModelsEmbed(t *testing.T) {
+	softTimeout, hardTimeout := getTimeouts(t)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// TODO use info API eventually
+	var maxVram uint64
+	var err error
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err = strconv.ParseUint(s, 10, 64)
+		if err != nil {
+			t.Fatalf("invalid  OLLAMA_MAX_VRAM %v", err)
+		}
+	} else {
+		slog.Warn("No VRAM info available, testing all models, so larger ones might timeout...")
+	}
+
+	data, err := ioutil.ReadFile(filepath.Join("testdata", "embed.json"))
+	if err != nil {
+		t.Fatalf("failed to open test data file: %s", err)
+	}
+	testCase := map[string][]float64{}
+	err = json.Unmarshal(data, &testCase)
+	if err != nil {
+		t.Fatalf("failed to load test data: %s", err)
+	}
+	for model, expected := range testCase {
+
+		t.Run(model, func(t *testing.T) {
+			if time.Now().Sub(started) > softTimeout {
+				t.Skip("skipping remaining tests to avoid excessive runtime")
+			}
+			if err := PullIfMissing(ctx, client, model); err != nil {
+				t.Fatalf("pull failed %s", err)
+			}
+			if maxVram > 0 {
+				resp, err := client.List(ctx)
+				if err != nil {
+					t.Fatalf("list models failed %v", err)
+				}
+				for _, m := range resp.Models {
+					if m.Name == model && float32(m.Size)*1.2 > float32(maxVram) {
+						t.Skipf("model %s is too large for available VRAM: %s > %s", model, format.HumanBytes(m.Size), format.HumanBytes(int64(maxVram)))
+					}
+				}
+			}
+			req := api.EmbeddingRequest{
+				Model:  model,
+				Prompt: "why is the sky blue?",
+				Options: map[string]interface{}{
+					"temperature": 0,
+					"seed":        123,
+				},
+			}
+			resp, err := client.Embeddings(ctx, &req)
+			if err != nil {
+				t.Fatalf("embeddings call failed %s", err)
+			}
+			if len(resp.Embedding) == 0 {
+				t.Errorf("zero length embedding response")
+			}
+			if len(expected) != len(resp.Embedding) {
+				expStr := make([]string, len(resp.Embedding))
+				for i, v := range resp.Embedding {
+					expStr[i] = fmt.Sprintf("%0.6f", v)
+				}
+				// When adding new models, use this output to populate the testdata/embed.json
+				fmt.Printf("expected\n%s\n", strings.Join(expStr, ", "))
+				t.Fatalf("expected %d, got %d", len(expected), len(resp.Embedding))
+			}
+			sim := cosineSimilarity(resp.Embedding, expected)
+			if sim < 0.99 {
+				t.Fatalf("expected %v, got %v (similarity: %f)", expected[0:5], resp.Embedding[0:5], sim)
+			}
+		})
+	}
+
+}

integration/utils_test.go

@@ -24,9 +24,14 @@ import (
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/lifecycle"
+	"github.com/ollama/ollama/format"
 	"github.com/stretchr/testify/require"
 )
 
+const (
+	smol = "llama3.2:1b"
+)
+
 func Init() {
 	lifecycle.InitLogging()
 }
@@ -140,7 +145,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 
 	showCtx, cancel := context.WithDeadlineCause(
 		ctx,
-		time.Now().Add(10*time.Second),
+		time.Now().Add(20*time.Second),
 		fmt.Errorf("show for existing model %s took too long", modelName),
 	)
 	defer cancel()
@@ -157,7 +162,7 @@ func PullIfMissing(ctx context.Context, client *api.Client, modelName string) er
 	}
 	slog.Info("model missing", "model", modelName)
 
-	stallDuration := 30 * time.Second // This includes checksum verification, which can take a while on larger models
+	stallDuration := 60 * time.Second // This includes checksum verification, which can take a while on larger models, and slower systems
 	stallTimer := time.NewTimer(stallDuration)
 	fn := func(resp api.ProgressResponse) error {
 		// fmt.Print(".")
@@ -283,51 +288,51 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 }
 
 // Generate a set of requests
-// By default each request uses orca-mini as the model
+// By default each request uses llama3.2 as the model
 func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 	return []api.GenerateRequest{
 			{
-				Model:     "orca-mini",
+				Model:     smol,
 				Prompt:    "why is the ocean blue?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				Prompt:    "why is the color of dirt brown?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				Prompt:    "what is the origin of the us thanksgiving holiday?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				Prompt:    "what is the origin of independence day?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
 			}, {
-				Model:     "orca-mini",
+				Model:     smol,
 				Prompt:    "what is the composition of air?",
 				Stream:    &stream,
 				KeepAlive: &api.Duration{Duration: 10 * time.Second},
-				Options: map[string]interface{}{
+				Options: map[string]any{
 					"seed":        42,
 					"temperature": 0.0,
 				},
@@ -341,3 +346,15 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }
+
+func skipUnderMinVRAM(t *testing.T, gb uint64) {
+	// TODO use info API in the future
+	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
+		maxVram, err := strconv.ParseUint(s, 10, 64)
+		require.NoError(t, err)
+		// Don't hammer on small VRAM cards...
+		if maxVram < gb*format.GibiByte {
+			t.Skip("skipping with small VRAM to avoid timeouts")
+		}
+	}
+}

kvcache/cache.go

@@ -4,6 +4,7 @@ import (
 	"errors"
 
 	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
 )
 
 var (
@@ -29,22 +30,44 @@ type Cache interface {
 	// cache implementation used.
 	Put(ctx ml.Context, key, value ml.Tensor)
 
+	// SetConfig controls optimizations (mostly backend-specific) that may transform
+	// the output of the cache to work better with specific kernels. If not called,
+	// the backend settings will be used. This works well when calling Attention.
+	//
+	// The config can be overridden by models, especially if they require vanilla
+	// output when implementing their own version of attention. To do this, pass
+	// an empty ml.CacheConfig.
+	//
+	// Most models will not need to use this.
+	SetConfig(ml.CacheConfig)
+
 	// ** cache management **
 
-	// Init sets up runtime parameters
-	Init(backend ml.Backend, dtype ml.DType, capacity int32)
+	// Init sets up runtime parameters.
+	// backend: Used to allocate cache data storage and execute management operations (such as defrag)
+	// dtype: The data type for storing cache entries
+	// maxSequences: The maximum number of sequences stored in the cache - across all batches
+	// capacity: The number of cache entries to store, per sequence
+	// maxBatch: The maximum number of tokens that can occur in a single batch
+	Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int)
 
 	// Close closes the cache and frees resources associated with it
 	Close()
 
 	// StartForward is called before the start of the model's forward pass.
 	// For each token in the coming batch, there must be a corresponding
-	// entry in positions and seqs.
-	StartForward(ctx ml.Context, positions []int32, seqs []int) error
+	// entry in positions and seqs. reserve is to preallocate memory
+	// without actually storing data in the cache.
+	StartForward(ctx ml.Context, batch input.Batch, reserve bool) error
 
 	// CopyPrefix copies tokens in the range [0, len) from srcSeq to dstSeq
 	CopyPrefix(srcSeq, dstSeq int, len int32)
 
+	// CanResume returns true if the cache can continue with the next token at
+	// the given position and sequence. Assumes that the caller has already
+	// verified the contents of the cache.
+	CanResume(seq int, pos int32) bool
+
 	// Remove deletes tokens in the range [beginIndex, endIndex) from seq. Set
 	// endIndex to math.MaxInt32 to remove everything starting at beginIndex.
 	//

kvcache/causal.go

@@ -8,6 +8,7 @@ import (
 	"slices"
 
 	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
 )
 
 type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error)
@@ -19,8 +20,13 @@ type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, e
 // The mask is of shape history size, batch size
 type Causal struct {
 	DType      ml.DType
-	Capacity   int32
 	windowSize int32
+	chunkSize  int32
+
+	opts CausalOptions
+
+	// config controls mostly backend-specific optimizations
+	config *ml.CacheConfig
 
 	// ** current forward pass **
 
@@ -39,6 +45,12 @@ type Causal struct {
 	// locations in the cache that are needed for this batch
 	curCellRange cellRange
 
+	// curSequences is the sequences corresponding to this pass's entries in the cache
+	curSequences []int
+
+	// curPositions is the positions corresponding to this pass's entries in the cache
+	curPositions []int32
+
 	// ** cache metadata **
 
 	// for each possible location in the cache, stores the position and set of sequences
@@ -52,8 +64,8 @@ type Causal struct {
 
 	shiftFn      shiftFn
 	backend      ml.Backend
-	cacheCtx     ml.Context
-	keys, values []ml.Tensor
+	ctxs         map[int]ml.Context
+	keys, values map[int]ml.Tensor
 }
 
 type cacheCell struct {
@@ -67,67 +79,140 @@ type cellRange struct {
 }
 
 func NewCausalCache(shift shiftFn) *Causal {
-	return &Causal{windowSize: math.MaxInt32, shiftFn: shift}
+	return &Causal{
+		windowSize: math.MaxInt32,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
+	}
 }
 
 func NewSWACache(windowSize int32, shift shiftFn) *Causal {
-	return &Causal{windowSize: windowSize, shiftFn: shift}
+	return &Causal{
+		windowSize: windowSize,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
+	}
 }
 
-func (c *Causal) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
+func NewChunkedAttentionCache(chunkSize int32, shift shiftFn) *Causal {
+	return &Causal{
+		windowSize: math.MaxInt32,
+		chunkSize:  chunkSize,
+		shiftFn:    shift,
+		ctxs:       make(map[int]ml.Context),
+		keys:       make(map[int]ml.Tensor),
+		values:     make(map[int]ml.Tensor),
+	}
+}
+
+func (c *Causal) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
+	if c.config == nil {
+		var config ml.CacheConfig
+		if cc, ok := backend.(ml.BackendCacheConfig); ok {
+			config = cc.CacheConfig()
+		}
+		c.config = &config
+	}
+
+	if c.config.CachePadding == 0 {
+		c.config.CachePadding = 1
+	}
+
+	if c.config.MaskBatchPadding == 0 {
+		c.config.MaskBatchPadding = 1
+	}
+
+	if c.config.MaskDType == ml.DTypeOther {
+		c.config.MaskDType = ml.DTypeF32
+	}
+
+	var cacheSize int
+	if c.windowSize == math.MaxInt32 || capacity < int(c.windowSize) {
+		cacheSize = maxSequences * capacity
+	} else {
+		cacheSize = (maxSequences * int(c.windowSize)) + maxBatch
+	}
+	cacheSize = roundUp(cacheSize, c.config.CachePadding)
+	c.cells = make([]cacheCell, cacheSize)
+
 	c.DType = dtype
-	c.Capacity = capacity
-	c.cells = make([]cacheCell, capacity)
 	c.cellRanges = make(map[int]cellRange)
 	c.backend = backend
-	c.cacheCtx = backend.NewContext()
+}
+
+func (c *Causal) SetConfig(config ml.CacheConfig) {
+	if c.config != nil {
+		panic("config cannot be changed after being previously set, either by the model or backend")
+	}
+
+	c.config = &config
 }
 
 func (c *Causal) Close() {
-	c.cacheCtx.Close()
+	for _, ctx := range c.ctxs {
+		ctx.Close()
+	}
 }
 
-func (c *Causal) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
-	c.curBatchSize = len(positions)
+func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	c.curBatchSize = len(batch.Positions)
+	c.curSequences = batch.Sequences
+	c.curPositions = batch.Positions
+	c.opts.Except = nil
+
+	if !reserve {
+		c.updateSlidingWindow()
+
+		var err error
+		c.curLoc, err = c.findStartLoc()
+		if errors.Is(err, ErrKvCacheFull) {
+			c.defrag()
+			c.curLoc, err = c.findStartLoc()
+		}
+		if err != nil {
+			return err
+		}
+
+		c.curCellRange = newRange()
+		for i, pos := range batch.Positions {
+			seq := batch.Sequences[i]
+
+			c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
+
+			seqRange, ok := c.cellRanges[seq]
+			if !ok {
+				seqRange = newRange()
+			}
+
+			if c.curLoc+i > seqRange.max {
+				seqRange.max = c.curLoc + i
+			}
+			if seqRange.max > c.curCellRange.max {
+				c.curCellRange.max = seqRange.max
+			}
+
+			if c.curLoc+i < seqRange.min {
+				seqRange.min = c.curLoc + i
+			}
+			if seqRange.min < c.curCellRange.min {
+				c.curCellRange.min = seqRange.min
+			}
+			c.cellRanges[seq] = seqRange
+		}
+	} else {
+		// If we are reserving memory, don't update any of the cache metadata but set the size
+		// to the worst case.
+		c.curLoc = 0
+		c.curCellRange.min = 0
+		c.curCellRange.max = len(c.cells) - 1
+	}
 
 	var err error
-	c.curLoc, err = c.findStartLoc()
-	if errors.Is(err, ErrKvCacheFull) {
-		c.defrag()
-		c.curLoc, err = c.findStartLoc()
-	}
-	if err != nil {
-		return err
-	}
-
-	c.curCellRange = newRange()
-	for i, pos := range positions {
-		seq := seqs[i]
-
-		c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
-
-		seqRange, ok := c.cellRanges[seq]
-		if !ok {
-			seqRange = newRange()
-		}
-
-		if c.curLoc+i > seqRange.max {
-			seqRange.max = c.curLoc + i
-		}
-		if seqRange.max > c.curCellRange.max {
-			c.curCellRange.max = seqRange.max
-		}
-
-		if c.curLoc+i < seqRange.min {
-			seqRange.min = c.curLoc + i
-		}
-		if seqRange.min < c.curCellRange.min {
-			c.curCellRange.min = seqRange.min
-		}
-		c.cellRanges[seq] = seqRange
-	}
-
-	c.curMask, err = c.buildMask(ctx, positions, seqs)
+	c.curMask, err = c.buildMask(ctx)
 
 	return err
 }
@@ -154,39 +239,139 @@ func (c *Causal) findStartLoc() (int, error) {
 		}
 	}
 
-	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, c.Capacity)
+	return 0, fmt.Errorf("%w (length: %v)", ErrKvCacheFull, len(c.cells))
+}
+
+func (c *Causal) updateSlidingWindow() {
+	if c.windowSize == math.MaxInt32 {
+		return
+	}
+
+	// create a map of unique sequences to the lowest position in that sequence
+	lowestPos := make(map[int]int32)
+	for i := range c.curPositions {
+		seq := c.curSequences[i]
+
+		pos, ok := lowestPos[seq]
+		if !ok {
+			pos = c.curPositions[i]
+		} else if c.curPositions[i] < pos {
+			pos = c.curPositions[i]
+		}
+
+		lowestPos[seq] = pos
+	}
+
+	// delete any entries that are beyond the window of the oldest position in the sequence
+	for seq, pos := range lowestPos {
+		oldRange, ok := c.cellRanges[seq]
+		if !ok {
+			continue
+		}
+
+		newRange := newRange()
+
+		for i := oldRange.min; i <= oldRange.max; i++ {
+			if slices.Contains(c.cells[i].sequences, seq) {
+				if c.cells[i].pos < pos-c.windowSize {
+					c.cells[i].sequences = slices.DeleteFunc(c.cells[i].sequences, func(s int) bool { return s == seq })
+				} else {
+					newRange.min = min(newRange.min, i)
+					newRange.max = max(newRange.max, i)
+				}
+			}
+		}
+
+		c.cellRanges[seq] = newRange
+	}
+}
+
+func roundDown(length, pad int) int {
+	return (length / pad) * pad
+}
+
+func roundUp(length, pad int) int {
+	return ((length + pad - 1) / pad) * pad
 }
 
 // Builds a mask of history x batch indicating whether for each token in the batch the
 // token in the history should apply. This is based on both the sequence and causality (the
 // position of the history is not ahead of the token in the batch).
-func (c *Causal) buildMask(ctx ml.Context, positions []int32, seqs []int) (ml.Tensor, error) {
-	// TODO(jessegross): This does not do padding, which is required for flash attention
-	len := c.curCellRange.max - c.curCellRange.min + 1
-	mask := make([]float32, c.curBatchSize*len)
+func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
+	// Align and pad the two dimensions as required by the backend
+	batchSize := roundUp(c.curBatchSize, c.config.MaskBatchPadding)
+
+	c.curCellRange.min = roundDown(c.curCellRange.min, c.config.CachePadding)
+	c.curCellRange.max = roundUp(c.curCellRange.max+1, c.config.CachePadding) - 1
+
+	length := c.curCellRange.max - c.curCellRange.min + 1
+	mask := make([]float32, batchSize*length)
 
 	for i := range c.curBatchSize {
+		enabled := !slices.Contains(c.opts.Except, i)
 		for j := c.curCellRange.min; j <= c.curCellRange.max; j++ {
-			if !slices.Contains(c.cells[j].sequences, seqs[i]) || c.cells[j].pos > positions[i] ||
-				c.cells[j].pos < positions[i]-c.windowSize {
-				mask[i*len+(j-c.curCellRange.min)] = float32(math.Inf(-1))
+			if !slices.Contains(c.cells[j].sequences, c.curSequences[i]) ||
+				(enabled && c.cells[j].pos > c.curPositions[i]) ||
+				c.chunkSize > 0 && c.cells[j].pos < c.curPositions[i]-c.curPositions[i]%c.chunkSize ||
+				c.cells[j].pos < c.curPositions[i]-c.windowSize {
+				mask[i*length+(j-c.curCellRange.min)] = float32(math.Inf(-1))
 			}
 		}
 	}
 
-	return ctx.FromFloatSlice(mask, len, c.curBatchSize)
+	// Mask out any padding tokens we added. For padding that we added to the cache history, this
+	// has already been masked out because the sequence doesn't match.
+	for i := c.curBatchSize * length; i < len(mask); i++ {
+		mask[i] = float32(math.Inf(-1))
+	}
+
+	maskTensor, err := ctx.Input().FromFloatSlice(mask, length, batchSize)
+	if err != nil {
+		return nil, err
+	}
+
+	if c.config.MaskDType != ml.DTypeF32 {
+		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
+		ctx.Forward(maskTensor.Copy(ctx, out))
+		maskTensor = out
+	}
+
+	return maskTensor, nil
 }
 
-func moveCell(ctx ml.Context, objs []ml.Tensor, src, dst, len int) {
-	for _, obj := range objs {
-		if obj == nil {
+func (c *Causal) moveCells(ctx ml.Context, src, dst, length int) {
+	for i, key := range c.keys {
+		if key == nil {
 			continue
 		}
 
-		srcView := obj.View(ctx, obj.Stride(2)*src, obj.Dim(0)*obj.Dim(1)*len)
-		dstView := obj.View(ctx, obj.Stride(2)*dst, obj.Dim(0)*obj.Dim(1)*len)
+		kHeadDim := key.Dim(0)
+		numKVHeads := key.Dim(1)
+		rowSize := key.Stride(2)
 
-		ctx.Forward(srcView.Copy(ctx, dstView))
+		kSrcView := key.View(ctx, rowSize*src, kHeadDim*numKVHeads*length)
+		kDstView := key.View(ctx, rowSize*dst, kHeadDim*numKVHeads*length)
+
+		value := c.values[i]
+		var vSrcView, vDstView ml.Tensor
+		if c.config.PermutedV {
+			vHeadDim := value.Dim(1)
+			elemSize := value.Stride(0)
+
+			vSrcView = value.View(ctx, elemSize*src, length, len(c.cells)*elemSize, vHeadDim*numKVHeads)
+			vDstView = value.View(ctx, elemSize*dst, length, len(c.cells)*elemSize, vHeadDim*numKVHeads)
+		} else {
+			vHeadDim := value.Dim(0)
+			rowSize := value.Stride(2)
+
+			vSrcView = value.View(ctx, rowSize*src, vHeadDim*numKVHeads*length)
+			vDstView = value.View(ctx, rowSize*dst, vHeadDim*numKVHeads*length)
+		}
+
+		ctx.Forward(
+			kSrcView.Copy(ctx, kDstView),
+			vSrcView.Copy(ctx, vDstView),
+		)
 	}
 }
 
@@ -210,7 +395,8 @@ func (c *Causal) defrag() {
 	ctx := c.backend.NewContext()
 
 	// For every move, 6 tensors are required per layer (2 views and a
-	// copy for each of k and v).
+	// copy for each of k and v). We also need to refer to the original
+	// k and v cache tensors - once per layer, not per move.
 	layers := 0
 	for _, key := range c.keys {
 		if key == nil {
@@ -219,7 +405,7 @@ func (c *Causal) defrag() {
 		layers++
 	}
 
-	maxMoves := ctx.MaxTensors() / (6 * layers)
+	maxMoves := (ctx.MaxGraphNodes() - 2*layers) / (6 * layers)
 	moves := 0
 
 	var pendingSrc, pendingDst, pendingLen int
@@ -238,8 +424,7 @@ func (c *Causal) defrag() {
 							pendingLen++
 							break
 						} else {
-							moveCell(ctx, c.keys, pendingSrc, pendingDst, pendingLen)
-							moveCell(ctx, c.values, pendingSrc, pendingDst, pendingLen)
+							c.moveCells(ctx, pendingSrc, pendingDst, pendingLen)
 							moves++
 						}
 					}
@@ -263,8 +448,7 @@ func (c *Causal) defrag() {
 	}
 
 	if pendingLen > 0 {
-		moveCell(ctx, c.keys, pendingSrc, pendingDst, pendingLen)
-		moveCell(ctx, c.values, pendingSrc, pendingDst, pendingLen)
+		c.moveCells(ctx, pendingSrc, pendingDst, pendingLen)
 		moves++
 	}
 
@@ -293,45 +477,107 @@ func (c *Causal) defrag() {
 }
 
 func (c *Causal) SetLayer(layer int) {
-	if layer >= len(c.keys) {
-		c.keys = append(c.keys, make([]ml.Tensor, layer-len(c.keys)+1)...)
-		c.values = append(c.values, make([]ml.Tensor, layer-len(c.values)+1)...)
-	}
-
 	c.curLayer = layer
 }
 
+type CausalOptions struct {
+	// Enabled controls whether the causal mask is generated for a particular index in a batch
+	Except []int
+}
+
+// SetCausal disables causal mask generation for a particular range of indicies in
+// the current batch for subsequent calls to Get. The state resets for the next forward pass.
+func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
+	if !slices.Equal(c.opts.Except, opts.Except) {
+		c.opts = opts
+		if ctx != nil {
+			var err error
+			c.curMask, err = c.buildMask(ctx)
+			if err != nil {
+				// This error should never occur because we have previously built a mask with the same shape
+				panic(fmt.Errorf("SetCausal: %w", err))
+			}
+		}
+	}
+}
+
 func (c *Causal) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
 	key := c.keys[c.curLayer]
 	value := c.values[c.curLayer]
 
-	key = key.View(ctx, key.Stride(2)*c.curCellRange.min,
-		key.Dim(0), key.Stride(1),
-		key.Dim(1), key.Stride(2),
-		c.curMask.Dim(0),
+	kHeadDim := key.Dim(0)
+	numKVHeads := key.Dim(1)
+	rowSize := key.Stride(2)
+	cachedSize := c.curMask.Dim(0)
+
+	key = key.View(ctx, rowSize*c.curCellRange.min,
+		kHeadDim, key.Stride(1),
+		numKVHeads, key.Stride(2),
+		cachedSize,
 	)
 
-	value = value.View(ctx, key.Stride(2)*c.curCellRange.min,
-		value.Dim(0), value.Stride(1),
-		value.Dim(1), value.Stride(2),
-		c.curMask.Dim(0),
-	)
+	if c.config.PermutedV {
+		vHeadDim := value.Dim(1)
+		elemSize := value.Stride(0)
+
+		value = value.View(ctx, elemSize*c.curCellRange.min,
+			cachedSize, value.Stride(1),
+			vHeadDim, value.Stride(2),
+			numKVHeads,
+		)
+	} else {
+		vHeadDim := value.Dim(0)
+		rowSize := value.Stride(2)
+
+		value = value.View(ctx, rowSize*c.curCellRange.min,
+			vHeadDim, value.Stride(1),
+			numKVHeads, value.Stride(2),
+			cachedSize,
+		)
+	}
 
 	return key, value, c.curMask
 }
 
 func (c *Causal) Put(ctx ml.Context, key, value ml.Tensor) {
-	if c.curBatchSize != key.Dim(2) {
-		panic(fmt.Errorf("inconsistent batch sizes (layer: %v, batch size: %v layer batch size: %v)", c.curLayer, c.curBatchSize, key.Dim(2)))
+	kHeadDim := key.Dim(0)
+	vHeadDim := value.Dim(0)
+	numKVHeads := key.Dim(1)
+	batchSize := key.Dim(2)
+
+	if c.curBatchSize != batchSize {
+		panic(fmt.Errorf("inconsistent batch sizes (layer: %v, batch size: %v layer batch size: %v)", c.curLayer, c.curBatchSize, batchSize))
 	}
 
-	if c.keys[c.curLayer] == nil || c.values[c.curLayer] == nil {
-		c.keys[c.curLayer] = c.cacheCtx.Zeros(c.DType, key.Dim(0), key.Dim(1), int(c.Capacity))
-		c.values[c.curLayer] = c.cacheCtx.Zeros(c.DType, value.Dim(0), value.Dim(1), int(c.Capacity))
+	if _, ok := c.ctxs[c.curLayer]; !ok {
+		c.ctxs[c.curLayer] = c.backend.NewContextSize(2).Layer(c.curLayer)
 	}
 
-	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer].View(ctx, c.keys[c.curLayer].Stride(2)*c.curLoc, key.Dim(0)*key.Dim(1)*key.Dim(2))))
-	ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, c.values[c.curLayer].Stride(2)*c.curLoc, value.Dim(0)*value.Dim(1)*value.Dim(2))))
+	if _, ok := c.keys[c.curLayer]; !ok {
+		c.keys[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, kHeadDim, numKVHeads, len(c.cells))
+	}
+
+	if _, ok := c.values[c.curLayer]; !ok {
+		if c.config.PermutedV {
+			c.values[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, len(c.cells), vHeadDim, numKVHeads)
+		} else {
+			c.values[c.curLayer] = c.ctxs[c.curLayer].Zeros(c.DType, vHeadDim, numKVHeads, len(c.cells))
+		}
+	}
+
+	rowSize := c.keys[c.curLayer].Stride(2)
+	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer].View(ctx, rowSize*c.curLoc, kHeadDim*numKVHeads*batchSize)))
+
+	if c.config.PermutedV {
+		elemSize := c.values[c.curLayer].Stride(0)
+
+		value = value.Permute(ctx, 1, 2, 0, 3)
+		ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, elemSize*c.curLoc, batchSize, len(c.cells)*elemSize, vHeadDim*numKVHeads)))
+	} else {
+		rowSize := c.values[c.curLayer].Stride(2)
+
+		ctx.Forward(value.Copy(ctx, c.values[c.curLayer].View(ctx, rowSize*c.curLoc, vHeadDim*numKVHeads*batchSize)))
+	}
 }
 
 func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
@@ -357,6 +603,35 @@ func (c *Causal) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	c.cellRanges[dstSeq] = seqRange
 }
 
+func (c *Causal) CanResume(seq int, pos int32) bool {
+	if c.windowSize == math.MaxInt32 {
+		return true
+	}
+
+	seqRange, ok := c.cellRanges[seq]
+	if !ok {
+		return false
+	}
+
+	// for sliding window, check that the window of the new sequence is contained in
+	// the window of what we are storing
+	var last int32 = -1
+	for i := seqRange.min; i <= seqRange.max; i++ {
+		if slices.Contains(c.cells[i].sequences, seq) {
+			last = max(last, c.cells[i].pos)
+		}
+	}
+
+	if last == -1 {
+		return false
+	}
+
+	lastWindowStart := max(0, last-c.windowSize)
+	posWindowStart := max(0, pos-c.windowSize)
+
+	return posWindowStart >= lastWindowStart
+}
+
 func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 	if c.shiftFn == nil {
 		return ErrNotSupported
@@ -377,7 +652,7 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 		}
 	}
 
-	kShift, err := ctx.FromIntSlice(offsets, len(offsets))
+	kShift, err := ctx.Input().FromIntSlice(offsets, len(offsets))
 	if err != nil {
 		return err
 	}
@@ -387,9 +662,13 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 			continue
 		}
 
-		key = key.View(ctx, key.Stride(2)*seqRange.min,
-			key.Dim(0), key.Stride(1),
-			key.Dim(1), key.Stride(2),
+		kHeadDim := key.Dim(0)
+		numKVHeads := key.Dim(1)
+		rowSize := key.Stride(2)
+
+		key = key.View(ctx, rowSize*seqRange.min,
+			kHeadDim, key.Stride(1),
+			numKVHeads, key.Stride(2),
 			size,
 		)
 
@@ -407,6 +686,12 @@ func (c *Causal) shift(seq int, beginIndex, offset int32) error {
 }
 
 func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
+	// TODO(jessegross): We should check to see if removing the middle of the sequence will
+	// cause the sliding window to encompass tokens that we no longer have. If so, then we
+	// should return an error, which will trigger the runner to evaluate the full history and
+	// rebuild the window. However, if we have multimodal inputs in our history, this reuse
+	// results in use after free, so we don't do it for now.
+
 	var offset int32
 	if endIndex != math.MaxInt32 {
 		offset = beginIndex - endIndex
@@ -421,8 +706,7 @@ func (c *Causal) Remove(seq int, beginIndex, endIndex int32) error {
 			} else {
 				if c.cells[i].pos >= endIndex {
 					if slices.ContainsFunc(c.cells[i].sequences, func(s int) bool { return s != seq }) {
-						// TODO(jessegross): Need to be careful about data shared between sequences
-						return errors.New("shifting on cells shared by multiple sequences not yet implemented")
+						return errors.New("shifting cells shared by multiple sequences not supported")
 					}
 
 					c.cells[i].pos += offset

kvcache/causal_test.go

@@ -6,6 +6,7 @@ import (
 	"testing"
 
 	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
 )
 
 type testCase struct {
@@ -24,7 +25,7 @@ func TestStore(t *testing.T) {
 	cache := NewCausalCache(nil)
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF16, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
@@ -57,11 +58,11 @@ func TestSWA(t *testing.T) {
 	cache := NewSWACache(1, nil)
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF32, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
-			name:          "SlidingWindow",
+			name:          "FirstBatch",
 			in:            []float32{1, 2, 3, 4},
 			inShape:       []int{1, 1, 4},
 			seqs:          []int{0, 0, 0, 0},
@@ -70,17 +71,85 @@ func TestSWA(t *testing.T) {
 			expectedShape: []int{1, 1, 4},
 			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0},
 		},
+		{
+			name:          "SecondBatch",
+			in:            []float32{5, 6},
+			inShape:       []int{1, 1, 2},
+			seqs:          []int{0, 0},
+			pos:           []int32{4, 5},
+			expected:      []float32{5, 6, 3, 4},
+			expectedShape: []int{1, 1, 4},
+			expectedMask:  []float32{0, float32(math.Inf(-1)), float32(math.Inf(-1)), 0, 0, 0, float32(math.Inf(-1)), float32(math.Inf(-1))},
+		},
 	}
 
 	testCache(t, backend, cache, tests)
 }
 
+func TestChunkedAttention(t *testing.T) {
+	cache := NewChunkedAttentionCache(2, nil)
+	defer cache.Close()
+
+	var b testBackend
+	cache.Init(&b, ml.DTypeF16, 1, 16, 16)
+
+	x := float32(math.Inf(-1))
+
+	testCache(
+		t, &b, cache,
+		[]testCase{
+			{
+				name:          "FirstBatch",
+				in:            []float32{1, 2, 3, 4},
+				inShape:       []int{1, 1, 4},
+				seqs:          []int{0, 0, 0, 0},
+				pos:           []int32{0, 1, 2, 3},
+				expected:      []float32{1, 2, 3, 4},
+				expectedShape: []int{1, 1, 4},
+				expectedMask: []float32{
+					0, x, x, x,
+					0, 0, x, x,
+					x, x, 0, x,
+					x, x, 0, 0,
+				},
+			},
+			{
+				name:          "SecondBatch",
+				in:            []float32{5, 6, 7},
+				inShape:       []int{1, 1, 3},
+				seqs:          []int{0, 0, 0},
+				pos:           []int32{4, 5, 6},
+				expected:      []float32{1, 2, 3, 4, 5, 6, 7},
+				expectedShape: []int{1, 1, 7},
+				expectedMask: []float32{
+					x, x, x, x, 0, x, x,
+					x, x, x, x, 0, 0, x,
+					x, x, x, x, x, x, 0,
+				},
+			},
+			{
+				name:          "ThirdBatch",
+				in:            []float32{8, 9},
+				inShape:       []int{1, 1, 2},
+				seqs:          []int{0, 0},
+				pos:           []int32{7, 8},
+				expected:      []float32{1, 2, 3, 4, 5, 6, 7, 8, 9},
+				expectedShape: []int{1, 1, 9},
+				expectedMask: []float32{
+					x, x, x, x, x, x, 0, 0, x,
+					x, x, x, x, x, x, x, x, 0,
+				},
+			},
+		},
+	)
+}
+
 func TestSequences(t *testing.T) {
 	backend := &testBackend{}
 	cache := NewCausalCache(nil)
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF16, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
@@ -115,7 +184,7 @@ func TestRemove(t *testing.T) {
 	})
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF16, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
@@ -180,7 +249,7 @@ func TestDefrag(t *testing.T) {
 	})
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF16, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
@@ -228,7 +297,7 @@ func TestCopy(t *testing.T) {
 	cache := NewCausalCache(func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) { return key, nil })
 	defer cache.Close()
 
-	cache.Init(backend, ml.DTypeF16, 16)
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
 
 	tests := []testCase{
 		{
@@ -269,7 +338,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			context := backend.NewContext()
 			defer context.Close()
 
-			err := cache.StartForward(context, test.pos, test.seqs)
+			err := cache.StartForward(context, input.Batch{Positions: test.pos, Sequences: test.seqs}, false)
 			if err != nil {
 				panic(err)
 			}
@@ -280,38 +349,111 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 
 			out, _, mask := cache.Get(context)
 
-			context.Forward(out)
-			context.Forward(mask)
-			context.Compute(out, mask)
+			context.Forward(out, mask).Compute(out, mask)
 
-			if !slices.Equal(out.Floats(), test.expected) || !slices.Equal(out.Shape(), test.expectedShape) || !slices.Equal(mask.Floats(), test.expectedMask) {
-				t.Errorf("TestCache: have %v (shape %v); want %v (shape %v); mask: have %v (shape %v) want %v", out.Floats(), out.Shape(), test.expected, test.expectedShape, mask.Floats(), mask.Shape(), test.expectedMask)
+			if !slices.Equal(out.Floats(), test.expected) {
+				t.Errorf("TestCache: have %v; want %v", out.Floats(), test.expected)
+			}
+
+			if !slices.Equal(out.Shape(), test.expectedShape) {
+				t.Errorf("TestCache: has shape %v; want %v", out.Shape(), test.expectedShape)
+			}
+
+			if !slices.Equal(mask.Floats(), test.expectedMask) {
+				t.Errorf("TestCache: have mask: have %v want %v", mask.Floats(), test.expectedMask)
 			}
 		})
 	}
 }
 
-type testBackend struct{}
+func TestCanResume(t *testing.T) {
+	backend := &testBackend{}
+	windowSize := int32(4)
+	cache := NewSWACache(windowSize, nil)
+	defer cache.Close()
 
-func (b *testBackend) Config() ml.Config {
-	panic("not implemented")
+	cache.Init(backend, ml.DTypeF16, 1, 16, 16)
+
+	context := backend.NewContext()
+	defer context.Close()
+
+	err := cache.StartForward(context, input.Batch{
+		Positions: []int32{0, 1, 2, 3},
+		Sequences: []int{0, 0, 0, 0},
+	}, false)
+	if err != nil {
+		t.Fatalf("StartForward failed: %v", err)
+	}
+
+	cache.SetLayer(0)
+	tensor, _ := context.FromFloatSlice([]float32{1, 2, 3, 4}, 1, 1, 4)
+	cache.Put(context, tensor, tensor)
+
+	// with window size 4, nothing has slid out of the window yet
+	if !cache.CanResume(0, 0) {
+		t.Errorf("CanResume(0, 0) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 1) {
+		t.Errorf("CanResume(0, 1) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 2) {
+		t.Errorf("CanResume(0, 2) = false, want true (within window)")
+	}
+	if !cache.CanResume(0, 3) {
+		t.Errorf("CanResume(0, 3) = false, want true (latest position)")
+	}
+
+	// shift window by adding position 4
+	err = cache.StartForward(context, input.Batch{
+		Positions: []int32{4, 5},
+		Sequences: []int{0, 0},
+	}, false)
+	if err != nil {
+		t.Fatalf("StartForward failed: %v", err)
+	}
+
+	cache.SetLayer(0)
+	tensor, _ = context.FromFloatSlice([]float32{5, 6}, 1, 1, 2)
+	cache.Put(context, tensor, tensor)
+
+	// only the latest position has overlapping windows
+	if cache.CanResume(0, 0) {
+		t.Errorf("after shift: CanResume(0, 0) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 1) {
+		t.Errorf("after shift: CanResume(0, 1) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 2) {
+		t.Errorf("after shift: CanResume(0, 2) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 3) {
+		t.Errorf("after shift: CanResume(0, 3) = true, want false (outside window)")
+	}
+	if cache.CanResume(0, 4) {
+		t.Errorf("after shift: CanResume(0, 4) = true, want false (outside window)")
+	}
+	if !cache.CanResume(0, 5) {
+		t.Errorf("after shift: CanResume(0, 5) = false, want true (latest position)")
+	}
 }
 
-func (b *testBackend) Get(name string) ml.Tensor {
-	panic("not implemented")
+type testBackend struct {
+	ml.Backend
 }
 
 func (b *testBackend) NewContext() ml.Context {
 	return &testContext{}
 }
 
-func (b *testBackend) SystemInfo() string {
-	return "not implemented"
+func (b *testBackend) NewContextSize(int) ml.Context {
+	return &testContext{}
 }
 
-type testContext struct{}
+type testContext struct {
+	ml.Context
+}
 
-func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
+func (c *testContext) Empty(dtype ml.DType, shape ...int) ml.Tensor {
 	total := 0
 
 	if len(shape) > 0 {
@@ -324,8 +466,12 @@ func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
 	return &testTensor{dtype: dtype, elementSize: 4, data: make([]float32, total), shape: shape}
 }
 
+func (c *testContext) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
+	return c.Empty(dtype, shape...)
+}
+
 func (c *testContext) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
-	t := c.Zeros(ml.DTypeF32, shape...).(*testTensor)
+	t := c.Empty(ml.DTypeF32, shape...).(*testTensor)
 
 	copy(t.data, s)
 
@@ -344,17 +490,35 @@ func (c *testContext) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
 	return out, nil
 }
 
-func (c *testContext) Forward(ml.Tensor) {}
+func (c *testContext) Arange(start, stop, step float32, dtype ml.DType) ml.Tensor {
+	s := make([]float32, 0, int((stop-start)/step))
+	for i := start; i < stop; i += step {
+		s = append(s, i)
+	}
+
+	out, _ := c.FromFloatSlice(s, len(s))
+	out.(*testTensor).dtype = dtype
+	return out
+}
+
+func (c *testContext) Input() ml.Context    { return c }
+func (c *testContext) Layer(int) ml.Context { return c }
+
+func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
 
 func (c *testContext) Compute(...ml.Tensor) {}
 
-func (c *testContext) MaxTensors() int {
+func (c *testContext) Reserve() error { return nil }
+
+func (c *testContext) MaxGraphNodes() int {
 	return 10
 }
 
 func (c *testContext) Close() {}
 
 type testTensor struct {
+	ml.Tensor
+
 	dtype       ml.DType
 	elementSize int
 	data        []float32
@@ -382,18 +546,22 @@ func (t *testTensor) DType() ml.DType {
 	return t.dtype
 }
 
-func (t *testTensor) Bytes() []byte {
-	panic("not implemented")
-}
-
 func (t *testTensor) Floats() []float32 {
 	out := make([]float32, len(t.data))
 	copy(out, t.data)
 	return out
 }
 
+func (t *testTensor) Neg(ctx ml.Context) ml.Tensor {
+	out := ctx.Empty(t.DType(), t.Shape()...).(*testTensor)
+	for i := range out.data {
+		out.data[i] = -t.data[i]
+	}
+	return out
+}
+
 func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	out := ctx.Zeros(t.DType(), t.Shape()...).(*testTensor)
+	out := ctx.Empty(t.DType(), t.Shape()...).(*testTensor)
 
 	for i := range out.data {
 		out.data[i] = t.data[i] + t2.(*testTensor).data[i]
@@ -402,58 +570,6 @@ func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	return out
 }
 
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) LayerNorm(ctx ml.Context, weight, bias ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RMSNorm(ctx ml.Context, weight ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim uint32, base, scale float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Tanh(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) GELU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) SILU(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	offset /= t.elementSize
 
@@ -470,40 +586,12 @@ func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 
 	context := &testContext{}
 
-	view := context.Zeros(t.dtype, s...).(*testTensor)
+	view := context.Empty(t.dtype, s...).(*testTensor)
 	view.data = t.data[offset : offset+len(view.data)]
 
 	return view
 }
 
-func (t *testTensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	copy(t2.(*testTensor).data, t.data)
 	return nil

kvcache/encoder.go

@@ -1,7 +1,10 @@
 package kvcache
 
 import (
+	"fmt"
+
 	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
 )
 
 // Encoder cache stores K and V tensors that are position independent
@@ -11,6 +14,9 @@ import (
 //
 // Not currently safe for multiple sequences
 type EncoderCache struct {
+	// config controls mostly backend-specific optimizations
+	config *ml.CacheConfig
+
 	// ** current forward pass **
 
 	// the active layer for Get and Put
@@ -21,6 +27,11 @@ type EncoderCache struct {
 	// anything will be stored)
 	curPos int32
 
+	// curReserve indicates that this forward pass is only for
+	// memory reservation and we should not update our metadata
+	// based on it.
+	curReserve bool
+
 	// ** cache metadata **
 
 	// was something stored in the cache?
@@ -30,36 +41,65 @@ type EncoderCache struct {
 	encoderPos int32
 
 	// ** cache data storage **
-
-	cacheCtx     ml.Context
-	keys, values []ml.Tensor
+	backend      ml.Backend
+	ctxs         map[int]ml.Context
+	keys, values map[int]ml.Tensor
 }
 
 func NewEncoderCache() *EncoderCache {
-	return &EncoderCache{}
+	return &EncoderCache{
+		ctxs:   make(map[int]ml.Context),
+		keys:   make(map[int]ml.Tensor),
+		values: make(map[int]ml.Tensor),
+	}
 }
 
-func (c *EncoderCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
-	c.cacheCtx = backend.NewContext()
+func (c *EncoderCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
+	if c.config == nil {
+		var config ml.CacheConfig
+		if cc, ok := backend.(ml.BackendCacheConfig); ok {
+			config = cc.CacheConfig()
+		}
+		c.config = &config
+	}
+
+	if maxSequences > 1 {
+		panic(fmt.Errorf("encoder cache does not support multiple sequences; requested: %v", maxSequences))
+	}
+
+	if c.config.CachePadding != 0 && c.config.CachePadding != 1 {
+		panic(fmt.Errorf("encoder cache is unable to enforce requested CachePadding (%v)", c.config.CachePadding))
+	}
+
+	c.backend = backend
+}
+
+func (c *EncoderCache) SetConfig(config ml.CacheConfig) {
+	if c.config != nil {
+		panic("config cannot be changed after being previously set, either by the model or backend")
+	}
+
+	c.config = &config
 }
 
 func (c *EncoderCache) Close() {
-	c.cacheCtx.Close()
+	for _, ctx := range c.ctxs {
+		ctx.Close()
+	}
 }
 
-func (c *EncoderCache) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
-	// The image is always in the first position
-	c.curPos = positions[0]
+func (c *EncoderCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
+	// We work with the most recent image
+	if len(batch.Multimodal) > 0 {
+		c.curPos = batch.Positions[batch.Multimodal[len(batch.Multimodal)-1].Index]
+	}
+
+	c.curReserve = reserve
 
 	return nil
 }
 
 func (c *EncoderCache) SetLayer(layer int) {
-	if layer >= len(c.keys) {
-		c.keys = append(c.keys, make([]ml.Tensor, layer-len(c.keys)+1)...)
-		c.values = append(c.values, make([]ml.Tensor, layer-len(c.values)+1)...)
-	}
-
 	c.curLayer = layer
 }
 
@@ -72,22 +112,41 @@ func (c *EncoderCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
 }
 
 func (c *EncoderCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.encoderPos = c.curPos
-	c.encoderCached = true
-
-	if c.keys[c.curLayer] == nil || c.values[c.curLayer] == nil {
-		c.keys[c.curLayer] = c.cacheCtx.Zeros(key.DType(), key.Shape()...)
-		c.values[c.curLayer] = c.cacheCtx.Zeros(value.DType(), value.Shape()...)
+	if !c.curReserve {
+		c.encoderPos = c.curPos
+		c.encoderCached = true
 	}
 
-	ctx.Forward(key.Copy(ctx, c.keys[c.curLayer]))
-	ctx.Forward(value.Copy(ctx, c.values[c.curLayer]))
+	if c.config.PermutedV {
+		value = value.Permute(ctx, 1, 2, 0, 3)
+	}
+
+	if _, ok := c.ctxs[c.curLayer]; !ok {
+		c.ctxs[c.curLayer] = c.backend.NewContextSize(2).Layer(c.curLayer)
+	}
+
+	if _, ok := c.keys[c.curLayer]; !ok {
+		c.keys[c.curLayer] = c.ctxs[c.curLayer].Empty(key.DType(), key.Shape()...)
+	}
+
+	if _, ok := c.values[c.curLayer]; !ok {
+		c.values[c.curLayer] = c.ctxs[c.curLayer].Empty(value.DType(), value.Shape()...)
+	}
+
+	ctx.Forward(
+		key.Copy(ctx, c.keys[c.curLayer]),
+		value.Copy(ctx, c.values[c.curLayer]),
+	)
 }
 
 func (c *EncoderCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	panic("encoder cache does not support multiple sequences")
 }
 
+func (c *EncoderCache) CanResume(seq int, pos int32) bool {
+	return true
+}
+
 func (c *EncoderCache) Remove(seq int, beginIndex, endIndex int32) error {
 	if c.encoderPos >= beginIndex && c.encoderPos < endIndex {
 		c.encoderCached = false

kvcache/wrapper.go

@@ -4,6 +4,7 @@ import (
 	"math"
 
 	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/input"
 )
 
 // Wrapper cache is a container for multiple types of caches,
@@ -22,9 +23,15 @@ func NewWrapperCache(caches ...Cache) *WrapperCache {
 	}
 }
 
-func (c *WrapperCache) Init(backend ml.Backend, dtype ml.DType, capacity int32) {
+func (c *WrapperCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {
 	for _, cache := range c.caches {
-		cache.Init(backend, dtype, capacity)
+		cache.Init(backend, dtype, maxSequences, capacity, maxBatch)
+	}
+}
+
+func (c *WrapperCache) SetConfig(config ml.CacheConfig) {
+	for _, cache := range c.caches {
+		cache.SetConfig(config)
 	}
 }
 
@@ -34,14 +41,14 @@ func (c *WrapperCache) Close() {
 	}
 }
 
-func (c *WrapperCache) StartForward(ctx ml.Context, positions []int32, seqs []int) error {
+func (c *WrapperCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	for i, cache := range c.caches {
-		err := cache.StartForward(ctx, positions, seqs)
+		err := cache.StartForward(ctx, batch, reserve)
 		if err != nil {
 			// unwind on error - Remove with endIndex set to math.MaxInt32 does not fail
 			for j := i - 1; j >= 0; j-- {
-				for k := range positions {
-					_ = c.caches[j].Remove(seqs[k], positions[k], math.MaxInt32)
+				for k := range batch.Positions {
+					_ = c.caches[j].Remove(batch.Sequences[k], batch.Positions[k], math.MaxInt32)
 				}
 			}
 			return err
@@ -80,6 +87,16 @@ func (c *WrapperCache) CopyPrefix(srcSeq, dstSeq int, len int32) {
 	}
 }
 
+func (c *WrapperCache) CanResume(seq int, pos int32) bool {
+	for _, cache := range c.caches {
+		if !cache.CanResume(seq, pos) {
+			return false
+		}
+	}
+
+	return true
+}
+
 func (c *WrapperCache) Remove(seq int, beginIndex, endIndex int32) error {
 	// If the one of these fails, the caller is supposed to retry with endIndex set to math.MaxInt32, which should not fail
 	for _, cache := range c.caches {

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "46e3556e01b824e52395fb050b29804b6cff2a7c";
+char const *LLAMA_COMMIT = "2016f07bd106c73699ecbaace80f55db5ed95dac";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/.rsync-filter

@@ -13,6 +13,7 @@ include include/llama-*.*
 include examples/
 include examples/llava/
 include examples/llava/clip.*
+include examples/llava/clip-impl.*
 include examples/llava/llava.*
 include src/
 include src/llama.*

llama/llama.cpp/common/common.cpp

@@ -2,12 +2,11 @@
 #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
 #endif
 
+#include "ggml.h"
+#include "gguf.h"
+
 #include "common.h"
 #include "log.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
-#include "json-schema-to-grammar.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -49,31 +48,11 @@
 #include <sys/stat.h>
 #include <unistd.h>
 #endif
-#if defined(LLAMA_USE_CURL)
-#include <curl/curl.h>
-#include <curl/easy.h>
-#include <future>
-#endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if defined(LLAMA_USE_CURL)
-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#else
-#include <sys/syslimits.h>
-#endif
-#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-#endif // LLAMA_USE_CURL
-
-using json = nlohmann::ordered_json;
-
 //
 // CPU utils
 //
@@ -464,6 +443,53 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+std::string regex_escape(const std::string & s) {
+    static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
+    return std::regex_replace(s, special_chars, "\\$0");
+}
+
+std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
+    std::ostringstream result;
+    for (size_t i = 0; i < values.size(); ++i) {
+        if (i > 0) {
+            result << separator;
+        }
+        result << values[i];
+    }
+    return result.str();
+}
+
+std::vector<std::string> string_split(const std::string & str, const std::string & delimiter) {
+    std::vector<std::string> parts;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        parts.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    parts.push_back(str.substr(start));
+
+    return parts;
+}
+
+std::string string_repeat(const std::string & str, size_t n) {
+    if (n == 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(str.length() * n);
+
+    for (size_t i = 0; i < n; ++i) {
+        result += str;
+    }
+
+    return result;
+}
+
 std::string string_from(bool value) {
     return value ? "true" : "false";
 }
@@ -804,7 +830,7 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#ifdef __linux__
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else {
@@ -814,7 +840,9 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
-#endif // __linux__
+#else
+#  error Unknown architecture
+#endif
         cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
     }
@@ -835,45 +863,39 @@ std::string fs_get_cache_file(const std::string & filename) {
 //
 // Model utils
 //
+
 struct common_init_result common_init_from_params(common_params & params) {
     common_init_result iparams;
     auto mparams = common_model_params_to_llama(params);
 
-    llama_model * model = nullptr;
-
-    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
-        model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams);
-    } else if (!params.model_url.empty()) {
-        model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
-    } else {
-        model = llama_load_model_from_file(params.model.c_str(), mparams);
-    }
-
+    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
     if (model == NULL) {
-        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
         return iparams;
     }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     if (params.reranking) {
         bool ok = true;
 
-        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
+        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
             ok = false;
         }
 
-        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
+        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
             ok = false;
         }
 
         if (!ok) {
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
@@ -881,39 +903,40 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     auto cparams = common_context_params_to_llama(params);
 
-    llama_context * lctx = llama_new_context_with_model(model, cparams);
+    llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
-        llama_free_model(model);
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        llama_model_free(model);
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
-        LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
+    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+        LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
 
     if (!params.control_vectors.empty()) {
         if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
-        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_n_layer(model);
+        if (params.control_vector_layer_end   <= 0) params.control_vector_layer_end   = llama_model_n_layer(model);
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
 
-        int err = llama_control_vector_apply(lctx,
-                                             cvec.data.data(),
-                                             cvec.data.size(),
-                                             cvec.n_embd,
-                                             params.control_vector_layer_start,
-                                             params.control_vector_layer_end);
+        int err = llama_apply_adapter_cvec(
+                lctx,
+                cvec.data.data(),
+                cvec.data.size(),
+                cvec.n_embd,
+                params.control_vector_layer_start,
+                params.control_vector_layer_end);
         if (err) {
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
 
             return iparams;
         }
@@ -921,12 +944,12 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
-        llama_lora_adapter_ptr lora;
-        lora.reset(llama_lora_adapter_init(model, la.path.c_str()));
+        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());
             llama_free(lctx);
-            llama_free_model(model);
+            llama_model_free(model);
             return iparams;
         }
 
@@ -935,17 +958,17 @@ struct common_init_result common_init_from_params(common_params & params) {
     }
 
     if (!params.lora_init_without_apply) {
-        common_lora_adapters_apply(lctx, params.lora_adapters);
+        common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
-    if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
-        LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
+    if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
+        LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sampling.ignore_eos = false;
     }
 
     if (params.sampling.ignore_eos) {
-        for (llama_token i = 0; i < llama_n_vocab(model); i++) {
-            if (llama_token_is_eog(model, i)) {
+        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+            if (llama_vocab_is_eog(vocab, i)) {
                 LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
                 params.sampling.logit_bias.push_back({i, -INFINITY});
             }
@@ -965,9 +988,12 @@ struct common_init_result common_init_from_params(common_params & params) {
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
+        llama_set_warmup(lctx, true);
+
         std::vector<llama_token> tmp;
-        llama_token bos = llama_token_bos(model);
-        llama_token eos = llama_token_eos(model);
+        llama_token bos = llama_vocab_bos(vocab);
+        llama_token eos = llama_vocab_eos(vocab);
+
         // some models (e.g. T5) don't have a BOS token
         if (bos != LLAMA_TOKEN_NULL) {
             tmp.push_back(bos);
@@ -982,7 +1008,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_encoder(model)) {
             llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
-            if (decoder_start_token_id == -1) {
+            if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
                 decoder_start_token_id = bos;
             }
             tmp.clear();
@@ -991,9 +1017,10 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_cache_clear(lctx);
+        llama_kv_self_clear(lctx);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
+        llama_set_warmup(lctx, false);
     }
 
     iparams.model.reset(model);
@@ -1002,11 +1029,24 @@ struct common_init_result common_init_from_params(common_params & params) {
     return iparams;
 }
 
-void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora) {
-    llama_lora_adapter_clear(ctx);
+std::string get_model_endpoint() {
+    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
+    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
+    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
+    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
+    std::string model_endpoint = "https://huggingface.co/";
+    if (endpoint_env) {
+        model_endpoint = endpoint_env;
+        if (model_endpoint.back() != '/') model_endpoint += '/';
+    }
+    return model_endpoint;
+}
+
+void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
+    llama_clear_adapter_lora(ctx);
     for (auto & la : lora) {
         if (la.scale != 0.0f) {
-            llama_lora_adapter_set(ctx, la.ptr, la.scale);
+            llama_set_adapter_lora(ctx, la.ptr, la.scale);
         }
     }
 }
@@ -1017,16 +1057,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     if (!params.devices.empty()) {
         mparams.devices = params.devices.data();
     }
+
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
-    mparams.rpc_servers     = params.rpc_servers.c_str();
+
     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_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1034,6 +1076,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
         mparams.kv_overrides = params.kv_overrides.data();
     }
 
+    if (params.tensor_buft_overrides.empty()) {
+        mparams.tensor_buft_overrides = NULL;
+    } else {
+        GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+        mparams.tensor_buft_overrides = params.tensor_buft_overrides.data();
+    }
+
     return mparams;
 }
 
@@ -1093,373 +1142,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
     return tpp;
 }
 
-#ifdef LLAMA_USE_CURL
-
-#define CURL_MAX_RETRY 3
-#define CURL_RETRY_DELAY_SECONDS 2
-
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
-    int remaining_attempts = max_attempts;
-
-    while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
-
-        CURLcode res = curl_easy_perform(curl);
-        if (res == CURLE_OK) {
-            return true;
-        }
-
-        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
-        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
-
-        remaining_attempts--;
-        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-    }
-
-    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
-
-    return false;
-}
-
-static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
-    // Initialize libcurl
-    std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
-    if (!curl) {
-        LOG_ERR("%s: error initializing libcurl\n", __func__);
-        return false;
-    }
-
-    bool force_download = false;
-
-    // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-
-    // Check if hf-token or bearer-token was specified
-    if (!hf_token.empty()) {
-      std::string auth_header = "Authorization: Bearer ";
-      auth_header += hf_token.c_str();
-      struct curl_slist *http_headers = NULL;
-      http_headers = curl_slist_append(http_headers, auth_header.c_str());
-      curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
-    }
-
-#if defined(_WIN32)
-    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
-    //   operating system. Currently implemented under MS-Windows.
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
-
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
-    std::string etag;
-    std::string last_modified;
-
-    if (file_exists) {
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
-            }
-        }
-    } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
-    }
-
-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
-    };
-
-    common_load_model_from_url_headers headers;
-
-    {
-        typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-        auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-            common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
-
-            static std::regex header_regex("([^:]+): (.*)\r\n");
-            static std::regex etag_regex("ETag", std::regex_constants::icase);
-            static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-            std::string header(buffer, n_items);
-            std::smatch match;
-            if (std::regex_match(header, match, header_regex)) {
-                const std::string & key = match[1];
-                const std::string & value = match[2];
-                if (std::regex_match(key, match, etag_regex)) {
-                    headers->etag = value;
-                } else if (std::regex_match(key, match, last_modified_regex)) {
-                    headers->last_modified = value;
-                }
-            }
-            return n_items;
-        };
-
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-        }
-    }
-
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
-        if (!etag.empty() && etag != headers.etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
-            should_download = true;
-        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
-            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
-            should_download = true;
-        }
-    }
-    if (should_download) {
-        std::string path_temporary = path + ".downloadInProgress";
-        if (file_exists) {
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
-            }
-        }
-
-        // Set the output file
-
-        struct FILE_deleter {
-            void operator()(FILE * f) const {
-                fclose(f);
-            }
-        };
-
-        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
-        if (!outfile) {
-            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
-            return false;
-        }
-
-        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
-        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
-            return fwrite(data, size, nmemb, (FILE *)fd);
-        };
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
-
-        //  display download progress
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
-
-        // helper function to hide password in URL
-        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
-            std::size_t protocol_pos = url.find("://");
-            if (protocol_pos == std::string::npos) {
-                return url;  // Malformed URL
-            }
-
-            std::size_t at_pos = url.find('@', protocol_pos + 3);
-            if (at_pos == std::string::npos) {
-                return url;  // No password in URL
-            }
-
-            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
-        };
-
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
-        if (!was_perform_successful) {
-            return false;
-        }
-
-        long http_code = 0;
-        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code < 200 || http_code >= 400) {
-            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-            return false;
-        }
-
-        // Causes file to be closed explicitly here before we rename it.
-        outfile.reset();
-
-        // Write the updated JSON metadata file.
-        metadata.update({
-            {"url", url},
-            {"etag", headers.etag},
-            {"lastModified", headers.last_modified}
-        });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
-
-        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
-        }
-    }
-
-    return true;
-}
-
-struct llama_model * common_load_model_from_url(
-        const std::string & model_url,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // Basic validation of the model_url
-    if (model_url.empty()) {
-        LOG_ERR("%s: invalid model_url\n", __func__);
-        return NULL;
-    }
-
-    if (!common_download_file(model_url, local_path, hf_token)) {
-        return NULL;
-    }
-
-    // check for additional GGUFs split to download
-    int n_split = 0;
-    {
-        struct gguf_init_params gguf_params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ NULL,
-        };
-        auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params);
-        if (!ctx_gguf) {
-            LOG_ERR("\n%s:  failed to load input GGUF from %s\n", __func__, local_path.c_str());
-            return NULL;
-        }
-
-        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-        if (key_n_split >= 0) {
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-        }
-
-        gguf_free(ctx_gguf);
-    }
-
-    if (n_split > 1) {
-        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-
-        // Verify the first split file format
-        // and extract split URL and PATH prefixes
-        {
-            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split);
-                return NULL;
-            }
-
-            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) {
-                LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split);
-                return NULL;
-            }
-        }
-
-        // Prepare download in parallel
-        std::vector<std::future<bool>> futures_download;
-        for (int idx = 1; idx < n_split; idx++) {
-            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool {
-                char split_path[PATH_MAX] = {0};
-                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
-
-                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
-                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
-
-                return common_download_file(split_url, split_path, hf_token);
-            }, idx));
-        }
-
-        // Wait for all downloads to complete
-        for (auto & f : futures_download) {
-            if (!f.get()) {
-                return NULL;
-            }
-        }
-    }
-
-    return llama_load_model_from_file(local_path.c_str(), params);
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & repo,
-        const std::string & remote_path,
-        const std::string & local_path,
-        const std::string & hf_token,
-        const struct llama_model_params & params) {
-    // construct hugging face model url:
-    //
-    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
-    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
-    //
-    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
-    //
-
-    std::string model_url = "https://huggingface.co/";
-    model_url += repo;
-    model_url += "/resolve/main/";
-    model_url += remote_path;
-
-    return common_load_model_from_url(model_url, local_path, hf_token, params);
-}
-
-#else
-
-struct llama_model * common_load_model_from_url(
-        const std::string & /*model_url*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
-    return nullptr;
-}
-
-struct llama_model * common_load_model_from_hf(
-        const std::string & /*repo*/,
-        const std::string & /*remote_path*/,
-        const std::string & /*local_path*/,
-        const std::string & /*hf_token*/,
-        const struct llama_model_params & /*params*/) {
-    LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
-    return nullptr;
-}
-
-#endif // LLAMA_USE_CURL
-
 //
 // Batch utils
 //
@@ -1556,21 +1238,23 @@ std::vector<llama_token> common_tokenize(
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
-    return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    return common_tokenize(vocab, text, add_special, parse_special);
 }
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_model * model,
+    const struct llama_vocab * vocab,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special) {
     // upper limit for the number of tokens
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+        int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -1579,12 +1263,18 @@ std::vector<llama_token> common_tokenize(
 }
 
 std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    return common_token_to_piece(vocab, token, special);
+}
+
+std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
     std::string piece;
     piece.resize(piece.capacity());  // using string internal cache, 15 bytes + '\n'
-    const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+    const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
     if (n_chars < 0) {
         piece.resize(-n_chars);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
+        int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
         GGML_ASSERT(check == -n_chars);
     }
     else {
@@ -1594,13 +1284,19 @@ std::string common_token_to_piece(const struct llama_context * ctx, llama_token
     return piece;
 }
 
-std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
+std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    return common_detokenize(vocab, tokens, special);
+}
+
+std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
     std::string text;
     text.resize(std::max(text.capacity(), tokens.size()));
-    int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+    int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
     if (n_chars < 0) {
         text.resize(-n_chars);
-        n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
+        n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
         GGML_ASSERT(n_chars <= (int32_t)text.size());  // whitespace trimming is performed after per-token detokenization
     }
 
@@ -1610,103 +1306,6 @@ std::string common_detokenize(llama_context * ctx, const std::vector<llama_token
     return text;
 }
 
-//
-// Chat template utils
-//
-
-std::string common_get_builtin_chat_template(const struct llama_model * model) {
-    static const char * template_key = "tokenizer.chat_template";
-    // call with NULL buffer to get the total size of the string
-    int32_t res = llama_model_meta_val_str(model, template_key, NULL, 0);
-    if (res > 0) {
-        std::vector<char> model_template(res + 1, 0);
-        llama_model_meta_val_str(model, template_key, model_template.data(), model_template.size());
-        return std::string(model_template.data(), model_template.size() - 1);
-    }
-    return "";
-}
-
-bool common_chat_verify_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
-    return res >= 0;
-}
-
-std::string common_chat_apply_template(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<common_chat_msg> & msgs,
-        bool add_ass) {
-    int alloc_size = 0;
-    bool fallback = false; // indicate if we must fallback to default chatml
-    std::vector<llama_chat_message> chat;
-    for (auto & msg : msgs) {
-        chat.push_back({msg.role.c_str(), msg.content.c_str()});
-        alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
-    }
-
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector<char> buf(alloc_size);
-
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
-
-    // error: chat template is not supported
-    if (res < 0) {
-        if (ptr_tmpl != nullptr) {
-            // if the custom "tmpl" is not supported, we throw an error
-            // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
-            throw std::runtime_error("this custom template is not supported");
-        } else {
-            // If the built-in template is not supported, we default to chatml
-            res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
-            fallback = true;
-        }
-    }
-
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(
-            fallback ? nullptr : model,
-            fallback ? "chatml" : ptr_tmpl,
-            chat.data(), chat.size(), add_ass, buf.data(), buf.size());
-    }
-
-    std::string formatted_chat(buf.data(), res);
-    return formatted_chat;
-}
-
-std::string common_chat_format_single(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<common_chat_msg> & past_msg,
-        const common_chat_msg & new_msg,
-        bool add_ass) {
-    std::ostringstream ss;
-    auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
-    std::vector<common_chat_msg> chat_new(past_msg);
-    // if the past_msg ends with a newline, we must preserve it in the formatted version
-    if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
-        ss << "\n";
-    };
-    // format chat with new_msg
-    chat_new.push_back(new_msg);
-    auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
-    // get the diff part
-    ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
-    return ss.str();
-}
-
-std::string common_chat_format_example(const struct llama_model * model,
-        const std::string & tmpl) {
-    std::vector<common_chat_msg> msgs = {
-        {"system",    "You are a helpful assistant"},
-        {"user",      "Hello"},
-        {"assistant", "Hi there"},
-        {"user",      "How are you?"},
-    };
-    return common_chat_apply_template(model, tmpl, msgs, true);
-}
-
 //
 // KV cache utils
 //
@@ -1966,4 +1565,3 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
-

llama/llama.cpp/common/common.h

@@ -4,6 +4,7 @@
 
 #include "llama-cpp.h"
 
+#include <set>
 #include <string>
 #include <vector>
 #include <sstream>
@@ -24,11 +25,11 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct common_lora_adapter_info {
+struct common_adapter_lora_info {
     std::string path;
     float scale;
 
-    struct llama_lora_adapter * ptr;
+    struct llama_adapter_lora * ptr;
 };
 
 using llama_tokens = std::vector<llama_token>;
@@ -103,6 +104,25 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
+enum common_conversation_mode {
+    COMMON_CONVERSATION_MODE_DISABLED = 0,
+    COMMON_CONVERSATION_MODE_ENABLED  = 1,
+    COMMON_CONVERSATION_MODE_AUTO     = 2,
+};
+
+enum common_grammar_trigger_type {
+    COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+};
+
+struct common_grammar_trigger {
+    common_grammar_trigger_type type;
+    std::string value;
+    llama_token token = LLAMA_TOKEN_NULL;
+};
+
 // sampling parameters
 struct common_params_sampling {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@@ -128,6 +148,7 @@ struct common_params_sampling {
     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
     bool    ignore_eos         = false;
@@ -148,7 +169,10 @@ struct common_params_sampling {
         COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
-    std::string grammar; // optional BNF-like grammar to constrain sampling
+    std::string                         grammar; // optional BNF-like grammar to constrain sampling
+    bool                                grammar_lazy = false;
+    std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
+    std::set<llama_token>               preserved_tokens;
 
     std::vector<llama_logit_bias> logit_bias; // logit biases to apply
 
@@ -156,28 +180,40 @@ struct common_params_sampling {
     std::string print() const;
 };
 
+struct common_params_model {
+    std::string path    = ""; // model local path                                           // NOLINT
+    std::string url     = ""; // model url to download                                      // NOLINT
+    std::string hf_repo = ""; // HF repo                                                    // NOLINT
+    std::string hf_file = ""; // HF file                                                    // NOLINT
+};
+
 struct common_params_speculative {
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
     int32_t n_ctx        =     0; // draft context size
     int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_min        =     0; // minimum number of draft tokens to use for speculative decoding
     int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     float   p_split      =  0.1f; // speculative decoding split probability
-    float   p_min        =  0.9f; // minimum speculative decoding probability (greedy)
+    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    std::string model = ""; // draft model for speculative decoding                          // NOLINT
+    struct common_params_model model;
 };
 
 struct common_params_vocoder {
-    std::string hf_repo = ""; // HF repo                                                     // NOLINT
-    std::string hf_file = ""; // HF file                                                     // NOLINT
+    struct common_params_model model;
 
-    std::string model     = ""; // model path                                                // NOLINT
-    std::string model_url = ""; // model url to download                                     // NOLINT
+    std::string speaker_file = ""; // speaker file path                                      // NOLINT
+
+    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
+};
+
+enum common_reasoning_format {
+    COMMON_REASONING_FORMAT_NONE,
+    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
 };
 
 struct common_params {
@@ -226,13 +262,12 @@ struct common_params {
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
 
-    std::string model                = ""; // model path                                                    // NOLINT
+    struct common_params_model model;
+
     std::string model_alias          = ""; // model alias                                                   // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
+    std::string system_prompt        = "";                                                                  // NOLINT
     std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
@@ -240,14 +275,14 @@ struct common_params {
     std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
     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
-    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // 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;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
-    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
-    std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
+    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
+    std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -271,11 +306,11 @@ struct common_params {
     bool   kl_divergence    = false; // compute KL divergence
 
     bool usage             = false; // print usage
+    bool completion        = false; // print source-able completion script
     bool use_color         = false; // use color to distinguish generations and inputs
     bool special           = false; // enable special token output
     bool interactive       = false; // interactive mode
     bool interactive_first = false; // wait for user input immediately
-    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
@@ -298,11 +333,15 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
 
+    bool single_turn       = false; // single turn chat conversation
+
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
+    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
+
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    struct common_params_model mmproj;
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -322,7 +361,9 @@ struct common_params {
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
+    bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
 
     std::vector<std::string> api_keys;
 
@@ -360,8 +401,6 @@ struct common_params {
     int32_t i_pos  = -1;  // position of the passkey in the junk text
 
     // imatrix params
-    std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
-
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
@@ -373,16 +412,16 @@ struct common_params {
     int n_pca_batch = 100;
     int n_pca_iterations = 1000;
     dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_outfile       = "control_vector.gguf";
     std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 
-    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
-
     // batched-bench params
     bool batched_bench_output_jsonl = false;
+
+    // common params
+    std::string out_file; // output filename for all example programs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -401,13 +440,13 @@ bool set_process_priority(enum ggml_sched_priority prio);
 //
 
 #ifdef __GNUC__
-#ifdef __MINGW32__
-#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
 #else
-#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-#else
-#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#    define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
 #endif
 
 LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
@@ -416,8 +455,14 @@ std::string string_format(const char * fmt, ...);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
+std::string string_join(const std::vector<std::string> & values, const std::string & separator);
+std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
+std::string string_repeat(const std::string & str, size_t n);
+
 void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
 
+std::string regex_escape(const std::string & s);
+
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
@@ -454,6 +499,11 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
+static bool string_ends_with(const std::string & str,
+                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}
+
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
@@ -481,7 +531,7 @@ struct common_init_result {
     llama_model_ptr   model;
     llama_context_ptr context;
 
-    std::vector<llama_lora_adapter_ptr> lora;
+    std::vector<llama_adapter_lora_ptr> lora;
 };
 
 struct common_init_result     common_init_from_params(common_params & params);
@@ -490,20 +540,10 @@ struct llama_model_params     common_model_params_to_llama  (      common_params
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
-struct llama_model * common_load_model_from_url(
-    const std::string & model_url,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-struct llama_model * common_load_model_from_hf(
-    const std::string & repo,
-    const std::string & remote_path,
-    const std::string & local_path,
-    const std::string & hf_token,
-    const struct llama_model_params & params);
-
 // clear LoRA adapters from context, then apply new list of adapters
-void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_info> & lora);
+void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
+
+std::string                   get_model_endpoint();
 
 //
 // Batch utils
@@ -541,7 +581,7 @@ std::vector<llama_token> common_tokenize(
                         bool   parse_special = false);
 
 std::vector<llama_token> common_tokenize(
-    const struct llama_model * model,
+    const struct llama_vocab * vocab,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
@@ -553,48 +593,23 @@ std::string common_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
+std::string common_token_to_piece(
+          const struct llama_vocab * vocab,
+                       llama_token   token,
+                       bool          special = true);
+
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
 // optionally renders special/control tokens
 std::string common_detokenize(
-                         llama_context * ctx,
+            const struct llama_context * ctx,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-//
-// Chat template utils
-//
-
-// same with llama_chat_message, but uses std::string
-struct common_chat_msg {
-    std::string role;
-    std::string content;
-};
-
-// Get the built-in chat template for the model. Return empty string if not present.
-std::string common_get_builtin_chat_template(const struct llama_model * model);
-
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool common_chat_verify_template(const std::string & tmpl);
-
-// CPP wrapper for llama_chat_apply_template
-// If the built-in template is not supported, we default to chatml
-// If the custom "tmpl" is not supported, we throw an error
-std::string common_chat_apply_template(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<common_chat_msg> & chat,
-        bool add_ass);
-
-// Format single message, while taking into account the position of that message in chat history
-std::string common_chat_format_single(const struct llama_model * model,
-        const std::string & tmpl,
-        const std::vector<common_chat_msg> & past_msg,
-        const common_chat_msg & new_msg,
-        bool add_ass);
-
-// Returns an example of formatted chat
-std::string common_chat_format_example(const struct llama_model * model,
-        const std::string & tmpl);
+std::string common_detokenize(
+              const struct llama_vocab * vocab,
+        const std::vector<llama_token> & tokens,
+                                  bool   special = true);
 
 //
 // KV cache utils

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -1,4 +1,6 @@
 #include "json-schema-to-grammar.h"
+#include "common.h"
+
 #include <algorithm>
 #include <fstream>
 #include <map>
@@ -11,11 +13,6 @@
 
 using json = nlohmann::ordered_json;
 
-template <typename Iterator>
-static std::string join(Iterator begin, Iterator end, const std::string & separator);
-
-static std::string repeat(const std::string & str, size_t n);
-
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
@@ -128,8 +125,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
                 if (sub_len > 0) {
                     auto from_sub = from.substr(i + 1);
                     auto to_sub = to.substr(i + 1);
-                    auto sub_zeros = repeat("0", sub_len);
-                    auto sub_nines = repeat("9", sub_len);
+                    auto sub_zeros = string_repeat("0", sub_len);
+                    auto sub_nines = string_repeat("9", sub_len);
 
                     auto to_reached = false;
                     out << "(";
@@ -188,8 +185,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         auto max_digits = max_s.length();
 
         for (auto digits = min_digits; digits < max_digits; digits++) {
-            uniform_range(min_s, repeat("9", digits));
-            min_s = "1" + repeat("0", digits);
+            uniform_range(min_s, string_repeat("9", digits));
+            min_s = "1" + string_repeat("0", digits);
             out << " | ";
         }
         uniform_range(min_s, max_s);
@@ -267,7 +264,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
     throw std::runtime_error("At least one of min_value or max_value must be set");
 }
 
-const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}";
+const std::string SPACE_RULE = "| \" \" | \"\\n\"{1,2} [ \\t]{0,20}";
 
 struct BuiltinRule {
     std::string content;
@@ -318,49 +315,6 @@ std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
 std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
 std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
-template <typename Iterator>
-std::string join(Iterator begin, Iterator end, const std::string & separator) {
-    std::ostringstream result;
-    if (begin != end) {
-        result << *begin;
-        for (Iterator it = begin + 1; it != end; ++it) {
-            result << separator << *it;
-        }
-    }
-    return result.str();
-}
-
-static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
-    std::vector<std::string> tokens;
-    size_t start = 0;
-    size_t end = str.find(delimiter);
-
-    while (end != std::string::npos) {
-        tokens.push_back(str.substr(start, end - start));
-        start = end + delimiter.length();
-        end = str.find(delimiter, start);
-    }
-
-    tokens.push_back(str.substr(start));
-
-    return tokens;
-}
-
-static std::string repeat(const std::string & str, size_t n) {
-    if (n == 0) {
-        return "";
-    }
-
-    std::string result;
-    result.reserve(str.length() * n);
-
-    for (size_t i = 0; i < n; ++i) {
-        result += str;
-    }
-
-    return result;
-}
-
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
     std::smatch match;
     std::string result;
@@ -389,6 +343,7 @@ static std::string format_literal(const std::string & literal) {
 
 class SchemaConverter {
 private:
+    friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
     std::function<json(const std::string &)> _fetch_json;
     bool _dotall;
     std::unordered_map<std::string, std::string> _rules;
@@ -418,7 +373,7 @@ private:
         for (size_t i = 0; i < alt_schemas.size(); i++) {
             rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
         }
-        return join(rules.begin(), rules.end(), " | ");
+        return string_join(rules, " | ");
     }
 
     std::string _visit_pattern(const std::string & pattern, const std::string & name) {
@@ -481,7 +436,7 @@ private:
                 for (const auto & item : ret) {
                     results.push_back(to_rule(item));
                 }
-                return std::make_pair(join(results.begin(), results.end(), " "), false);
+                return std::make_pair(string_join(results, " "), false);
             };
 
             while (i < length) {
@@ -539,7 +494,7 @@ private:
                     }
                     curly_brackets += '}';
                     i++;
-                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
+                    auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                     int min_times = 0;
                     int max_times = std::numeric_limits<int>::max();
                     try {
@@ -854,7 +809,7 @@ public:
                             return;
                         }
                         std::string pointer = ref.substr(ref.find('#') + 1);
-                        std::vector<std::string> tokens = split(pointer, "/");
+                        std::vector<std::string> tokens = string_split(pointer, "/");
                         for (size_t i = 1; i < tokens.size(); ++i) {
                             std::string sel = tokens[i];
                             if (target.is_null() || !target.contains(sel)) {
@@ -905,7 +860,7 @@ public:
             for (const auto & v : schema["enum"]) {
                 enum_values.push_back(_generate_constant_rule(v));
             }
-            return _add_rule(rule_name, "(" + join(enum_values.begin(), enum_values.end(), " | ") + ") space");
+            return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ") space");
         } else if ((schema_type.is_null() || schema_type == "object")
                 && (schema.contains("properties") ||
                     (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
@@ -1019,10 +974,10 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
+            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
-            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
+            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());
         }
     }
 
@@ -1035,11 +990,35 @@ public:
     }
 };
 
-std::string json_schema_to_grammar(const json & schema) {
-    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
-    auto copy = schema;
-    converter.resolve_refs(copy, "input");
-    converter.visit(copy, "");
+std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
+#ifdef LLAMA_USE_LLGUIDANCE
+    if (!force_gbnf) {
+        return "%llguidance {}\nstart: %json " + schema.dump();
+    }
+#else
+    (void)force_gbnf;
+#endif // LLAMA_USE_LLGUIDANCE
+    return build_grammar([&](const common_grammar_builder & callbacks) {
+        auto copy = schema;
+        callbacks.resolve_refs(copy);
+        callbacks.add_schema("", copy);
+    });
+}
+
+std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
+    SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall);
+    common_grammar_builder builder {
+        /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
+            return converter._add_rule(name, rule);
+        },
+        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
+            return converter.visit(schema, name == "root" ? "" : name);
+        },
+        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
+            converter.resolve_refs(schema, "");
+        }
+    };
+    cb(builder);
     converter.check_errors();
     return converter.format_grammar();
 }

llama/llama.cpp/common/json-schema-to-grammar.h

@@ -5,4 +5,17 @@
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);
+std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
+                                   bool force_gbnf = false);
+
+struct common_grammar_builder {
+    std::function<std::string(const std::string &, const std::string &)> add_rule;
+    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
+    std::function<void(nlohmann::ordered_json &)> resolve_refs;
+};
+
+struct common_grammar_options {
+    bool dotall = false;
+};
+
+std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

llama/llama.cpp/common/log.cpp

@@ -1,5 +1,6 @@
 #include "log.h"
 
+#include <chrono>
 #include <condition_variable>
 #include <cstdarg>
 #include <cstdio>
@@ -14,16 +15,6 @@ void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
-#define LOG_COL_DEFAULT "\033[0m"
-#define LOG_COL_BOLD    "\033[1m"
-#define LOG_COL_RED     "\033[31m"
-#define LOG_COL_GREEN   "\033[32m"
-#define LOG_COL_YELLOW  "\033[33m"
-#define LOG_COL_BLUE    "\033[34m"
-#define LOG_COL_MAGENTA "\033[35m"
-#define LOG_COL_CYAN    "\033[36m"
-#define LOG_COL_WHITE   "\033[37m"
-
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -206,6 +197,7 @@ public:
                 vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
             }
 #endif
+            va_end(args_copy);
         }
 
         entry.level = level;

llama/llama.cpp/common/log.h

@@ -2,9 +2,20 @@
 
 #include "ggml.h" // for ggml_log_level
 
+#define LOG_CLR_TO_EOL  "\033[K\r"
+#define LOG_COL_DEFAULT "\033[0m"
+#define LOG_COL_BOLD    "\033[1m"
+#define LOG_COL_RED     "\033[31m"
+#define LOG_COL_GREEN   "\033[32m"
+#define LOG_COL_YELLOW  "\033[33m"
+#define LOG_COL_BLUE    "\033[34m"
+#define LOG_COL_MAGENTA "\033[35m"
+#define LOG_COL_CYAN    "\033[36m"
+#define LOG_COL_WHITE   "\033[37m"
+
 #ifndef __GNUC__
 #    define LOG_ATTRIBUTE_FORMAT(...)
-#elif defined(__MINGW32__)
+#elif defined(__MINGW32__) && !defined(__clang__)
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
 #else
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))

llama/llama.cpp/common/sampling.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <unordered_map>
+#include <algorithm>
 
 // the ring buffer works similarly to std::deque, but with a fixed capacity
 // TODO: deduplicate with llama-impl.h
@@ -113,7 +114,10 @@ struct common_sampler {
     void set_logits(struct llama_context * ctx, int idx) {
         const auto * logits = llama_get_logits_ith(ctx, idx);
 
-        const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+        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);
 
@@ -131,24 +135,87 @@ std::string common_params_sampling::print() const {
     snprintf(result, sizeof(result),
             "\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, temp = %.3f\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",
             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, temp,
+            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
 }
 
 struct common_sampler * common_sampler_init(const struct llama_model * model, const 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;
 
+    struct llama_sampler * grmr;
+    if (params.grammar.compare(0, 11, "%llguidance") == 0) {
+#ifdef LLAMA_USE_LLGUIDANCE
+        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> patterns_at_start;
+        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));
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN:
+                case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START:
+                {
+                    const auto & pattern = trigger.value;
+                    (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern);
+                    break;
+                }
+                case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN:
+                {
+                    const auto token = trigger.token;
+                    trigger_tokens.push_back(token);
+                    break;
+                }
+                default:
+                    GGML_ASSERT(false && "unknown trigger type");
+            }
+        }
+
+        std::vector<std::string> trigger_patterns;
+        if (!patterns_at_start.empty()) {
+            trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*");
+        }
+        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) {
+            trigger_patterns_c.push_back(regex.c_str());
+        }
+
+        grmr = params.grammar_lazy
+             ? 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())
+             :      llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
+        if (!grmr) {
+            return nullptr;
+        }
+    }
+
     auto * result = new common_sampler {
         /* .params = */ params,
-        /* .grmr   = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
+        /* .grmr   = */ grmr,
         /* .chain  = */ llama_sampler_chain_init(lparams),
         /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur    = */ {},
@@ -157,56 +224,62 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     llama_sampler_chain_add(result->chain,
             llama_sampler_init_logit_bias(
-                llama_n_vocab(model),
+                llama_vocab_n_tokens(vocab),
                 params.logit_bias.size(),
                 params.logit_bias.data()));
 
     if (params.mirostat == 0) {
-        for (const auto & cnstr : params.samplers) {
-            switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char *> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
+        if (params.top_n_sigma >= 0) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+        } else {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                        {
+                            std::vector<const char *> c_breakers;
+                            c_breakers.reserve(params.dry_sequence_breakers.size());
+                            for (const auto & str : params.dry_sequence_breakers) {
+                                c_breakers.push_back(str.c_str());
+                            }
 
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (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:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                    break;
-                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                    break;
-                case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                    break;
-                default:
-                    GGML_ASSERT(false && "unknown sampler type");
+                            llama_sampler_chain_add(result->chain, 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:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    case COMMON_SAMPLER_TYPE_INFILL:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                        break;
+                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
             }
         }
         llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
     } else if (params.mirostat == 1) {
         llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
     } else if (params.mirostat == 2) {
         llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
         llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));

llama/llama.cpp/common/sampling.h

@@ -102,3 +102,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
 
 std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);
+
+llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
+                const char * grammar_kind, const char * grammar_data);

llama/llama.cpp/examples/llava/clip-impl.h

@@ -0,0 +1,341 @@
+#include "ggml.h"
+#include "gguf.h"
+#include "clip.h"
+
+#include "clip.h"
+
+#include <climits>
+#include <cstdarg>
+#include <string>
+#include <map>
+#include <sstream>
+#include <vector>
+#include <memory>
+
+// Internal header for clip.cpp
+
+#define KEY_FTYPE               "general.file_type"
+#define KEY_NAME                "general.name"
+#define KEY_DESCRIPTION         "general.description"
+#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
+#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
+#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
+#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
+#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
+#define KEY_USE_GELU            "clip.use_gelu"
+#define KEY_USE_SILU            "clip.use_silu"
+#define KEY_N_EMBD              "clip.%s.embedding_length"
+#define KEY_N_FF                "clip.%s.feed_forward_length"
+#define KEY_N_BLOCK             "clip.%s.block_count"
+#define KEY_N_HEAD              "clip.%s.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.%s.projection_dim"
+#define KEY_TOKENS              "tokenizer.ggml.tokens"
+#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_IMAGE_SIZE          "clip.vision.image_size"
+#define KEY_PATCH_SIZE          "clip.vision.patch_size"
+#define KEY_IMAGE_MEAN          "clip.vision.image_mean"
+#define KEY_IMAGE_STD           "clip.vision.image_std"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+
+#define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
+#define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
+#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+
+
+//
+// tensor name constants
+//
+
+#define TN_POS_EMBD        "%s.position_embd.weight"
+#define TN_CLASS_EMBD      "v.class_embd"
+#define TN_PATCH_EMBD      "v.patch_embd.weight"  // not rename tensor with ".0" postfix for backwrad compat
+#define TN_PATCH_EMBD_1    "v.patch_embd.weight.1"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
+#define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
+#define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
+#define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
+#define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
+#define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_LN_1            "%s.blk.%d.ln1.%s"
+#define TN_LN_2            "%s.blk.%d.ln2.%s"
+#define TN_LN_PRE          "%s.pre_ln.%s"
+#define TN_LN_POST         "%s.post_ln.%s"
+#define TN_LLAVA_PROJ      "mm.%d.%s"
+#define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
+#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
+#define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
+#define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+
+// mimicpmv
+#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
+#define TN_MINICPMV_QUERY      "resampler.query"
+#define TN_MINICPMV_PROJ       "resampler.proj.weight"
+#define TN_MINICPMV_KV_PROJ    "resampler.kv.weight"
+#define TN_MINICPMV_ATTN       "resampler.attn.%s.%s"
+#define TN_MINICPMV_LN         "resampler.ln_%s.%s"
+
+#define TN_GLM_ADAPER_CONV      "adapter.conv.%s"
+#define TN_GLM_ADAPTER_LINEAR   "adapter.linear.linear.%s"
+#define TN_GLM_ADAPTER_NORM_1   "adapter.linear.norm1.%s"
+#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
+#define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
+#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
+#define TN_GLM_BOI_W            "adapter.boi"
+#define TN_GLM_EOI_W            "adapter.eoi"
+
+enum projector_type {
+    PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
+    PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
+    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_GLM_EDGE,
+    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_UNKNOWN,
+};
+
+static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
+    { PROJECTOR_TYPE_MLP,       "mlp" },
+    { PROJECTOR_TYPE_LDP,       "ldp" },
+    { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
+    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
+    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+};
+
+static projector_type clip_projector_type_from_string(const std::string & str) {
+    for (const auto & pair : PROJECTOR_TYPE_NAMES) {
+        if (pair.second == str) {
+            return pair.first;
+        }
+    }
+    return PROJECTOR_TYPE_UNKNOWN;
+}
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+//
+// logging
+//
+
+static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+struct clip_logger_state {
+    ggml_log_level verbosity_thold;
+    ggml_log_callback log_callback;
+    void * log_callback_user_data;
+};
+
+extern struct clip_logger_state g_logger_state;
+
+static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
+    if (format == NULL) {
+        return;
+    }
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        free(buffer2);
+    }
+    va_end(args_copy);
+}
+
+static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    clip_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+#define LOG_TMPL(level, ...) \
+    do { \
+        if ((level) >= g_logger_state.verbosity_thold) { \
+            clip_log_internal((level), __VA_ARGS__); \
+        } \
+    } while (0)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  __VA_ARGS__)
+
+//
+// cpp wrappers
+//
+
+// wrapper for clip_image_size
+struct clip_image_size_deleter {
+    void operator()(clip_image_size * val) { clip_image_size_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
+// wrapper for clip_image_u8
+struct clip_image_u8_deleter {
+    void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
+};
+typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
+
+// wrapper for clip_image_f32
+struct clip_image_f32_deleter {
+    void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
+
+struct clip_image_u8_batch {
+    std::vector<clip_image_u8_ptr> entries;
+};
+
+struct clip_image_f32_batch {
+    std::vector<clip_image_f32_ptr> entries;
+};
+
+//
+// common utils
+//
+
+static std::string string_format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), buf.size());
+}
+
+static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+// split string by a `std::string delim` instead of `char delim`
+static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t pos = 0;
+    std::string token;
+    while ((pos = s.find(delimiter)) != std::string::npos) {
+        token = s.substr(0, pos);
+        tokens.push_back(token);
+        s.erase(0, pos + delimiter.length());
+    }
+    tokens.push_back(s);
+    return tokens;
+}
+
+//
+// gguf utils
+//
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return string_format("unknown type %d", type);
+    }
+}
+
+static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        string_replace_all(val, "\\", "\\\\");
+                        string_replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}
+
+//
+// API used internally with mtmd
+//
+
+projector_type clip_get_projector_type(const struct clip_ctx * ctx);

llama/llama.cpp/examples/llava/clip.h

@@ -1,6 +1,7 @@
 #ifndef CLIP_H
 #define CLIP_H
 
+#include "ggml.h"
 #include <stddef.h>
 #include <stdint.h>
 
@@ -29,32 +30,34 @@ struct clip_image_size {
     int height;
 };
 
-struct clip_image_u8_batch {
-    struct clip_image_u8 * data;
-    size_t size;
+struct clip_image_f32;
+struct clip_image_u8_batch;
+struct clip_image_f32_batch;
+
+struct clip_context_params {
+    bool use_gpu;
+    enum ggml_log_level verbosity;
 };
 
-struct clip_image_f32_batch {
-    struct clip_image_f32 * data;
-    size_t size;
-};
+// deprecated, use clip_init
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
 
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params);
 
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
 
-CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
 
 // TODO: should be enum, not string
 CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
 
 CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
 CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
@@ -64,15 +67,32 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
 CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
 
-CLIP_API struct clip_image_size * clip_image_size_init();
-CLIP_API struct clip_image_u8  * clip_image_u8_init ();
-CLIP_API struct clip_image_f32 * clip_image_f32_init();
+CLIP_API struct clip_image_size      * clip_image_size_init();
+CLIP_API struct clip_image_u8        * clip_image_u8_init ();
+CLIP_API struct clip_image_f32       * clip_image_f32_init();
+CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava
 
+// nx, ny are the output image dimensions
+CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
+
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
 CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
+// use for accessing underlay data of clip_image_f32_batch
+CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
+CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
+CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
+CLIP_API struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
+
+/**
+ * Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
+ * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
+ */
+CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
+
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
@@ -89,10 +109,16 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
+
+CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
 
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
+
 #ifdef __cplusplus
 }
 #endif

llama/llama.cpp/examples/llava/llava.cpp

@@ -10,6 +10,7 @@
 #include <cstring>
 #include <limits>
 #include <vector>
+#include <memory>
 
 #if defined(LLAVA_LOG_OFF)
 #   define LOG_INF(...)
@@ -45,6 +46,17 @@ struct clip_image_grid_shape {
     int second;
 };
 
+// convenience cpp wrapper
+struct clip_image_f32_batch_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_f32_batch, clip_image_f32_batch_deleter> clip_image_f32_batch_ptr;
+
+struct clip_image_size_deleter {
+    void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); }
+};
+typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
+
 /**
  * Selects the best resolution from a list of possible resolutions based on the original size.
  *
@@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
         struct ggml_context * ctx;
     } model;
 
-    const int32_t image_size = clip_image_size(ctx_clip);
-    const int32_t patch_size = clip_patch_size(ctx_clip);
+    const int32_t image_size = clip_get_image_size(ctx_clip);
+    const int32_t patch_size = clip_get_patch_size(ctx_clip);
 
     int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
 
@@ -216,7 +228,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     return true;
 }
 
-static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
+static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) {
     int width = image->nx;
     int height = image->ny;
     int num_patches = (height / patch_size) * (width / patch_size);
@@ -246,12 +258,9 @@ static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int p
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
     // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
-    clip_image_f32_batch img_res_v;
-    img_res_v.size = 0;
-    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+    clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init());
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) {
         LOG_ERR("%s: unable to preprocess image\n", __func__);
-        delete[] img_res_v.data;
         return false;
     }
 
@@ -259,66 +268,83 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
     const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
 
+    const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get());
+
     if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        struct clip_image_size * load_image_size = clip_image_size_init();
+        image_embd_v.resize(n_imgs);
+        clip_image_size load_image_size;
 
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        for (size_t i = 0; i < n_imgs; i++) {
             const int64_t t_img_enc_step_start_us = ggml_time_us();
-            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            int patch_size=14;
-            load_image_size->width = img_res_v.data[i].nx;
-            load_image_size->height = img_res_v.data[i].ny;
-            clip_add_load_image_size(ctx_clip, load_image_size);
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            int patch_size = 14;
+            load_image_size.width = nx;
+            load_image_size.height = ny;
+            clip_add_load_image_size(ctx_clip, &load_image_size);
 
             bool encoded = false;
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             if (clip_is_qwen2vl(ctx_clip)) {
-                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+                encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]);
             }
             else {
-                int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
-                if (has_minicpmv_projector == 2) {
-                    encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
-                }
-                else if (has_minicpmv_projector == 3) {
-                    encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
-                }
+                encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]);
             }
 
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 return false;
             }
             const int64_t t_img_enc_steop_batch_us = ggml_time_us();
-            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
+            int nx = clip_image_f32_batch_nx(img_res_v.get(), i);
+            int ny = clip_image_f32_batch_ny(img_res_v.get(), i);
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             std::memcpy(
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
-                clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
+                clip_embd_nbytes_by_img(ctx_clip, nx, ny));
+            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
             free(image_embd_v[i]);
         }
         image_embd_v.clear();
-        load_image_size->width = img->nx;
-        load_image_size->height = img->ny;
+        load_image_size.width = img->nx;
+        load_image_size.height = img->ny;
+        clip_add_load_image_size(ctx_clip, &load_image_size);
+        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height);
+    }
+    else if (clip_is_glm(ctx_clip)){
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        load_image_size->width  = clip_image_f32_batch_nx(img_res_v.get(), 0);
+        load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0);
         clip_add_load_image_size(ctx_clip, load_image_size);
-        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+        int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2);
+        *n_img_pos = (pos * pos + 2);
+        if (!encoded){
+            LOG_ERR("Unable to encode image \n");
+            return false;
+        }
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
         *n_img_pos = clip_n_patches(ctx_clip);
-        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
-        delete[] img_res_v.data;
+        clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
 
@@ -329,31 +355,28 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // spatial_unpad llava-1.6 type embedding
         // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
         std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        for (size_t i = 0; i < img_res_v.size; i++) {
+        image_embd_v.resize(n_imgs);
+        for (size_t i = 0; i < n_imgs; i++) {
+            clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i);
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
+        const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip);
 
         std::vector<std::pair<int, int>> grid_pinpoints;
-        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+        for (size_t i = 0; i < num_gridpoints; i += 2) {
             grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
         }
 
-        // free all img_res_v - not needed anymore
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
-
-        const int32_t image_size = clip_image_size(ctx_clip);
+        const int32_t image_size = clip_get_image_size(ctx_clip);
 
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 
@@ -384,7 +407,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
 
 bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
         // make sure that the correct mmproj was used, i.e., compare apples to apples
-    int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
+    int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
         LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
@@ -394,10 +417,14 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 }
 
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
-    int num_max_patches = 6;
+    // Granite vision uses up to 10 patches + base patch
+    int num_max_patches = 11;
     if (clip_is_minicpmv(ctx_clip)) {
         num_max_patches = 10;
     }
+    if (clip_is_glm(ctx_clip)) {
+        num_max_patches = 1;
+    }
     float * image_embd;
     if (clip_is_qwen2vl(ctx_clip)) {
         // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
@@ -457,7 +484,7 @@ struct llava_embd_batch {
 };
 
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
-    int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
+    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llama));
 
     for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
         int n_eval = image_embed->n_image_pos - i;

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

@@ -9,7 +9,7 @@
 #include "llama.h"
 
 struct llama_model_deleter {
-    void operator()(llama_model * model) { llama_free_model(model); }
+    void operator()(llama_model * model) { llama_model_free(model); }
 };
 
 struct llama_context_deleter {
@@ -20,11 +20,11 @@ struct llama_sampler_deleter {
     void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
 };
 
-struct llama_lora_adapter_deleter {
-    void operator()(llama_lora_adapter * lora_adapter) { llama_lora_adapter_free(lora_adapter); }
+struct llama_adapter_lora_deleter {
+    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(adapter); }
 };
 
 typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
 typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
 typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;
-typedef std::unique_ptr<llama_lora_adapter, llama_lora_adapter_deleter> llama_lora_adapter_ptr;
+typedef std::unique_ptr<llama_adapter_lora, llama_adapter_lora_deleter> llama_adapter_lora_ptr;

llama/llama.cpp/include/llama.h

@@ -34,7 +34,6 @@
 
 #define LLAMA_DEFAULT_SEED 0xFFFFFFFF
 
-// TODO: use everywhere in the implementation
 #define LLAMA_TOKEN_NULL -1
 
 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
@@ -57,10 +56,11 @@ extern "C" {
     // TODO: show sample usage
     //
 
-    // struct llama_vocab; // TODO: add in the future
+    struct llama_vocab;
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
+    struct llama_kv_cache;
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -106,6 +106,11 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
         LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
         LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM  = 28,
+        LLAMA_VOCAB_PRE_TYPE_GPT4O          = 29,
+        LLAMA_VOCAB_PRE_TYPE_SUPERBPE       = 30,
+        LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
+        LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
+        LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
     };
 
     enum llama_rope_type {
@@ -214,7 +219,7 @@ extern "C" {
         LLAMA_SPLIT_MODE_ROW   = 2, // split layers and KV across GPUs, use tensor parallelism if supported
     };
 
-    // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979)
+    // TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979)
     typedef struct llama_token_data {
         llama_token id; // token id
         float logit;    // log-odds of the token
@@ -277,10 +282,18 @@ extern "C" {
         };
     };
 
+    struct llama_model_tensor_buft_override {
+        const char * pattern;
+        ggml_backend_buffer_type_t buft;
+    };
+
     struct llama_model_params {
         // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
         ggml_backend_dev_t * devices;
 
+        // 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
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
@@ -290,9 +303,6 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
-        // comma separated list of RPC servers to use for offloading
-        const char * rpc_servers;
-
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
         // If it returns false, model loading is immediately aborted.
@@ -312,7 +322,7 @@ extern "C" {
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
-    //       https://github.com/ggerganov/llama.cpp/pull/7544
+    //       https://github.com/ggml-org/llama.cpp/pull/7544
     struct llama_context_params {
         uint32_t n_ctx;             // text context, 0 = from model
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
@@ -325,7 +335,7 @@ extern "C" {
         enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
         enum llama_attention_type    attention_type;    // attention type to use for embeddings
 
-        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
+        // ref: https://github.com/ggml-org/llama.cpp/pull/2054
         float    rope_freq_base;   // RoPE base frequency, 0 = from model
         float    rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
         float    yarn_ext_factor;  // YaRN extrapolation mix factor, negative = from model
@@ -359,17 +369,18 @@ extern "C" {
 
     // model quantization parameters
     typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // token embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;                      // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;               // quantize to this llama_ftype
+        enum ggml_type output_tensor_type;    // output tensor type
+        enum ggml_type token_embedding_type;  // token embeddings tensor type
+        bool allow_requantize;                // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor;          // quantize output.weight
+        bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                            // quantize all tensors to the default type
+        bool keep_split;                      // quantize to the same number of shards
+        void * imatrix;                       // pointer to importance matrix data
+        void * kv_overrides;                  // pointer to vector containing overrides
+        void * tensor_types;                  // pointer to vector containing tensor types
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -388,11 +399,10 @@ extern "C" {
     } llama_chat_message;
 
     // lora adapter
-    // TODO: rename to llama_adapter_lora
-    struct llama_lora_adapter;
+    struct llama_adapter_lora;
 
     // Helpers for getting default parameters
-    // TODO: update API to start accepting pointers to params structs (https://github.com/ggerganov/llama.cpp/discussions/9172)
+    // TODO: update API to start accepting pointers to params structs (https://github.com/ggml-org/llama.cpp/discussions/9172)
     LLAMA_API struct llama_model_params          llama_model_default_params(void);
     LLAMA_API struct llama_context_params        llama_context_default_params(void);
     LLAMA_API struct llama_sampler_chain_params  llama_sampler_chain_default_params(void);
@@ -403,31 +413,53 @@ extern "C" {
     // Call once at the start of the program
     LLAMA_API void llama_backend_init(void);
 
+    // Call once at the end of the program - currently only used for MPI
+    LLAMA_API void llama_backend_free(void);
+
     //optional:
     LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
     // Optional: an auto threadpool gets created in ggml if not passed explicitly
     LLAMA_API void llama_attach_threadpool(
-               struct   llama_context * ctx,
-            ggml_threadpool_t   threadpool,
-            ggml_threadpool_t   threadpool_batch);
+            struct llama_context * ctx,
+               ggml_threadpool_t   threadpool,
+               ggml_threadpool_t   threadpool_batch);
+
     LLAMA_API void llama_detach_threadpool(struct llama_context * ctx);
 
-    // Call once at the end of the program - currently only used for MPI
-    LLAMA_API void llama_backend_free(void);
+    DEPRECATED(LLAMA_API struct llama_model * llama_load_model_from_file(
+                             const char * path_model,
+              struct llama_model_params   params),
+            "use llama_model_load_from_file instead");
 
-    LLAMA_API struct llama_model * llama_load_model_from_file(
+    // Load the model from a file
+    // If the file is split into multiple parts, the file name must follow this pattern: <name>-%05d-of-%05d.gguf
+    // If the split file name does not follow this pattern, use llama_model_load_from_splits
+    LLAMA_API struct llama_model * llama_model_load_from_file(
                              const char * path_model,
               struct llama_model_params   params);
 
-    // TODO: rename to llama_model_free
-    LLAMA_API void llama_free_model(struct llama_model * model);
+    // Load the model from multiple splits (support custom naming scheme)
+    // The paths must be in the correct order
+    LLAMA_API struct llama_model * llama_model_load_from_splits(
+                             const char ** paths,
+                                 size_t    n_paths,
+              struct llama_model_params    params);
 
-    // TODO: rename to llama_init_from_model
-    LLAMA_API struct llama_context * llama_new_context_with_model(
+    DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
+            "use llama_model_free instead");
+
+    LLAMA_API void llama_model_free(struct llama_model * model);
+
+    LLAMA_API struct llama_context * llama_init_from_model(
                      struct llama_model * model,
             struct llama_context_params   params);
 
+    DEPRECATED(LLAMA_API struct llama_context * llama_new_context_with_model(
+                     struct llama_model * model,
+            struct llama_context_params   params),
+            "use llama_init_from_model instead");
+
     // TODO (jmorganca): this should most likely be passed in as part of a batch
     // and not set on the context for all batches.
     LLAMA_API void llama_set_cross_attention(struct llama_context * ctx, bool cross_attn_state);
@@ -449,20 +481,32 @@ extern "C" {
     LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);
 
-    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
-    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_head     (const struct llama_model * model);
+    DEPRECATED(LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model), "use llama_model_n_ctx_train instead");
+    DEPRECATED(LLAMA_API int32_t llama_n_embd     (const struct llama_model * model), "use llama_model_n_embd instead");
+    DEPRECATED(LLAMA_API int32_t llama_n_layer    (const struct llama_model * model), "use llama_model_n_layer instead");
+    DEPRECATED(LLAMA_API int32_t llama_n_head     (const struct llama_model * model), "use llama_model_n_head instead");
 
-    LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
+    DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");
 
-    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model * model);
+    LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
+    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type
+
+    LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
+    LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);
+
+    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_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);
 
     // Get the model's RoPE frequency scaling factor
-    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
+    LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
+
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_vocab * vocab);
+
+    LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab);
 
     // Functions to access the model's GGUF metadata scalar values
     // - The functions return the length of the string on success, or -1 on failure
@@ -488,6 +532,10 @@ extern "C" {
     // Returns the total size of all the tensors in the model in bytes
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
 
+    // Get the default chat template. Returns nullptr if not available
+    // If name is NULL, returns the default chat template
+    LLAMA_API const char * llama_model_chat_template(const struct llama_model * model, const char * name);
+
     // Returns the total number of parameters in the model
     LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
 
@@ -515,34 +563,31 @@ extern "C" {
     //
 
     // Load a LoRA adapter from file
-    // TODO: rename to llama_adapter_lora_init
-    LLAMA_API struct llama_lora_adapter * llama_lora_adapter_init(
+    LLAMA_API struct llama_adapter_lora * llama_adapter_lora_init(
             struct llama_model * model,
             const char * path_lora);
 
+    // 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);
+
+    // The following functions operate on a llama_context, hence the naming: llama_verb_...
+
     // Add a loaded LoRA adapter to given context
     // This will not modify model's weight
-    // TODO: rename to llama_set_adapter_lora
-    LLAMA_API int32_t llama_lora_adapter_set(
+    LLAMA_API int32_t llama_set_adapter_lora(
             struct llama_context * ctx,
-            struct llama_lora_adapter * adapter,
+            struct llama_adapter_lora * adapter,
             float scale);
 
     // Remove a specific LoRA adapter from given context
     // Return -1 if the adapter is not present in the context
-    // TODO: rename to llama_rm_adapter_lora
-    LLAMA_API int32_t llama_lora_adapter_remove(
+    LLAMA_API int32_t llama_rm_adapter_lora(
             struct llama_context * ctx,
-            struct llama_lora_adapter * adapter);
+            struct llama_adapter_lora * adapter);
 
     // Remove all LoRA adapters from given context
-    // TODO: rename to llama_clear_adapter_lora
-    LLAMA_API void llama_lora_adapter_clear(struct llama_context * ctx);
-
-    // Manually free a LoRA adapter
-    // Note: loaded adapters will be free when the associated model is deleted
-    // TODO: rename to llama_adapter_lora_free
-    LLAMA_API void llama_lora_adapter_free(struct llama_lora_adapter * adapter);
+    LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx);
 
     // Apply a loaded control vector to a llama_context, or if data is NULL, clear
     // the currently loaded vector.
@@ -550,9 +595,8 @@ extern "C" {
     // to an n_embd x n_layers buffer starting from layer 1.
     // il_start and il_end are the layer range the vector should apply to (both inclusive)
     // See llama_control_vector_load in common to load a control vector.
-    // TODO: rename to llama_adapter_cvec_apply
-    LLAMA_API int32_t llama_control_vector_apply(
-            struct llama_context * lctx,
+    LLAMA_API int32_t llama_apply_adapter_cvec(
+            struct llama_context * ctx,
                      const float * data,
                           size_t   len,
                          int32_t   n_embd,
@@ -563,7 +607,7 @@ extern "C" {
     // KV cache
     //
 
-    // TODO: remove llama_kv_cache_view_* API
+    // TODO: start using struct llama_kv_cache
 
     // Information associated with an individual cell in the KV cache view.
     struct llama_kv_cache_view_cell {
@@ -618,13 +662,19 @@ extern "C" {
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
     // If a KV cell has multiple sequences assigned to it, it will be counted multiple times
-    LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx),
+            "use llama_kv_self_n_tokens instead");
 
     // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
-    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);
+    LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx),
+            "use llama_kv_self_used_cells instead");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_cache_clear(
+    LLAMA_API void llama_kv_self_clear(
             struct llama_context * ctx);
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
@@ -632,7 +682,7 @@ extern "C" {
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_cache_seq_rm(
+    LLAMA_API bool llama_kv_self_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -642,7 +692,7 @@ extern "C" {
     // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_cp(
+    LLAMA_API void llama_kv_self_seq_cp(
             struct llama_context * ctx,
                     llama_seq_id   seq_id_src,
                     llama_seq_id   seq_id_dst,
@@ -650,17 +700,17 @@ extern "C" {
                        llama_pos   p1);
 
     // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_cache_seq_keep(
+    LLAMA_API void llama_kv_self_seq_keep(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_add(
+    LLAMA_API void llama_kv_self_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -670,10 +720,10 @@ extern "C" {
     // Integer division of the positions by factor of `d > 1`
     // If the KV cache is RoPEd, the KV data is updated accordingly:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
+    //   - explicitly with llama_kv_self_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_div(
+    LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -681,24 +731,76 @@ extern "C" {
                              int   d);
 
     // Returns the largest position present in the KV cache for the specified sequence
-    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache
-    //       how to avoid this?
+                     llama_seq_id   seq_id);
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_cache_update()
-    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
-
-    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
+    //   - explicitly with llama_kv_self_update()
+    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
 
     // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+
+    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_clear(
+            struct llama_context * ctx),
+            "use llama_kv_self_clear instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_rm instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id_src,
+                    llama_seq_id   seq_id_dst,
+                       llama_pos   p0,
+                       llama_pos   p1),
+            "use llama_kv_self_seq_cp instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_keep instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_add(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                       llama_pos   delta),
+            "use llama_kv_self_seq_add instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d),
+            "use llama_kv_self_seq_div instead");
+
+    DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id),
+            "use llama_kv_self_seq_pos_max instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx),
+            "use llama_kv_self_defrag instead");
+
+    DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx),
+            "use llama_kv_self_can_shift instead");
+
+    DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx),
+            "use llama_kv_self_update instead");
+
 
     //
     // State / sessions
@@ -862,6 +964,10 @@ extern "C" {
     // If set to true, the model will only attend to the past tokens
     LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
 
+    // Set whether the model is in warmup mode or not
+    // If true, all model tensors are activated during llama_decode() to load and cache their weights.
+    LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup);
+
     // Set abort callback
     LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
 
@@ -908,41 +1014,60 @@ extern "C" {
     // Vocab
     //
 
-    LLAMA_API const char * llama_token_get_text(const struct llama_model * model, llama_token token);
+    LLAMA_API const char * llama_vocab_get_text(const struct llama_vocab * vocab, llama_token token);
 
-    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);
+    LLAMA_API float llama_vocab_get_score(const struct llama_vocab * vocab, llama_token token);
 
-    LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token);
+    LLAMA_API enum llama_token_attr llama_vocab_get_attr(const struct llama_vocab * vocab, llama_token token);
 
     // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
-    LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
+    LLAMA_API bool llama_vocab_is_eog(const struct llama_vocab * vocab, llama_token token);
 
     // Identify if Token Id is a control token or a render-able token
-    LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
+    LLAMA_API bool llama_vocab_is_control(const struct llama_vocab * vocab, llama_token token);
 
     // Special tokens
-    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
-    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
-    LLAMA_API llama_token llama_token_eot(const struct llama_model * model); // end-of-turn
-    LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
-    LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
-    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line
-    LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding
+    LLAMA_API llama_token llama_vocab_bos(const struct llama_vocab * vocab); // beginning-of-sentence
+    LLAMA_API llama_token llama_vocab_eos(const struct llama_vocab * vocab); // end-of-sentence
+    LLAMA_API llama_token llama_vocab_eot(const struct llama_vocab * vocab); // end-of-turn
+    LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator
+    LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line
+    LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding
 
-    LLAMA_API bool llama_add_bos_token(const struct llama_model * model);
-    LLAMA_API bool llama_add_eos_token(const struct llama_model * model);
+    LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
+    LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
 
-    // infill tokens
-    DEPRECATED(LLAMA_API llama_token llama_token_prefix(const struct llama_model * model), "use llama_token_fim_pre instead");
-    DEPRECATED(LLAMA_API llama_token llama_token_middle(const struct llama_model * model), "use llama_token_fim_mid instead");
-    DEPRECATED(LLAMA_API llama_token llama_token_suffix(const struct llama_model * model), "use llama_token_fim_suf instead");
+    LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab);
+    LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab);
+    LLAMA_API llama_token llama_vocab_fim_mid(const struct llama_vocab * vocab);
+    LLAMA_API llama_token llama_vocab_fim_pad(const struct llama_vocab * vocab);
+    LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab);
+    LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab);
 
-    LLAMA_API llama_token llama_token_fim_pre(const struct llama_model * model);
-    LLAMA_API llama_token llama_token_fim_suf(const struct llama_model * model);
-    LLAMA_API llama_token llama_token_fim_mid(const struct llama_model * model);
-    LLAMA_API llama_token llama_token_fim_pad(const struct llama_model * model);
-    LLAMA_API llama_token llama_token_fim_rep(const struct llama_model * model);
-    LLAMA_API llama_token llama_token_fim_sep(const struct llama_model * model);
+    DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead");
+    DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead");
+    DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead");
+    DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead");
+    DEPRECATED(LLAMA_API bool llama_token_is_control(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_control instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_bos(const struct llama_vocab * vocab), "use llama_vocab_bos instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_eos(const struct llama_vocab * vocab), "use llama_vocab_eos instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_eot(const struct llama_vocab * vocab), "use llama_vocab_eot instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_cls(const struct llama_vocab * vocab), "use llama_vocab_cls instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_sep(const struct llama_vocab * vocab), "use llama_vocab_sep instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_nl (const struct llama_vocab * vocab), "use llama_vocab_nl instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_pad(const struct llama_vocab * vocab), "use llama_vocab_pad instead");
+    DEPRECATED(LLAMA_API bool llama_add_bos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_bos instead");
+    DEPRECATED(LLAMA_API bool llama_add_eos_token(const struct llama_vocab * vocab), "use llama_vocab_get_add_eos instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_pre(const struct llama_vocab * vocab), "use llama_vocab_fim_pre instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_suf(const struct llama_vocab * vocab), "use llama_vocab_fim_suf instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_mid(const struct llama_vocab * vocab), "use llama_vocab_fim_mid instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_pad(const struct llama_vocab * vocab), "use llama_vocab_fim_pad instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_rep(const struct llama_vocab * vocab), "use llama_vocab_fim_rep instead");
+    DEPRECATED(LLAMA_API llama_token llama_token_fim_sep(const struct llama_vocab * vocab), "use llama_vocab_fim_sep instead");
+
+    // CLS is equivalent to BOS
+    DEPRECATED(LLAMA_API llama_token llama_vocab_cls(const struct llama_vocab * vocab), // classification
+            "use llama_vocab_bos instead");
 
     //
     // Tokenization
@@ -958,7 +1083,7 @@ extern "C" {
     /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
     ///                      as plaintext. Does not insert a leading space.
     LLAMA_API int32_t llama_tokenize(
-        const struct llama_model * model,
+        const struct llama_vocab * vocab,
                       const char * text,
                          int32_t   text_len,
                      llama_token * tokens,
@@ -972,7 +1097,7 @@ extern "C" {
     // User can skip up to 'lstrip' leading spaces before copying (useful when encoding/decoding multiple tokens with 'add_space_prefix')
     // @param special If true, special tokens are rendered in the output.
     LLAMA_API int32_t llama_token_to_piece(
-              const struct llama_model * model,
+              const struct llama_vocab * vocab,
                            llama_token   token,
                                   char * buf,
                                int32_t   length,
@@ -986,7 +1111,7 @@ extern "C" {
     /// @param remove_special Allow to remove BOS and EOS tokens if model is configured to do so.
     /// @param unparse_special If true, special tokens are rendered in the output.
     LLAMA_API int32_t llama_detokenize(
-        const struct llama_model * model,
+        const struct llama_vocab * vocab,
                const llama_token * tokens,
                          int32_t   n_tokens,
                             char * text,
@@ -1000,7 +1125,7 @@ extern "C" {
 
     /// Apply chat template. Inspired by hf apply_chat_template() on python.
     /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
-    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
+    /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
     /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
     /// @param chat Pointer to a list of multiple llama_chat_message
     /// @param n_msg Number of llama_chat_message in this chat
@@ -1009,7 +1134,6 @@ extern "C" {
     /// @param length The size of the allocated buffer
     /// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
     LLAMA_API int32_t llama_chat_apply_template(
-              const struct llama_model * model,
                             const char * tmpl,
        const struct llama_chat_message * chat,
                                 size_t   n_msg,
@@ -1057,7 +1181,6 @@ 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).
-    // TODO: in the future, the entire sampling API that uses llama_model should start using llama_vocab
     //
 
     typedef void * llama_sampler_context_t;
@@ -1076,11 +1199,12 @@ extern "C" {
     };
 
     struct llama_sampler {
-        struct llama_sampler_i  * iface;
-        llama_sampler_context_t   ctx;
+        const struct llama_sampler_i * iface;
+        llama_sampler_context_t        ctx;
     };
 
     // 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 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);
@@ -1110,7 +1234,7 @@ extern "C" {
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first.
     DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax    (void),
-        "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)");
+        "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
@@ -1118,7 +1242,7 @@ extern "C" {
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     LLAMA_API struct llama_sampler * llama_sampler_init_top_p      (float   p, size_t min_keep);
 
-    /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
+    /// @details Minimum P sampling as described in https://github.com/ggml-org/llama.cpp/pull/3841
     LLAMA_API struct llama_sampler * llama_sampler_init_min_p      (float   p, size_t min_keep);
 
     /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
@@ -1133,6 +1257,9 @@ extern "C" {
     /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335
     LLAMA_API struct llama_sampler * llama_sampler_init_xtc        (float   p, float   t,     size_t min_keep, uint32_t seed);
 
+    /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641
+    LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float   n);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -1156,11 +1283,39 @@ extern "C" {
                                float   tau,
                                float   eta);
 
+    /// @details Intializes a GBNF grammar, see grammars/README.md for details.
+    /// @param vocab The vocabulary that this grammar will be used with.
+    /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails.
+    /// @param grammar_root The name of the start symbol for the grammar.
     LLAMA_API struct llama_sampler * llama_sampler_init_grammar(
-            const struct llama_model * model,
+            const struct llama_vocab * vocab,
                           const char * grammar_str,
                           const char * grammar_root);
 
+    DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
+            const struct llama_vocab * vocab,
+                          const char * grammar_str,
+                          const char * grammar_root,
+                         const char ** trigger_words,
+                                size_t num_trigger_words,
+                   const llama_token * trigger_tokens,
+                                size_t num_trigger_tokens),
+        "use llama_sampler_init_grammar_lazy_patterns instead");
+
+
+    /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639
+    /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group.
+    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included.
+    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
+
     /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
                              int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
@@ -1169,8 +1324,9 @@ extern "C" {
                                float   penalty_present); // 0.0 = disabled
 
     ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
-    LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
-            const struct llama_model *  model,
+    LLAMA_API struct llama_sampler * llama_sampler_init_dry(
+            const struct llama_vocab *  vocab,
+                             int32_t    n_ctx_train,
                                float    dry_multiplier,
                                float    dry_base,
                              int32_t    dry_allowed_length,
@@ -1204,7 +1360,7 @@ extern "C" {
     // 3. discard non-EOG tokens with low prob
     // 4. if no tokens are left -> pick EOT
     //
-    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_model * model);
+    LLAMA_API struct llama_sampler * llama_sampler_init_infill(const struct llama_vocab * vocab);
 
     // Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
     LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);

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

@@ -1,15 +1,16 @@
 #include "llama-adapter.h"
 
+#include "llama-impl.h"
+#include "llama-mmap.h"
 #include "llama-model.h"
 
-#include <algorithm>
 #include <map>
 #include <cassert>
 #include <stdexcept>
 
 // vec
 
-struct ggml_tensor * llama_control_vector::tensor_for(int il) const {
+ggml_tensor * llama_adapter_cvec::tensor_for(int il) const {
     if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
         return nullptr;
     }
@@ -17,7 +18,7 @@ struct ggml_tensor * llama_control_vector::tensor_for(int il) const {
     return tensors[il];
 }
 
-struct ggml_tensor * llama_control_vector::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const {
+ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const {
     ggml_tensor * layer_dir = tensor_for(il);
     if (layer_dir != nullptr) {
         cur = ggml_add(ctx, cur, layer_dir);
@@ -26,19 +27,19 @@ struct ggml_tensor * llama_control_vector::apply_to(struct ggml_context * ctx, s
     return cur;
 }
 
-static bool llama_control_vector_init(struct llama_control_vector & cvec, const llama_model & model) {
+bool llama_adapter_cvec::init(const llama_model & model) {
     const auto & hparams = model.hparams;
 
-    GGML_ASSERT(cvec.tensors.empty());
-    GGML_ASSERT(cvec.ctxs.empty());
-    GGML_ASSERT(cvec.bufs.empty());
+    GGML_ASSERT(tensors.empty());
+    GGML_ASSERT(ctxs.empty());
+    GGML_ASSERT(bufs.empty());
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
     auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * {
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ hparams.n_layer*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -50,7 +51,7 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
             }
 
             ctx_map[buft] = ctx;
-            cvec.ctxs.emplace_back(ctx);
+            ctxs.emplace_back(ctx);
 
             return ctx;
         }
@@ -59,21 +60,21 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
     };
 
     // make tensors
-    cvec.tensors.reserve(hparams.n_layer);
-    cvec.tensors.push_back(nullptr); // there's never a tensor for layer 0
+    tensors.reserve(hparams.n_layer);
+    tensors.push_back(nullptr); // there's never a tensor for layer 0
     for (size_t il = 1; il < hparams.n_layer; il++) {
-        ggml_backend_buffer_type_t buft = llama_model_select_buft(model, il);
+        ggml_backend_buffer_type_t buft = model.select_buft(il);
         ggml_context * ctx = ctx_for_buft(buft);
         if (!ctx) {
             LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__);
             return false;
         }
         ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
-        cvec.tensors.push_back(tensor);
+        tensors.push_back(tensor);
     }
 
     // allocate tensors / buffers and zero
-    cvec.bufs.reserve(ctx_map.size());
+    bufs.reserve(ctx_map.size());
     for (auto it : ctx_map) {
         ggml_backend_buffer_type_t buft = it.first;
         ggml_context * ctx = it.second;
@@ -83,14 +84,13 @@ static bool llama_control_vector_init(struct llama_control_vector & cvec, const
             return false;
         }
         ggml_backend_buffer_clear(buf, 0);
-        cvec.bufs.emplace_back(buf);
+        bufs.emplace_back(buf);
     }
 
     return true;
 }
 
-int32_t llama_control_vector_apply(
-        struct llama_control_vector & cvec,
+bool llama_adapter_cvec::apply(
         const llama_model & model,
         const float * data,
         size_t len,
@@ -101,40 +101,40 @@ int32_t llama_control_vector_apply(
 
     if (data == nullptr) {
         // disable the current control vector (but leave allocated for later)
-        cvec.layer_start = -1;
-        cvec.layer_end   = -1;
-        return 0;
+        layer_start = -1;
+        layer_end   = -1;
+        return true;
     }
 
     if (n_embd != (int) hparams.n_embd) {
         LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__);
-        return 1;
+        return false;
     }
 
-    if (cvec.tensors.empty()) {
-        if (!llama_control_vector_init(cvec, model)) {
-            return 1;
+    if (tensors.empty()) {
+        if (!init(model)) {
+            return false;
         }
     }
 
-    cvec.layer_start = il_start;
-    cvec.layer_end   = il_end;
+    layer_start = il_start;
+    layer_end   = il_end;
 
     for (size_t il = 1; il < hparams.n_layer; il++) {
-        assert(cvec.tensors[il] != nullptr);
+        assert(tensors[il] != nullptr);
 
         const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present
         if (off + n_embd <= len) {
-            ggml_backend_tensor_set(cvec.tensors[il], data + off, 0, n_embd * ggml_element_size(cvec.tensors[il]));
+            ggml_backend_tensor_set(tensors[il], data + off, 0, n_embd * ggml_element_size(tensors[il]));
         }
     }
 
-    return 0;
+    return true;
 }
 
 // lora
 
-llama_lora_weight * llama_lora_adapter::get_weight(struct ggml_tensor * w) {
+llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) {
     const std::string name(w->name);
 
     const auto pos = ab_map.find(name);
@@ -145,15 +145,11 @@ llama_lora_weight * llama_lora_adapter::get_weight(struct ggml_tensor * w) {
     return nullptr;
 }
 
-void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
-    delete adapter;
-}
-
-static void llama_lora_adapter_init_impl(struct llama_model & model, const char * path_lora, struct llama_lora_adapter & adapter) {
+static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) {
     LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora);
 
     ggml_context * ctx_init;
-    struct gguf_init_params meta_gguf_params = {
+    gguf_init_params meta_gguf_params = {
         /* .no_alloc = */ true,
         /* .ctx      = */ &ctx_init,
     };
@@ -204,7 +200,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             // add a new context
-            struct ggml_init_params params = {
+            ggml_init_params params = {
                 /*.mem_size   =*/ n_tensors*ggml_tensor_overhead(),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
@@ -221,7 +217,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     };
 
     // bundle lora_a and lora_b into pairs
-    std::map<std::string, llama_lora_weight> ab_map;
+    std::map<std::string, llama_adapter_lora_weight> ab_map;
     auto str_endswith = [](const std::string & str, const std::string & suffix) {
         return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
     };
@@ -231,52 +227,100 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
         if (str_endswith(name, ".lora_a")) {
             replace_all(name, ".lora_a", "");
             if (ab_map.find(name) == ab_map.end()) {
-                ab_map[name] = llama_lora_weight(cur, nullptr);
+                ab_map[name] = llama_adapter_lora_weight(cur, nullptr);
             } else {
                 ab_map[name].a = cur;
             }
         } else if (str_endswith(name, ".lora_b")) {
             replace_all(name, ".lora_b", "");
             if (ab_map.find(name) == ab_map.end()) {
-                ab_map[name] = llama_lora_weight(nullptr, cur);
+                ab_map[name] = llama_adapter_lora_weight(nullptr, cur);
             } else {
                 ab_map[name].b = cur;
             }
+        } else if (str_endswith(name, "_norm.weight")) {
+            // TODO: add support for norm vector
+            // for now, we don't really care because most adapters still work fine without it
+            continue;
         } else {
             throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
         }
     }
 
+    // get extra buffer types of the CPU
+    // TODO: a more general solution for non-CPU extra buft should be imlpemented in the future
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/12593#pullrequestreview-2718659948
+    std::vector<ggml_backend_buffer_type_t> buft_extra;
+    {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_extra.emplace_back(*extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    }
+
     // add tensors
     for (auto & it : ab_map) {
         const std::string & name = it.first;
-        llama_lora_weight & w = it.second;
+        llama_adapter_lora_weight & w = it.second;
+        bool is_token_embd = str_endswith(name, "token_embd.weight");
 
         if (!w.a || !w.b) {
             throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
         }
 
         // device buft and device ctx
-        auto * model_tensor = llama_model_get_tensor(model, name.c_str());
+        const auto * model_tensor = model.get_tensor(name.c_str());
         if (!model_tensor) {
-            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
+            throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
         }
 
-        struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
-        // validate tensor shape
-        if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
-            throw std::runtime_error("tensor '" + name + "' has incorrect shape");
+        auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer);
+
+        // do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case
+        for (auto & ex : buft_extra) {
+            if (ex == buft) {
+                LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+                auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+                buft = ggml_backend_dev_buffer_type(cpu_dev);
+
+                break;
+            }
         }
-        if (w.a->ne[1] != w.b->ne[0]) {
-            throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+
+        LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+        ggml_context * dev_ctx = ctx_for_buft(buft);
+        // validate tensor shape
+        if (is_token_embd) {
+            // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()
+            if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) {
+                throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
+            }
+        } else {
+            if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) {
+                throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
+            }
+            if (w.a->ne[1] != w.b->ne[0]) {
+                throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
+            }
         }
 
         // save tensor to adapter
-        struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
-        struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
+        ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a);
+        ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b);
         ggml_set_name(tensor_a, w.a->name);
         ggml_set_name(tensor_b, w.b->name);
-        adapter.ab_map[name] = llama_lora_weight(tensor_a, tensor_b);
+        adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b);
     }
 
     // allocate tensors / buffers and zero
@@ -299,7 +343,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     {
         llama_file gguf_file(path_lora, "rb");
         std::vector<uint8_t> read_buf;
-        auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) {
+        auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) {
             size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name));
             size_t size = ggml_nbytes(orig);
             read_buf.resize(size);
@@ -318,11 +362,11 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
     LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2);
 }
 
-struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model, const char * path_lora) {
-    struct llama_lora_adapter * adapter = new llama_lora_adapter();
+llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) {
+    llama_adapter_lora * adapter = new llama_adapter_lora();
 
     try {
-        llama_lora_adapter_init_impl(*model, path_lora, *adapter);
+        llama_adapter_lora_init_impl(*model, path_lora, *adapter);
         return adapter;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
@@ -332,3 +376,7 @@ struct llama_lora_adapter * llama_lora_adapter_init(struct llama_model * model,
 
     return nullptr;
 }
+
+void llama_adapter_lora_free(llama_adapter_lora * adapter) {
+    delete adapter;
+}

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

@@ -1,66 +1,76 @@
 #pragma once
 
-#include "llama-impl.h"
-#include "llama-hparams.h"
+#include "llama.h"
 
 #include "ggml-cpp.h"
 
+#include <string>
 #include <unordered_map>
 #include <vector>
 
+// TODO: pimpl
+
 //
 // llama_adapter_cvec
 //
 
-// TODO: rename to llama_adapter_cvec
-struct llama_control_vector {
-    std::vector<ggml_context_ptr> ctxs;
-    std::vector<ggml_backend_buffer_ptr> bufs;
+struct llama_adapter_cvec {
+    ggml_tensor * tensor_for(int il) const;
 
-    std::vector<struct ggml_tensor *> tensors; // per layer
+    ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const;
+
+    bool apply(
+            const llama_model & model,
+            const float * data,
+            size_t len,
+            int32_t n_embd,
+            int32_t il_start,
+            int32_t il_end);
+
+private:
+    bool init(const llama_model & model);
 
     int32_t layer_start = -1;
     int32_t layer_end   = -1;
 
-    struct ggml_tensor * tensor_for(int il) const;
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
 
-    struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int  il) const;
+    std::vector<ggml_tensor *> tensors; // per layer
 };
 
-int32_t llama_control_vector_apply(
-        struct llama_control_vector & cvec,
-        const llama_model & model,
-        const float * data,
-        size_t len,
-        int32_t n_embd,
-        int32_t il_start,
-        int32_t il_end);
-
 //
 // llama_adapter_lora
 //
 
-// TODO: rename to llama_adapter_lora_weight
-struct llama_lora_weight {
-    struct ggml_tensor * a = nullptr;
-    struct ggml_tensor * b = nullptr;
+struct llama_adapter_lora_weight {
+    ggml_tensor * a = nullptr;
+    ggml_tensor * b = nullptr;
 
-    llama_lora_weight() = default;
-    llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
+    // get actual scale based on rank and alpha
+    float get_scale(float alpha, float adapter_scale) const {
+        const float rank  = (float) b->ne[0];
+        const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
+        return scale;
+    }
+
+    llama_adapter_lora_weight() = default;
+    llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
 };
 
-// TODO: rename to llama_adapter_lora
-struct llama_lora_adapter {
+struct llama_adapter_lora {
     // map tensor name to lora_a_b
-    std::unordered_map<std::string, struct llama_lora_weight> ab_map;
+    std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
 
     std::vector<ggml_context_ptr> ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
     float alpha;
 
-    llama_lora_adapter() = default;
-    ~llama_lora_adapter() = default;
+    llama_adapter_lora() = default;
+    ~llama_adapter_lora() = default;
 
-    llama_lora_weight * get_weight(struct ggml_tensor * w);
+    llama_adapter_lora_weight * get_weight(ggml_tensor * w);
 };
+
+using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;

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

@@ -7,6 +7,7 @@
 static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_LLAMA,            "llama"            },
     { LLM_ARCH_MLLAMA,           "mllama"           },
+    { LLM_ARCH_LLAMA4,           "llama4"           },
     { LLM_ARCH_DECI,             "deci"             },
     { LLM_ARCH_FALCON,           "falcon"           },
     { LLM_ARCH_GROK,             "grok"             },
@@ -26,8 +27,11 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2,            "qwen2"            },
     { LLM_ARCH_QWEN2MOE,         "qwen2moe"         },
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
+    { LLM_ARCH_QWEN3,            "qwen3"            },
+    { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
     { LLM_ARCH_PHI2,             "phi2"             },
     { LLM_ARCH_PHI3,             "phi3"             },
+    { LLM_ARCH_PHIMOE,           "phimoe"           },
     { LLM_ARCH_PLAMO,            "plamo"            },
     { LLM_ARCH_CODESHELL,        "codeshell"        },
     { LLM_ARCH_ORION,            "orion"            },
@@ -36,6 +40,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_MINICPM3,         "minicpm3"         },
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
+    { LLM_ARCH_GEMMA3,           "gemma3"           },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
     { LLM_ARCH_XVERSE,           "xverse"           },
@@ -50,6 +55,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DEEPSEEK,         "deepseek"         },
     { LLM_ARCH_DEEPSEEK2,        "deepseek2"        },
     { LLM_ARCH_CHATGLM,          "chatglm"          },
+    { LLM_ARCH_GLM4,             "glm4"             },
     { LLM_ARCH_BITNET,           "bitnet"           },
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
@@ -57,11 +63,17 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
+    { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
+    { LLM_ARCH_RWKV7,            "rwkv7"            },
+    { LLM_ARCH_ARWKV7,           "arwkv7"           },
     { LLM_ARCH_GRANITE,          "granite"          },
     { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
     { LLM_ARCH_CHAMELEON,        "chameleon"        },
     { LLM_ARCH_SOLAR,            "solar"            },
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
+    { LLM_ARCH_PLM,              "plm"              },
+    { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -70,6 +82,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_GENERAL_ARCHITECTURE,         "general.architecture"                  },
     { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version"          },
     { LLM_KV_GENERAL_ALIGNMENT,            "general.alignment"                     },
+    { LLM_KV_GENERAL_FILE_TYPE,            "general.file_type"                     },
     { LLM_KV_GENERAL_NAME,                 "general.name"                          },
     { LLM_KV_GENERAL_AUTHOR,               "general.author"                        },
     { LLM_KV_GENERAL_VERSION,              "general.version"                       },
@@ -107,25 +120,33 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
     { LLM_KV_RESIDUAL_SCALE,                    "%s.residual_scale"                    },
     { LLM_KV_EMBEDDING_SCALE,                   "%s.embedding_scale"                   },
+    { LLM_KV_TOKEN_SHIFT_COUNT,                 "%s.token_shift_count"                 },
+    { LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         "%s.interleave_moe_layer_step"         },
 
-    { LLM_KV_ATTENTION_HEAD_COUNT,               "%s.attention.head_count"               },
-    { LLM_KV_ATTENTION_HEAD_COUNT_KV,            "%s.attention.head_count_kv"            },
-    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,           "%s.attention.max_alibi_bias"           },
-    { LLM_KV_ATTENTION_CLAMP_KQV,                "%s.attention.clamp_kqv"                },
-    { LLM_KV_ATTENTION_KEY_LENGTH,               "%s.attention.key_length"               },
-    { LLM_KV_ATTENTION_VALUE_LENGTH,             "%s.attention.value_length"             },
-    { LLM_KV_ATTENTION_LAYERNORM_EPS,            "%s.attention.layer_norm_epsilon"       },
-    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,        "%s.attention.layer_norm_rms_epsilon"   },
-    { LLM_KV_ATTENTION_GROUPNORM_EPS,            "%s.attention.group_norm_epsilon"       },
-    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,         "%s.attention.group_norm_groups"        },
-    { LLM_KV_ATTENTION_CAUSAL,                   "%s.attention.causal"                   },
-    { LLM_KV_ATTENTION_Q_LORA_RANK,              "%s.attention.q_lora_rank"              },
-    { LLM_KV_ATTENTION_KV_LORA_RANK,             "%s.attention.kv_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_SCALE,                    "%s.attention.scale"                    },
-    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,    "%s.attention.block_skip_connection"    },
-    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,   "%s.attention.cross_attention_layers"   },
+    { LLM_KV_ATTENTION_HEAD_COUNT,                   "%s.attention.head_count"                   },
+    { LLM_KV_ATTENTION_HEAD_COUNT_KV,                "%s.attention.head_count_kv"                },
+    { LLM_KV_ATTENTION_MAX_ALIBI_BIAS,               "%s.attention.max_alibi_bias"               },
+    { LLM_KV_ATTENTION_CLAMP_KQV,                    "%s.attention.clamp_kqv"                    },
+    { LLM_KV_ATTENTION_KEY_LENGTH,                   "%s.attention.key_length"                   },
+    { LLM_KV_ATTENTION_VALUE_LENGTH,                 "%s.attention.value_length"                 },
+    { LLM_KV_ATTENTION_LAYERNORM_EPS,                "%s.attention.layer_norm_epsilon"           },
+    { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            "%s.attention.layer_norm_rms_epsilon"       },
+    { LLM_KV_ATTENTION_GROUPNORM_EPS,                "%s.attention.group_norm_epsilon"           },
+    { LLM_KV_ATTENTION_GROUPNORM_GROUPS,             "%s.attention.group_norm_groups"            },
+    { LLM_KV_ATTENTION_CAUSAL,                       "%s.attention.causal"                       },
+    { LLM_KV_ATTENTION_Q_LORA_RANK,                  "%s.attention.q_lora_rank"                  },
+    { LLM_KV_ATTENTION_KV_LORA_RANK,                 "%s.attention.kv_lora_rank"                 },
+    { LLM_KV_ATTENTION_DECAY_LORA_RANK,              "%s.attention.decay_lora_rank"              },
+    { LLM_KV_ATTENTION_ICLR_LORA_RANK,               "%s.attention.iclr_lora_rank"               },
+    { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" },
+    { 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_SCALE,                        "%s.attention.scale"                        },
+    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
+    { LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,       "%s.attention.cross_attention_layers"       },
+    { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
+    { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -179,6 +200,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,      "tokenizer.chat_template.%s"              },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
@@ -222,6 +245,35 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_LLAMA4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_MLLAMA,
         {
@@ -587,6 +639,45 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_QWEN3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_QWEN3MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,        "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,        "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PHI2,
         {
@@ -622,6 +713,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_PHIMOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FACTORS_LONG,  "rope_factors_long" },
+            { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,           "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_PLAMO,
         {
@@ -778,6 +890,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {
@@ -1011,6 +1144,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_Q_B,           "blk.%d.attn_q_b" },
             { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
             { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_K_B,           "blk.%d.attn_k_b" },
+            { LLM_TENSOR_ATTN_V_B,           "blk.%d.attn_v_b" },
             { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
@@ -1027,6 +1162,22 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
         },
     },
+    {
+        LLM_ARCH_PLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_KV_A_MQA,      "blk.%d.attn_kv_a_mqa" },
+            { LLM_TENSOR_ATTN_KV_A_NORM,     "blk.%d.attn_kv_a_norm" },
+            { LLM_TENSOR_ATTN_KV_B,          "blk.%d.attn_kv_b" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_CHATGLM,
         {
@@ -1036,12 +1187,34 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
             { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
         },
     },
+    {
+        LLM_ARCH_GLM4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_BITNET,
         {
@@ -1182,6 +1355,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TIME_MIX_LERP_V,           "blk.%d.time_mix_lerp_v" },
             { LLM_TENSOR_TIME_MIX_LERP_R,           "blk.%d.time_mix_lerp_r" },
             { LLM_TENSOR_TIME_MIX_LERP_G,           "blk.%d.time_mix_lerp_g" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
             { LLM_TENSOR_TIME_MIX_FIRST,            "blk.%d.time_mix_first" },
             { LLM_TENSOR_TIME_MIX_DECAY,            "blk.%d.time_mix_decay" },
             { LLM_TENSOR_TIME_MIX_DECAY_W1,         "blk.%d.time_mix_decay_w1" },
@@ -1199,6 +1373,100 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE,    "blk.%d.channel_mix_receptance" },
         },
     },
+    {
+        LLM_ARCH_RWKV6QWEN2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_LERP_X,           "blk.%d.time_mix_lerp_x" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_FIRST,            "blk.%d.time_mix_first" },
+            { LLM_TENSOR_TIME_MIX_DECAY,            "blk.%d.time_mix_decay" },
+            { LLM_TENSOR_TIME_MIX_DECAY_W1,         "blk.%d.time_mix_decay_w1" },
+            { LLM_TENSOR_TIME_MIX_DECAY_W2,         "blk.%d.time_mix_decay_w2" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_GATE,             "blk.%d.time_mix_gate" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_FFN_NORM,                  "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,                  "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,                  "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
+        },
+    },
+    {
+        LLM_ARCH_RWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_NORM_2,               "blk.%d.attn_norm_2" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_CHANNEL_MIX_LERP_K,        "blk.%d.channel_mix_lerp_k" },
+            { LLM_TENSOR_CHANNEL_MIX_KEY,           "blk.%d.channel_mix_key" },
+            { LLM_TENSOR_CHANNEL_MIX_VALUE,         "blk.%d.channel_mix_value" },
+        },
+    },
+    {
+        LLM_ARCH_ARWKV7,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,                "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,           "token_embd_norm" },
+            { LLM_TENSOR_OUTPUT_NORM,               "output_norm" },
+            { LLM_TENSOR_OUTPUT,                    "output" },
+            { LLM_TENSOR_ATTN_NORM,                 "blk.%d.attn_norm" },
+            { LLM_TENSOR_TIME_MIX_W0,               "blk.%d.time_mix_w0" },
+            { LLM_TENSOR_TIME_MIX_W1,               "blk.%d.time_mix_w1" },
+            { LLM_TENSOR_TIME_MIX_W2,               "blk.%d.time_mix_w2" },
+            { LLM_TENSOR_TIME_MIX_A0,               "blk.%d.time_mix_a0" },
+            { LLM_TENSOR_TIME_MIX_A1,               "blk.%d.time_mix_a1" },
+            { LLM_TENSOR_TIME_MIX_A2,               "blk.%d.time_mix_a2" },
+            { LLM_TENSOR_TIME_MIX_V0,               "blk.%d.time_mix_v0" },
+            { LLM_TENSOR_TIME_MIX_V1,               "blk.%d.time_mix_v1" },
+            { LLM_TENSOR_TIME_MIX_V2,               "blk.%d.time_mix_v2" },
+            { LLM_TENSOR_TIME_MIX_G1,               "blk.%d.time_mix_g1" },
+            { LLM_TENSOR_TIME_MIX_G2,               "blk.%d.time_mix_g2" },
+            { LLM_TENSOR_TIME_MIX_K_K,              "blk.%d.time_mix_k_k" },
+            { LLM_TENSOR_TIME_MIX_K_A,              "blk.%d.time_mix_k_a" },
+            { LLM_TENSOR_TIME_MIX_R_K,              "blk.%d.time_mix_r_k" },
+            { LLM_TENSOR_TIME_MIX_LERP_FUSED,       "blk.%d.time_mix_lerp_fused" },
+            { LLM_TENSOR_TIME_MIX_KEY,              "blk.%d.time_mix_key" },
+            { LLM_TENSOR_TIME_MIX_VALUE,            "blk.%d.time_mix_value" },
+            { LLM_TENSOR_TIME_MIX_RECEPTANCE,       "blk.%d.time_mix_receptance" },
+            { LLM_TENSOR_TIME_MIX_LN,               "blk.%d.time_mix_ln" },
+            { LLM_TENSOR_TIME_MIX_OUTPUT,           "blk.%d.time_mix_output" },
+            { LLM_TENSOR_FFN_NORM,                  "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,                  "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,                  "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,                    "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_GRANITE,
         {
@@ -1253,6 +1521,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
         },
     },
+    {
+        LLM_ARCH_SOLAR,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_BSKCN_TV,        "bskcn_tv" },
+        },
+    },
     {
         LLM_ARCH_WAVTOKENIZER_DEC,
         {
@@ -1279,23 +1565,44 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_SOLAR,
+        LLM_ARCH_BAILINGMOE,
         {
-            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-            { LLM_TENSOR_OUTPUT,          "output" },
-            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-            { LLM_TENSOR_BSKCN_TV,        "bskcn_tv" },
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FREQS,         "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_MISTRAL3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,  "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
+            { LLM_TENSOR_ATTN_NORM,   "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,      "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,      "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,      "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,    "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,    "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,    "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,      "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,    "blk.%d.ffn_down" },
+        }
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1333,23 +1640,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_K,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_V,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_QKV,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_OUT,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_DOWN_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_GATE_SHEXP,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_FFN_UP_SHEXP,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_A,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_Q_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_A_MQA,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_KV_B,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_Q,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_K,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_DEC_ATTN_V,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1376,6 +1668,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_OUT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_W2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_A2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_V2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G1,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_TIME_MIX_G2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W1,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_DECAY_W2,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_TIME_MIX_KEY,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -1394,12 +1692,19 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_TIME_MIX_LN,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_K,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_CHANNEL_MIX_LERP_R,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_K_A,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_TIME_MIX_R_K,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_W,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_K,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_V,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_R,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_LERP_G,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_LERP_FUSED,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_DECAY,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_W0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_A0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_TIME_MIX_V0,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
     {LLM_TENSOR_TIME_MIX_FIRST,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}},
     {LLM_TENSOR_ATTN_NORM,                  {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_NORM_2,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -1455,10 +1760,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_CONVNEXT_GAMMA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 };
 
-LLM_KV::LLM_KV(llm_arch arch) : arch(arch) {}
+LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
 std::string LLM_KV::operator()(llm_kv kv) const {
-    return ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
+        : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
 }
 
 std::string LLM_TN_IMPL::str() const {

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

@@ -10,6 +10,7 @@
 
 enum llm_arch {
     LLM_ARCH_LLAMA,
+    LLM_ARCH_LLAMA4,
     LLM_ARCH_MLLAMA,
     LLM_ARCH_DECI,
     LLM_ARCH_FALCON,
@@ -30,8 +31,11 @@ enum llm_arch {
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
     LLM_ARCH_QWEN2VL,
+    LLM_ARCH_QWEN3,
+    LLM_ARCH_QWEN3MOE,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
+    LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
@@ -40,6 +44,7 @@ enum llm_arch {
     LLM_ARCH_MINICPM3,
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
+    LLM_ARCH_GEMMA3,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
@@ -54,6 +59,7 @@ enum llm_arch {
     LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
+    LLM_ARCH_GLM4,
     LLM_ARCH_BITNET,
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
@@ -61,11 +67,17 @@ enum llm_arch {
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_EXAONE,
     LLM_ARCH_RWKV6,
+    LLM_ARCH_RWKV6QWEN2,
+    LLM_ARCH_RWKV7,
+    LLM_ARCH_ARWKV7,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
     LLM_ARCH_CHAMELEON,
     LLM_ARCH_SOLAR,
     LLM_ARCH_WAVTOKENIZER_DEC,
+    LLM_ARCH_MISTRAL3,
+    LLM_ARCH_PLM,
+    LLM_ARCH_BAILINGMOE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -74,6 +86,7 @@ enum llm_kv {
     LLM_KV_GENERAL_ARCHITECTURE,
     LLM_KV_GENERAL_QUANTIZATION_VERSION,
     LLM_KV_GENERAL_ALIGNMENT,
+    LLM_KV_GENERAL_FILE_TYPE,
     LLM_KV_GENERAL_NAME,
     LLM_KV_GENERAL_AUTHOR,
     LLM_KV_GENERAL_VERSION,
@@ -111,6 +124,8 @@ enum llm_kv {
     LLM_KV_TIME_DECAY_EXTRA_DIM,
     LLM_KV_RESIDUAL_SCALE,
     LLM_KV_EMBEDDING_SCALE,
+    LLM_KV_TOKEN_SHIFT_COUNT,
+    LLM_KV_INTERLEAVE_MOE_LAYER_STEP,
 
     LLM_KV_ATTENTION_HEAD_COUNT,
     LLM_KV_ATTENTION_HEAD_COUNT_KV,
@@ -125,11 +140,17 @@ enum llm_kv {
     LLM_KV_ATTENTION_CAUSAL,
     LLM_KV_ATTENTION_Q_LORA_RANK,
     LLM_KV_ATTENTION_KV_LORA_RANK,
+    LLM_KV_ATTENTION_DECAY_LORA_RANK,
+    LLM_KV_ATTENTION_ICLR_LORA_RANK,
+    LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK,
+    LLM_KV_ATTENTION_GATE_LORA_RANK,
     LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
     LLM_KV_ATTENTION_SLIDING_WINDOW,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
+    LLM_KV_ATTENTION_KEY_LENGTH_MLA,
+    LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -177,6 +198,8 @@ enum llm_kv {
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -241,6 +264,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,
+    LLM_TENSOR_POST_ATTN_NORM,
+    LLM_TENSOR_POST_MLP_NORM,
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -248,14 +273,27 @@ enum llm_tensor {
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
     LLM_TENSOR_SSM_OUT,
+    LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
     LLM_TENSOR_TIME_MIX_W2,
+    LLM_TENSOR_TIME_MIX_A0,
+    LLM_TENSOR_TIME_MIX_A1,
+    LLM_TENSOR_TIME_MIX_A2,
+    LLM_TENSOR_TIME_MIX_V0,
+    LLM_TENSOR_TIME_MIX_V1,
+    LLM_TENSOR_TIME_MIX_V2,
+    LLM_TENSOR_TIME_MIX_G1,
+    LLM_TENSOR_TIME_MIX_G2,
+    LLM_TENSOR_TIME_MIX_K_K,
+    LLM_TENSOR_TIME_MIX_K_A,
+    LLM_TENSOR_TIME_MIX_R_K,
     LLM_TENSOR_TIME_MIX_LERP_X,
     LLM_TENSOR_TIME_MIX_LERP_W,
     LLM_TENSOR_TIME_MIX_LERP_K,
     LLM_TENSOR_TIME_MIX_LERP_V,
     LLM_TENSOR_TIME_MIX_LERP_R,
     LLM_TENSOR_TIME_MIX_LERP_G,
+    LLM_TENSOR_TIME_MIX_LERP_FUSED,
     LLM_TENSOR_TIME_MIX_FIRST,
     LLM_TENSOR_TIME_MIX_DECAY,
     LLM_TENSOR_TIME_MIX_DECAY_W1,
@@ -275,6 +313,8 @@ enum llm_tensor {
     LLM_TENSOR_ATTN_Q_B,
     LLM_TENSOR_ATTN_KV_A_MQA,
     LLM_TENSOR_ATTN_KV_B,
+    LLM_TENSOR_ATTN_K_B,
+    LLM_TENSOR_ATTN_V_B,
     LLM_TENSOR_ATTN_Q_A_NORM,
     LLM_TENSOR_ATTN_KV_A_NORM,
     LLM_TENSOR_ATTN_SUB_NORM,
@@ -343,9 +383,10 @@ enum llm_tensor_layer {
 };
 
 struct LLM_KV {
-    LLM_KV(llm_arch arch);
+    LLM_KV(llm_arch arch, const char * suffix = nullptr);
 
     llm_arch arch;
+    const char * suffix;
 
     std::string operator()(llm_kv kv) const;
 };

llama/llama.cpp/src/llama-batch.h

@@ -42,9 +42,9 @@ struct llama_sbatch {
     bool logits_all; // TODO: remove once lctx.logits_all is removed too
 
     // sorted indices into the batch
-    std::vector<size_t> ids;
+    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<size_t> out_ids;
+    std::vector<int64_t> out_ids;
     std::vector<llama_sbatch_seq> seq;
 
     const llama_batch * batch = nullptr;

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

@@ -4,6 +4,7 @@
 
 #include <map>
 #include <sstream>
+#include <algorithm>
 
 #if __cplusplus >= 202000L
     #define LU8(x) (const char*)(u8##x)
@@ -35,6 +36,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
     { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
     { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
+    { "phi4",              LLM_CHAT_TEMPLATE_PHI_4             },
     { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
     { "zephyr",            LLM_CHAT_TEMPLATE_ZEPHYR            },
     { "monarch",           LLM_CHAT_TEMPLATE_MONARCH           },
@@ -50,12 +52,16 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
     { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
     { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
     { "megrez",            LLM_CHAT_TEMPLATE_MEGREZ            },
+    { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
+    { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
+    { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -73,7 +79,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return tmpl.find(haystack) != std::string::npos;
     };
     if (tmpl_contains("<|im_start|>")) {
-        return LLM_CHAT_TEMPLATE_CHATML;
+        return tmpl_contains("<|im_sep|>")
+            ? LLM_CHAT_TEMPLATE_PHI_4
+            : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -112,7 +120,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
-        return LLM_CHAT_TEMPLATE_FALCON_3;
+        return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -149,7 +157,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_MINICPM;
     } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_2;
-    } else if (tmpl_contains(LU8("'<｜Assistant｜>' + message['content'] + '<｜end▁of▁sentence｜>'"))) {
+    } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_3;
     } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
@@ -163,6 +171,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_GIGACHAT;
     } else if (tmpl_contains("<|role_start|>")) {
         return LLM_CHAT_TEMPLATE_MEGREZ;
+    } else if (tmpl_contains(" Ассистент:")) {
+        return LLM_CHAT_TEMPLATE_YANDEX;
+    } else if (tmpl_contains("<role>ASSISTANT</role>") && tmpl_contains("'HUMAN'")) {
+        return LLM_CHAT_TEMPLATE_BAILING;
+    } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA4;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -269,6 +283,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_PHI_4) {
+        // chatml template
+        for (auto message : chat) {
+            ss << "<|im_start|>" << message->role << "<|im_sep|>" << message->content << "<|im_end|>";
+        }
+        if (add_ass) {
+            ss << "<|im_start|>assistant<|im_sep|>";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {
         // Falcon 3
         for (auto message : chat) {
@@ -429,6 +451,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         for (auto message : chat) {
@@ -546,7 +576,51 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|role_start|>assistant<|role_end|>";
         }
-    } else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) {
+        // Yandex template ("\n\n" is defined as EOT token)
+
+        ss << "<s>";
+
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << " Пользователь: " << chat[i]->content << "\n\n";
+            } else if (role == "assistant") {
+                ss << " Ассистент: " << chat[i]->content << "\n\n";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << " Ассистент:[SEP]";
+        }
+    }  else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) {
+        // Bailing (Ling) template
+        for (auto message : chat) {
+            std::string role(message->role);
+
+            if (role == "user") {
+                role = "HUMAN";
+            } else {
+                std::transform(role.begin(), role.end(), role.begin(), ::toupper);
+            }
+
+            ss << "<role>" << role << "</role>" << message->content;
+        }
+
+        if (add_ass) {
+            ss << "<role>ASSISTANT</role>";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA4) {
+        // Llama 4
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|header_start|>" << role << "<|header_end|>\n\n" << trim(message->content) << "<|eot|>";
+        }
+        if (add_ass) {
+            ss << "<|header_start|>assistant<|header_end|>\n\n";
+        }
+    }  else {
         // template not supported
         return -1;
     }
@@ -564,4 +638,3 @@ int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
     }
     return (int32_t) LLM_CHAT_TEMPLATES.size();
 }
-

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

@@ -15,6 +15,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
     LLM_CHAT_TEMPLATE_PHI_3,
+    LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,
     LLM_CHAT_TEMPLATE_ZEPHYR,
     LLM_CHAT_TEMPLATE_MONARCH,
@@ -30,12 +31,16 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_LLAMA_3,
     LLM_CHAT_TEMPLATE_CHATGML_3,
     LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_MEGREZ,
+    LLM_CHAT_TEMPLATE_YANDEX,
+    LLM_CHAT_TEMPLATE_BAILING,
+    LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

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

@@ -3,66 +3,216 @@
 #include "llama.h"
 #include "llama-batch.h"
 #include "llama-cparams.h"
-#include "llama-model.h"
-#include "llama-kv-cache.h"
+#include "llama-graph.h"
 #include "llama-adapter.h"
+#include "llama-kv-cache.h"
 
 #include "ggml-cpp.h"
 
 #include <map>
-#include <unordered_map>
 #include <vector>
-#include <set>
+
+struct llama_model;
+struct llama_kv_cache;
+
+class llama_io_read_i;
+class llama_io_write_i;
 
 struct llama_context {
-    llama_context(const llama_model & model)
-        : model(model)
-        , t_start_us(model.t_start_us)
-        , t_load_us(model.t_load_us) {}
+    // init scheduler and compute buffers, reserve worst-case graphs
+    llama_context(
+            const llama_model & model,
+                  llama_context_params params);
 
-    const struct llama_model & model;
+    ~llama_context();
 
-    struct llama_cparams        cparams;
-    struct llama_sbatch         sbatch;  // TODO: revisit if needed
-    struct llama_kv_cache       kv_self;
-    struct llama_control_vector cvec;
+    void synchronize();
 
-    std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
+    const llama_model & get_model() const;
 
-    std::vector<ggml_backend_ptr> backends;
-    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
+    uint32_t n_ctx()         const;
+    uint32_t n_ctx_per_seq() const;
+    uint32_t n_batch()       const;
+    uint32_t n_ubatch()      const;
+    uint32_t n_seq_max()     const;
 
-    ggml_backend_t backend_cpu = nullptr;
+    uint32_t n_threads()       const;
+    uint32_t n_threads_batch() const;
 
-    ggml_threadpool_t threadpool       = nullptr;
-    ggml_threadpool_t threadpool_batch = nullptr;
+          llama_kv_cache * get_kv_self();
+    const llama_kv_cache * get_kv_self() const;
 
-    bool has_evaluated_once = false;
+    void kv_self_update();
 
-    mutable int64_t t_start_us;
-    mutable int64_t t_load_us;
-    mutable int64_t t_p_eval_us = 0;
-    mutable int64_t t_eval_us   = 0;
+    enum llama_pooling_type pooling_type() const;
 
-    mutable int64_t t_compute_start_us = 0;
-    mutable int64_t n_queued_tokens = 0;
+    float * get_logits();
+    float * get_logits_ith(int32_t i);
 
-    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
-    mutable int32_t n_eval   = 0; // number of eval calls
+    float * get_embeddings();
+    float * get_embeddings_ith(int32_t i);
+    float * get_embeddings_seq(llama_seq_id seq_id);
 
-    // host buffer for the model output (logits and embeddings)
-    ggml_backend_buffer_ptr buf_output;
+    void attach_threadpool(
+            ggml_threadpool_t threadpool,
+            ggml_threadpool_t threadpool_batch);
+
+    void detach_threadpool();
+
+    void set_n_threads(int32_t n_threads, int32_t n_threads_batch);
+
+    void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
+
+    void set_embeddings (bool value);
+    void set_causal_attn(bool value);
+    void set_warmup(bool value);
+    void set_cross_attn(bool value);
+
+    void set_adapter_lora(
+            llama_adapter_lora * adapter,
+            float scale);
+
+    bool rm_adapter_lora(
+            llama_adapter_lora * adapter);
+
+    void clear_adapter_lora();
+
+    bool apply_adapter_cvec(
+            const float * data,
+                 size_t   len,
+                int32_t   n_embd,
+                int32_t   il_start,
+                int32_t   il_end);
+
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
+
+    //
+    // state save/load
+    //
+
+    size_t state_get_size();
+    size_t state_get_data(      uint8_t * dst, size_t size);
+    size_t state_set_data(const uint8_t * src, size_t size);
+
+    size_t state_seq_get_size(llama_seq_id seq_id);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+
+    bool state_load_file(
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    bool state_save_file(
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    size_t state_seq_load_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    size_t state_seq_save_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    //
+    // perf
+    //
+
+    llama_perf_context_data perf_get_data() const;
+    void perf_reset();
+
+private:
+    //
+    // output
+    //
+
+    // Make sure enough space is available for outputs.
+    // Returns max number of outputs for which space was reserved.
+    int32_t output_reserve(int32_t n_outputs);
+
+    // make the outputs have the same order they had in the user-provided batch
+    // TODO: maybe remove this
+    void output_reorder();
+
+    //
+    // graph
+    //
+
+    int32_t graph_max_nodes() const;
+
+    // zero-out inputs and create the ctx_compute for the compute graph
+    ggml_cgraph * graph_init();
+
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
+    // returns the result of ggml_backend_sched_graph_compute_async execution
+    ggml_status graph_compute(
+            ggml_cgraph * gf,
+                   bool   batched);
+
+    llm_graph_cb graph_get_cb() const;
+
+    // used by kv_self_update()
+    ggml_tensor * build_rope_shift(
+        ggml_context * ctx0,
+        ggml_tensor * cur,
+        ggml_tensor * shift,
+        ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale,
+        ggml_backend_buffer * bbuf) const;
+
+    llm_graph_result_ptr build_kv_self_shift(
+            ggml_context * ctx0,
+            ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_kv_self_defrag(
+            ggml_context * ctx0,
+            ggml_cgraph * gf,
+            const std::vector<struct llama_kv_defrag_move> & moves) const;
+
+    // TODO: read/write lora adapters and cvec
+    size_t state_write_data(llama_io_write_i & io);
+    size_t state_read_data (llama_io_read_i  & io);
+
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+
+    //
+    // members
+    //
+
+    const llama_model & model;
+
+    llama_cparams       cparams;
+    llama_adapter_cvec  cvec;
+    llama_adapter_loras loras;
+    llama_sbatch        sbatch;
+
+    llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
+
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;
 
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
 
-    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
-    size_t  output_size = 0; // capacity (of tokens positions) for the output buffers
-    int32_t n_outputs   = 0; // number of actually-used outputs in the current ubatch or last logical batch
-
-    bool logits_all = false;
-
     // embeddings output (2-dimensional array: [n_outputs][n_embd])
     // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
     size_t  embd_size = 0; // capacity (of floats) for embeddings
@@ -72,59 +222,47 @@ struct llama_context {
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    // whether we are computing encoder output or decoder output
-    bool is_encoding = false;
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
 
-    // TODO: find a better way to accommodate mutli-dimension position encoding methods
-    // number of position id each token get, 1 for each token in most cases.
-    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
-    int n_pos_per_token = 1;
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 
-    // output of the encoder part of the encoder-decoder models
-    std::vector<float> embd_enc;
-    std::vector<std::set<llama_seq_id>> seq_ids_enc;
-
-    // memory buffers used to evaluate the model
-    std::vector<uint8_t> buf_compute_meta;
     ggml_backend_sched_ptr sched;
 
+    ggml_backend_t backend_cpu = nullptr;
+    std::vector<ggml_backend_ptr> backends;
+
+    ggml_context_ptr ctx_compute;
+
+    ggml_threadpool_t threadpool       = nullptr;
+    ggml_threadpool_t threadpool_batch = nullptr;
+
     ggml_abort_callback abort_callback      = nullptr;
     void *              abort_callback_data = nullptr;
 
-    // input tensors
-    struct ggml_tensor * inp_tokens;        // I32 [n_batch]
-    struct ggml_tensor * inp_embd;          // F32 [n_embd, n_batch]
-    struct ggml_tensor * inp_pos;           // I32 [n_batch]
-    struct ggml_tensor * inp_out_ids;       // I32 [n_outputs]
-    struct ggml_tensor * inp_KQ_mask;       // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_KQ_mask_swa;   // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_K_shift;       // I32 [kv_size]
-    struct ggml_tensor * inp_mean;          // F32 [n_batch, n_batch]
-    struct ggml_tensor * inp_cls;           // I32 [n_batch]
-    struct ggml_tensor * inp_s_copy;        // I32 [kv_size]
-    struct ggml_tensor * inp_s_mask;        // F32 [1, n_kv]
-    struct ggml_tensor * inp_s_seq;         // I32 [n_kv, n_batch]
-    struct ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
-    struct ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
-    struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+    std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    struct ggml_tensor * inp_cross_attn_state; // F32 [4, n_embd, 1061]
+    // buffer types used for the compute buffer of each backend
+    std::vector<ggml_backend_t>             backend_ptrs;
+    std::vector<ggml_backend_buffer_type_t> backend_buft;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    // host buffer for the model output (logits and embeddings)
+    ggml_backend_buffer_ptr buf_output;
+
+    bool has_evaluated_once = false;
+
+    // perf
+    mutable int64_t t_start_us  = 0;
+    mutable int64_t t_load_us   = 0;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens    = 0;
+
+    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    mutable int32_t n_eval   = 0; // number of eval calls
 };
-
-// TODO: make these methods of llama_context
-void llama_set_k_shift(struct llama_context & lctx);
-
-void llama_set_s_copy(struct llama_context & lctx);
-
-void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch);
-
-// Make sure enough space is available for outputs.
-// Returns max number of outputs for which space was reserved.
-size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs);
-
-// make the outputs have the same order they had in the user-provided batch
-void llama_output_reorder(struct llama_context & ctx);
-
-// For internal test use
-// TODO: remove
-const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

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

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool cross_attn;
+    bool warmup;
 
     enum llama_pooling_type pooling_type;
 

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

@@ -345,194 +345,194 @@ const char * llama_grammar_parser::parse_sequence(
     size_t last_sym_start = rule.size();
     const char * pos = src;
 
-        auto handle_repetitions = [&](int min_times, int max_times) {
+    auto handle_repetitions = [&](int min_times, int max_times) {
 
-            if (last_sym_start == rule.size()) {
-                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
-            }
+        if (last_sym_start == rule.size()) {
+            throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
+        }
 
-            // apply transformation to previous symbol (last_sym_start to end) according to
-            // the following rewrite rules:
-            // S{m,n} --> S S S (m times) S'(n-m)
-            //            S'(x)   ::= S S'(x-1) |
-            //            (... n-m definitions of these S' rules ...)
-            //            S'(1)   ::= S |
-            // S{m,} -->  S S S (m times) S'
-            //            S'     ::= S S' |
-            // S*     --> S{0,}
-            //        --> S'     ::= S S' |
-            // S+     --> S{1,}
-            //        --> S S'
-            //            S'     ::= S S' |
-            // S?     --> S{0,1}
-            //        --> S'
-            //            S'     ::= S |
+        // apply transformation to previous symbol (last_sym_start to end) according to
+        // the following rewrite rules:
+        // S{m,n} --> S S S (m times) S'(n-m)
+        //            S'(x)   ::= S S'(x-1) |
+        //            (... n-m definitions of these S' rules ...)
+        //            S'(1)   ::= S |
+        // S{m,} -->  S S S (m times) S'
+        //            S'     ::= S S' |
+        // S*     --> S{0,}
+        //        --> S'     ::= S S' |
+        // S+     --> S{1,}
+        //        --> S S'
+        //            S'     ::= S S' |
+        // S?     --> S{0,1}
+        //        --> S'
+        //            S'     ::= S |
 
-            llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end());
-            if (min_times == 0) {
-                rule.resize(last_sym_start);
-            } else {
-                // Repeat the previous elements (min_times - 1) times
-                for (int i = 1; i < min_times; i++) {
-                    rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
-                }
-            }
-
-            uint32_t last_rec_rule_id = 0;
-            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
-
-            llama_grammar_rule rec_rule(prev_rule);
-            for (int i = 0; i < n_opt; i++) {
-                rec_rule.resize(prev_rule.size());
-                uint32_t rec_rule_id = generate_symbol_id( rule_name);
-                if (i > 0 || max_times < 0) {
-                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
-                }
-                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
-                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
-                add_rule( rec_rule_id, rec_rule);
-                last_rec_rule_id = rec_rule_id;
-            }
-            if (n_opt > 0) {
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
-            }
-        };
-
-        while (*pos) {
-            if (*pos == '"') { // literal string
-                pos++;
-                last_sym_start = rule.size();
-                while (*pos != '"') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '[') { // char range(s)
-                pos++;
-                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
-                if (*pos == '^') {
-                    pos++;
-                    start_type = LLAMA_GRETYPE_CHAR_NOT;
-                }
-                last_sym_start = rule.size();
-                while (*pos != ']') {
-                    if (!*pos) {
-                        throw std::runtime_error("unexpected end of input");
-                    }
-                    auto char_pair = parse_char(pos);
-                         pos       = char_pair.second;
-                    enum llama_gretype type = last_sym_start < rule.size()
-                        ? LLAMA_GRETYPE_CHAR_ALT
-                        : start_type;
-
-                    rule.push_back({type, char_pair.first});
-                    if (pos[0] == '-' && pos[1] != ']') {
-                        if (!pos[1]) {
-                            throw std::runtime_error("unexpected end of input");
-                        }
-                        auto endchar_pair = parse_char(pos + 1);
-                             pos          = endchar_pair.second;
-                        rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
-                    }
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (is_word_char(*pos)) { // rule reference
-                const char * name_end    = parse_name(pos);
-                uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
-                pos = parse_space(name_end, is_nested);
-                last_sym_start = rule.size();
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
-            } else if (*pos == '(') { // grouping
-                // parse nested alternates into synthesized rule
-                pos = parse_space(pos + 1, true);
-                uint32_t sub_rule_id = generate_symbol_id(rule_name);
-                pos = parse_alternates(pos, rule_name, sub_rule_id, true);
-                last_sym_start = rule.size();
-                // output reference to synthesized rule
-                rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
-                if (*pos != ')') {
-                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
-                }
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '.') { // any char
-                last_sym_start = rule.size();
-                rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
-                pos = parse_space(pos + 1, is_nested);
-            } else if (*pos == '*') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, -1);
-            } else if (*pos == '+') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(1, -1);
-            } else if (*pos == '?') {
-                pos = parse_space(pos + 1, is_nested);
-                handle_repetitions(0, 1);
-            } else if (*pos == '{') {
-                pos = parse_space(pos + 1, is_nested);
-
-                if (!is_digit_char(*pos)) {
-                    throw std::runtime_error(std::string("expecting an int at ") + pos);
-                }
-                const char * int_end = parse_int(pos);
-                int min_times = std::stoul(std::string(pos, int_end - pos));
-                pos = parse_space(int_end, is_nested);
-
-                int max_times = -1;
-
-                if (*pos == '}') {
-                    max_times = min_times;
-                    pos = parse_space(pos + 1, is_nested);
-                } else if (*pos == ',') {
-                    pos = parse_space(pos + 1, is_nested);
-
-                    if (is_digit_char(*pos)) {
-                        const char * int_end = parse_int(pos);
-                        max_times = std::stoul(std::string(pos, int_end - pos));
-                        pos = parse_space(int_end, is_nested);
-                    }
-
-                    if (*pos != '}') {
-                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
-                    }
-                    pos = parse_space(pos + 1, is_nested);
-                } else {
-                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
-                }
-                handle_repetitions(min_times, max_times);
-            } else {
-                break;
+        llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end());
+        if (min_times == 0) {
+            rule.resize(last_sym_start);
+        } else {
+            // Repeat the previous elements (min_times - 1) times
+            for (int i = 1; i < min_times; i++) {
+                rule.insert(rule.end(), prev_rule.begin(), prev_rule.end());
             }
         }
-        return pos;
+
+        uint32_t last_rec_rule_id = 0;
+        auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+
+        llama_grammar_rule rec_rule(prev_rule);
+        for (int i = 0; i < n_opt; i++) {
+            rec_rule.resize(prev_rule.size());
+            uint32_t rec_rule_id = generate_symbol_id( rule_name);
+            if (i > 0 || max_times < 0) {
+                rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+            }
+            rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            rec_rule.push_back({LLAMA_GRETYPE_END, 0});
+            add_rule( rec_rule_id, rec_rule);
+            last_rec_rule_id = rec_rule_id;
+        }
+        if (n_opt > 0) {
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
+        }
+    };
+
+    while (*pos) {
+        if (*pos == '"') { // literal string
+            pos++;
+            last_sym_start = rule.size();
+            while (*pos != '"') {
+                if (!*pos) {
+                    throw std::runtime_error("unexpected end of input");
+                }
+                auto char_pair = parse_char(pos);
+                     pos       = char_pair.second;
+                rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '[') { // char range(s)
+            pos++;
+            enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+            if (*pos == '^') {
+                pos++;
+                start_type = LLAMA_GRETYPE_CHAR_NOT;
+            }
+            last_sym_start = rule.size();
+            while (*pos != ']') {
+                if (!*pos) {
+                    throw std::runtime_error("unexpected end of input");
+                }
+                auto char_pair = parse_char(pos);
+                     pos       = char_pair.second;
+                enum llama_gretype type = last_sym_start < rule.size()
+                    ? LLAMA_GRETYPE_CHAR_ALT
+                    : start_type;
+
+                rule.push_back({type, char_pair.first});
+                if (pos[0] == '-' && pos[1] != ']') {
+                    if (!pos[1]) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
+                    auto endchar_pair = parse_char(pos + 1);
+                         pos          = endchar_pair.second;
+                    rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                }
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (is_word_char(*pos)) { // rule reference
+            const char * name_end    = parse_name(pos);
+            uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
+            pos = parse_space(name_end, is_nested);
+            last_sym_start = rule.size();
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+        } else if (*pos == '(') { // grouping
+            // parse nested alternates into synthesized rule
+            pos = parse_space(pos + 1, true);
+            uint32_t sub_rule_id = generate_symbol_id(rule_name);
+            pos = parse_alternates(pos, rule_name, sub_rule_id, true);
+            last_sym_start = rule.size();
+            // output reference to synthesized rule
+            rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+            if (*pos != ')') {
+                throw std::runtime_error(std::string("expecting ')' at ") + pos);
+            }
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '.') { // any char
+            last_sym_start = rule.size();
+            rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
+            pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '*') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(0, -1);
+        } else if (*pos == '+') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(1, -1);
+        } else if (*pos == '?') {
+            pos = parse_space(pos + 1, is_nested);
+            handle_repetitions(0, 1);
+        } else if (*pos == '{') {
+            pos = parse_space(pos + 1, is_nested);
+
+            if (!is_digit_char(*pos)) {
+                throw std::runtime_error(std::string("expecting an int at ") + pos);
+            }
+            const char * int_end = parse_int(pos);
+            int min_times = std::stoul(std::string(pos, int_end - pos));
+            pos = parse_space(int_end, is_nested);
+
+            int max_times = -1;
+
+            if (*pos == '}') {
+                max_times = min_times;
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == ',') {
+                pos = parse_space(pos + 1, is_nested);
+
+                if (is_digit_char(*pos)) {
+                    const char * int_end = parse_int(pos);
+                    max_times = std::stoul(std::string(pos, int_end - pos));
+                    pos = parse_space(int_end, is_nested);
+                }
+
+                if (*pos != '}') {
+                    throw std::runtime_error(std::string("expecting '}' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                throw std::runtime_error(std::string("expecting ',' at ") + pos);
+            }
+            handle_repetitions(min_times, max_times);
+        } else {
+            break;
+        }
     }
+    return pos;
+}
 
 const char * llama_grammar_parser::parse_rule(const char * src) {
-        const char * name_end = parse_name(src);
-        const char * pos      = parse_space(name_end, false);
-        size_t       name_len = name_end - src;
-        uint32_t     rule_id  = get_symbol_id(src, name_len);
-        const std::string name(src, name_len);
+    const char * name_end = parse_name(src);
+    const char * pos      = parse_space(name_end, false);
+    size_t       name_len = name_end - src;
+    uint32_t     rule_id  = get_symbol_id(src, name_len);
+    const std::string name(src, name_len);
 
-        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
-            throw std::runtime_error(std::string("expecting ::= at ") + pos);
-        }
-        pos = parse_space(pos + 3, true);
-
-        pos = parse_alternates(pos, name, rule_id, false);
-
-        if (*pos == '\r') {
-            pos += pos[1] == '\n' ? 2 : 1;
-        } else if (*pos == '\n') {
-            pos++;
-        } else if (*pos) {
-            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
-        }
-        return parse_space(pos, true);
+    if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+        throw std::runtime_error(std::string("expecting ::= at ") + pos);
     }
+    pos = parse_space(pos + 3, true);
+
+    pos = parse_alternates(pos, name, rule_id, false);
+
+    if (*pos == '\r') {
+        pos += pos[1] == '\n' ? 2 : 1;
+    } else if (*pos == '\n') {
+        pos++;
+    } else if (*pos) {
+        throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+    }
+    return parse_space(pos, true);
+}
 
 bool llama_grammar_parser::parse(const char * src) {
     try {
@@ -560,7 +560,7 @@ bool llama_grammar_parser::parse(const char * src) {
             }
         }
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+        fprintf(stderr, "%s: error parsing grammar: %s\n\n%s\n", __func__, err.what(), src);
         rules.clear();
         return false;
     }
@@ -907,6 +907,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
         const llama_grammar_element ** rules,
         size_t n_rules,
         size_t start_rule_index) {
@@ -960,19 +961,35 @@ struct llama_grammar * llama_grammar_init_impl(
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        ollama_vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy =*/              false,
+        /* .awaiting_trigger = */ false,
+        /* .trigger_buffer = */   "",
+        /* .trigger_tokens   = */ {},
+        /* .trigger_patterns    = */ {},
+    };
 }
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     llama_grammar_parser parser;
 
     // if there is a grammar, parse it
-    if (!parser.parse(grammar_str)) {
-        return nullptr;
-    }
-
-    // will be empty (default) if there are parse errors
-    if (parser.rules.empty()) {
+    // rules will be empty (default) if there are parse errors
+    if (!parser.parse(grammar_str) || parser.rules.empty()) {
         fprintf(stderr, "%s: failed to parse grammar\n", __func__);
         return nullptr;
     }
@@ -1035,10 +1052,34 @@ struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab,
         }
     } while (true);
 
+    std::vector<llama_token>    vec_trigger_tokens;
+    std::vector<llama_grammar_trigger_pattern> vec_trigger_patterns;
+    for (size_t i = 0; i < num_trigger_tokens; i++) {
+        GGML_ASSERT(trigger_tokens != nullptr);
+        vec_trigger_tokens.push_back(trigger_tokens[i]);
+    }
+    for (size_t i = 0; i < num_trigger_patterns; i++) {
+        GGML_ASSERT(trigger_patterns != nullptr);
+        auto & trigger = vec_trigger_patterns.emplace_back();
+        trigger.pattern = trigger_patterns[i];
+        trigger.regex = std::regex(trigger.pattern);
+    }
+
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        ollama_vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy = */             lazy,
+        /* .awaiting_trigger = */ lazy,
+        /* .trigger_buffer = */   "",
+        std::move(vec_trigger_tokens),
+        std::move(vec_trigger_patterns),
+    };
 }
 
 void llama_grammar_free_impl(struct llama_grammar * grammar) {
@@ -1050,11 +1091,17 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
 }
 
 struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
-    llama_grammar * result = new llama_grammar {
+    auto * result = new llama_grammar {
         grammar.vocab,
+        grammar.o_vocab,
         grammar.rules,
         grammar.stacks,
         grammar.partial_utf8,
+        grammar.lazy,
+        grammar.awaiting_trigger,
+        grammar.trigger_buffer,
+        grammar.trigger_tokens,
+        grammar.trigger_patterns,
     };
 
     // redirect elements in stacks to point to new rules
@@ -1074,7 +1121,10 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
 }
 
 void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
-    GGML_ASSERT(grammar.vocab != nullptr);
+
+    if (grammar.awaiting_trigger) {
+        return;
+    }
 
     bool allow_eog = false;
     for (const auto & stack : grammar.stacks) {
@@ -1092,9 +1142,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 
     for (size_t i = 0; i < cur_p->size; ++i) {
         const llama_token id      = cur_p->data[i].id;
-        const std::string & piece = grammar.vocab->cache_token_to_piece.at(id);
+        const std::string piece = grammar.o_vocab ?
+            grammar.o_vocab->token_to_piece(id) :
+            grammar.vocab->token_to_piece(id);
 
-        if (llama_token_is_eog_impl(*grammar.vocab, id)) {
+        const bool is_eog = grammar.o_vocab ? grammar.o_vocab->is_eog(id) : grammar.vocab->is_eog(id);
+
+        if (is_eog) {
             if (!allow_eog) {
                 cur_p->data[i].logit = -INFINITY;
             }
@@ -1113,19 +1167,53 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 }
 
 void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
-    GGML_ASSERT(grammar.vocab != nullptr);
 
-    if (llama_token_is_eog_impl(*grammar.vocab, token)) {
+    const std::string piece = grammar.o_vocab ?
+        grammar.o_vocab->token_to_piece(token) :
+        grammar.vocab->token_to_piece(token);
+
+    if (grammar.awaiting_trigger) {
+        if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
+            grammar.awaiting_trigger = false;
+            grammar.trigger_buffer.clear();
+            llama_grammar_accept_str(grammar, piece);
+            LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
+            return;
+        } else {
+            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)) {
+                    grammar.awaiting_trigger = false;
+                    // get from the first match to the end of the string
+                    auto constrained_str = grammar.trigger_buffer.substr(match.position(1));
+                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
+                    grammar.trigger_buffer.clear();
+                    llama_grammar_accept_str(grammar, constrained_str);
+                    LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
+                    return;
+                }
+            }
+            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`)\n", token, piece.c_str());
+            return;
+        }
+    }
+
+    const bool is_eog = grammar.o_vocab ? grammar.o_vocab->is_eog(token) : grammar.vocab->is_eog(token);
+    if (is_eog) {
         for (const auto & stack : grammar.stacks) {
             if (stack.empty()) {
                 return;
             }
         }
-        GGML_ABORT("fatal error");
+        GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
     }
 
-    const std::string & piece = grammar.vocab->cache_token_to_piece.at(token);
+    llama_grammar_accept_str(grammar, piece);
+}
 
+void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
     const auto & code_points = decoded.first;
@@ -1135,5 +1223,32 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
     }
 
     grammar.partial_utf8 = decoded.second;
-    GGML_ASSERT(!grammar.stacks.empty());
+    if (grammar.stacks.empty()) {
+        throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
+    }
+}
+
+
+const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
+    try {
+        return token_to_piece_map.at(token);
+    } catch (const std::out_of_range&) {
+        throw std::runtime_error("Token not found in vocabulary: " + std::to_string(token));
+    }
+}
+
+void ollama_vocab::add_token_pieces(const uint32_t* tokens, size_t n_tokens, const char** pieces) {
+    for (size_t i = 0; i < n_tokens; i++) {
+        token_to_piece_map[tokens[i]] = pieces[i];
+    }
+}
+
+bool ollama_vocab::is_eog(const uint32_t token) const {
+    return special_eog_ids.count(token) > 0;
+}
+
+void ollama_vocab::set_eog_tokens(const uint32_t* tokens, size_t n_tokens) {
+    for (size_t i = 0; i < n_tokens; i++) {
+        special_eog_ids.insert(tokens[i]);
+    }
 }

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

@@ -3,10 +3,22 @@
 #include "llama.h"
 
 #include <map>
+#include <regex>
 #include <string>
 #include <vector>
+#include <set>
 
 struct llama_vocab;
+struct ollama_vocab {
+    std::map<uint32_t, std::string> token_to_piece_map;
+    std::set<uint32_t> special_eog_ids;
+
+    const std::string & token_to_piece(const uint32_t token) const;
+    void add_token_pieces(const uint32_t* tokens, size_t n_tokens, const char** pieces);
+    void set_eog_tokens(const uint32_t* tokens, size_t n_tokens);
+    bool is_eog(const uint32_t token) const;
+
+};
 
 // grammar element type
 enum llama_gretype {
@@ -105,15 +117,33 @@ struct llama_grammar_parser {
     void print(FILE * file);
 };
 
+struct llama_grammar_trigger_pattern {
+    std::string pattern;
+    std::regex  regex;
+};
+
 struct llama_grammar {
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
+    const ollama_vocab * o_vocab;
 
     const llama_grammar_rules  rules;  // TODO: shared ptr
           llama_grammar_stacks stacks;
 
     // buffer for partially generated UTF-8 sequence from accepted tokens
     llama_partial_utf8 partial_utf8;
+
+    // lazy grammars wait for trigger words or tokens before constraining the sampling.
+    // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // (useful e.g. for tool_choice=required)
+    bool                     lazy             = false;
+    bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
+    std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
+    std::vector<llama_grammar_trigger_pattern>
+                             trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
+                                                       // string, and the grammar will be given the string from the first match group onwards.
+
 };
 
 //
@@ -123,11 +153,21 @@ struct llama_grammar {
 // note: needed for tests (not great)
 struct llama_grammar * llama_grammar_init_impl(
         const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
         const llama_grammar_element ** rules,
         size_t n_rules,
         size_t start_rule_index);
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root);
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+        const struct ollama_vocab * ollama_vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_patterns,
+                            size_t num_trigger_patterns,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
 
 void llama_grammar_free_impl(struct llama_grammar * grammar);
 
@@ -141,3 +181,7 @@ void llama_grammar_apply_impl(
 void llama_grammar_accept_impl(
               struct llama_grammar & grammar,
                        llama_token   token);
+
+void llama_grammar_accept_str(
+              struct llama_grammar & grammar,
+                 const std::string & piece);

llama/llama.cpp/src/llama-graph.h

@@ -0,0 +1,608 @@
+#pragma once
+
+#include "llama-arch.h"
+#include "llama-hparams.h"
+#include "llama-adapter.h"
+
+#include <cstdint>
+#include <vector>
+#include <memory>
+#include <set>
+#include <functional>
+
+struct ggml_cgraph;
+struct ggml_context;
+struct ggml_tensor;
+
+struct llama_ubatch;
+struct llama_cparams;
+
+class llama_memory_i;
+class llama_kv_cache_unified;
+
+// certain models (typically multi-modal) can produce different types of graphs
+enum llm_graph_type {
+    LLM_GRAPH_TYPE_DEFAULT,
+    LLM_GRAPH_TYPE_ENCODER,
+    LLM_GRAPH_TYPE_DECODER,
+};
+
+enum llm_ffn_op_type {
+    LLM_FFN_SILU,
+    LLM_FFN_GELU,
+    LLM_FFN_RELU,
+    LLM_FFN_RELU_SQR,
+    LLM_FFN_SWIGLU,
+};
+
+enum llm_ffn_gate_type {
+    LLM_FFN_SEQ,
+    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up
+};
+
+enum llm_norm_type {
+    LLM_NORM,
+    LLM_NORM_RMS,
+    LLM_NORM_GROUP,
+};
+
+// TODO: tmp - need something better to pass the data from the encoder to the decoder
+struct llama_cross {
+    // the output embeddings from the encoder as a ggml tensor
+    // TODO: this needs more work to be correct, for now copy the embeddings data to host memory
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524
+    //ggml_tensor * t_embd = nullptr;
+
+    int64_t n_embd = 0;
+    int64_t n_enc  = 0;
+
+    // embeddings data copied to host memory (tmp)
+    std::vector<float> v_embd;
+
+    // needed to construct the cross-attention mask in the decoder
+    std::vector<std::set<llama_seq_id>> seq_ids_enc;
+};
+
+//
+// llm_graph_input
+//
+
+class llm_graph_input_i {
+public:
+    virtual ~llm_graph_input_i() = default;
+
+    virtual void set_input(const llama_ubatch * ubatch) = 0;
+};
+
+using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
+
+
+class llm_graph_input_embd : public llm_graph_input_i {
+public:
+    llm_graph_input_embd()          = default;
+    virtual ~llm_graph_input_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * tokens = nullptr; // I32 [n_batch]
+    ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
+};
+
+class llm_graph_input_pos : public llm_graph_input_i {
+public:
+    llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {}
+    virtual ~llm_graph_input_pos() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos = nullptr; // I32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+};
+
+// temperature tuning, used by llama4
+class llm_graph_input_attn_temp : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
+        : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    virtual ~llm_graph_input_attn_temp() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * attn_scale = nullptr; // F32 [n_batch]
+
+    const int64_t n_pos_per_token = 1;
+
+    const uint32_t n_attn_temp_floor_scale;
+    const float    f_attn_temp_scale;
+};
+
+class llm_graph_input_pos_bucket : public llm_graph_input_i {
+public:
+    llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {}
+    virtual ~llm_graph_input_pos_bucket() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
+
+    const llama_hparams & hparams;
+};
+
+class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
+public:
+    llm_graph_input_pos_bucket_kv(
+            const llama_hparams & hparams,
+            const llama_kv_cache_unified * kv_self) : hparams(hparams), kv_self(kv_self) {}
+    virtual ~llm_graph_input_pos_bucket_kv() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_out_ids : public llm_graph_input_i {
+public:
+    llm_graph_input_out_ids(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
+    virtual ~llm_graph_input_out_ids() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * out_ids; // I32 [n_outputs]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const int32_t n_outputs;
+};
+
+class llm_graph_input_mean : public llm_graph_input_i {
+public:
+    llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_mean() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * mean; // F32 [n_batch, n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_cls : public llm_graph_input_i {
+public:
+    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    virtual ~llm_graph_input_cls() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cls; // I32 [n_batch]
+
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_s_copy : public llm_graph_input_i {
+public:
+    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_copy() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_copy; // I32 [kv_size]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_s_mask : public llm_graph_input_i {
+public:
+    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    virtual ~llm_graph_input_s_mask() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_mask; // F32 [1, n_kv]
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_cross_embd : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_embd(
+            const llama_cross * cross) : cross(cross) {}
+    virtual ~llm_graph_input_cross_embd() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc]
+
+    const llama_cross * cross;
+};
+
+class llm_graph_input_attn_no_cache : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) :
+        hparams(hparams),
+        cparams(cparams) {
+    }
+    ~llm_graph_input_attn_no_cache() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
+
+    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch]
+    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+};
+
+class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_kv_unified(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_unified * kv_self) :
+        hparams(hparams),
+        cparams(cparams),
+        kv_self(kv_self) {
+    }
+    ~llm_graph_input_attn_kv_unified() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
+
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_kv_cache_unified * kv_self;
+};
+
+class llm_graph_input_attn_cross : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {}
+    ~llm_graph_input_attn_cross() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
+
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch]
+
+    const llama_cross * cross = nullptr;
+};
+
+class llm_graph_input_cross_attn_state : public llm_graph_input_i {
+public:
+    llm_graph_input_cross_attn_state()          = default;
+    virtual ~llm_graph_input_cross_attn_state() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * cross_attn_state; // F32 [4, n_embd, 1061]
+};
+
+//
+// llm_graph_result
+//
+
+// these objects deliver the result from the graph build process back to the llama_context
+// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their
+//   specific data, by calling the set_inputs() method
+// along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.
+//   these are used by the llama_context to extact the relevant data, based on the compute parameters
+
+class llm_graph_result_i {
+public:
+    virtual ~llm_graph_result_i() = default;
+
+    virtual ggml_tensor * get_logits()      = 0;
+    virtual ggml_tensor * get_embd()        = 0;
+    virtual ggml_tensor * get_embd_pooled() = 0;
+
+    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
+};
+
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
+
+
+class llm_graph_result : public llm_graph_result_i {
+public:
+    virtual ~llm_graph_result() = default;
+
+    ggml_tensor * get_logits()      override { return t_logits; }
+    ggml_tensor * get_embd()        override { return t_embd; }
+    ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
+
+    void set_inputs(const llama_ubatch * ubatch) override {
+        for (auto & input : inputs) {
+            input->set_input(ubatch);
+        }
+    }
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
+        inputs.emplace_back(std::move(input));
+        return inputs.back().get();
+    }
+
+    // important graph nodes
+    ggml_tensor * t_logits      = nullptr;
+    ggml_tensor * t_embd        = nullptr;
+    ggml_tensor * t_embd_pooled = nullptr;
+
+    std::vector<llm_graph_input_ptr> inputs;
+};
+
+//
+// llm_graph_context
+//
+
+// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
+using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
+
+struct llm_graph_params {
+    ggml_context * ctx;
+
+    const llm_arch arch;
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+    const llama_ubatch  & ubatch;
+
+    ggml_backend_sched * sched;
+    ggml_backend * backend_cpu;
+
+    const llama_adapter_cvec  * cvec;
+    const llama_adapter_loras * loras;
+    const llama_memory_i      * memory;
+    const llama_cross         * cross;
+
+    int32_t n_outputs;
+
+    const llm_graph_cb & cb;
+};
+
+struct llm_graph_context {
+    const llm_arch arch;
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+    const llama_ubatch  & ubatch;
+
+    const int64_t n_embd;
+    const int64_t n_layer;
+    const int64_t n_rot;
+    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
+    const int64_t n_ctx_per_seq;
+    const int64_t n_head;
+    const int64_t n_head_kv;
+    const int64_t n_embd_head_k;
+    const int64_t n_embd_k_gqa;
+    const int64_t n_embd_head_v;
+    const int64_t n_embd_v_gqa;
+    const int64_t n_expert;
+    const int64_t n_expert_used;
+
+    const float freq_base;
+    const float freq_scale;
+    const float ext_factor;
+    const float attn_factor;
+    const float beta_fast;
+    const float beta_slow;
+    const float norm_eps;
+    const float norm_rms_eps;
+
+    const int32_t n_tokens;
+    const int32_t n_outputs;
+    const int32_t n_ctx_orig; // yarn
+
+    const enum llama_pooling_type pooling_type;
+    const enum llama_rope_type    rope_type;
+
+    ggml_context * ctx0 = nullptr;
+
+    ggml_backend_sched * sched;
+
+    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+
+    const llama_adapter_cvec  * cvec;
+    const llama_adapter_loras * loras;
+    const llama_memory_i      * memory;
+    const llama_cross         * cross;
+
+    const llm_graph_cb & cb_func;
+
+    std::unique_ptr<llm_graph_result> res;
+
+    llm_graph_context(const llm_graph_params & params);
+
+    int64_t n_pos_per_token() const;
+
+    void cb(ggml_tensor * cur, const char * name, int il) const;
+
+    //
+    // common
+    //
+
+    ggml_tensor * build_cvec(
+             ggml_tensor * cur,
+                     int   il) const;
+
+    // do mat_mul, while optionally apply lora
+    ggml_tensor * build_lora_mm(
+              ggml_tensor * w,
+              ggml_tensor * cur) const;
+
+    // do mat_mul_id, while optionally apply lora
+    ggml_tensor * build_lora_mm_id(
+              ggml_tensor * w,   // ggml_tensor * as
+              ggml_tensor * cur, // ggml_tensor * b
+              ggml_tensor * ids) const;
+
+    ggml_tensor * build_norm(
+             ggml_tensor * cur,
+             ggml_tensor * mw,
+             ggml_tensor * mb,
+           llm_norm_type   type,
+                     int   il) const;
+
+    ggml_tensor * build_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * up,
+             ggml_tensor * up_b,
+             ggml_tensor * up_s,
+             ggml_tensor * gate,
+             ggml_tensor * gate_b,
+             ggml_tensor * gate_s,
+             ggml_tensor * down,
+             ggml_tensor * down_b,
+             ggml_tensor * down_s,
+             ggml_tensor * act_scales,
+         llm_ffn_op_type   type_op,
+       llm_ffn_gate_type   type_gate,
+                     int   il) const;
+
+    ggml_tensor * build_moe_ffn(
+             ggml_tensor * cur,
+             ggml_tensor * gate_inp,
+             ggml_tensor * up_exps,
+             ggml_tensor * gate_exps,
+             ggml_tensor * down_exps,
+             ggml_tensor * exp_probs_b,
+                 int64_t   n_expert,
+                 int64_t   n_expert_used,
+         llm_ffn_op_type   type_op,
+                    bool   norm_w,
+                    bool   scale_w,
+                   float   w_scale,
+            llama_expert_gating_func_type gating_op,
+                     int   il) const;
+
+    //
+    // inputs
+    //
+
+    ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const;
+    ggml_tensor * build_inp_pos() const;
+    ggml_tensor * build_inp_attn_scale() const;
+    ggml_tensor * build_inp_out_ids() const;
+    ggml_tensor * build_inp_mean() const;
+    ggml_tensor * build_inp_cls() const;
+    ggml_tensor * build_inp_s_copy() const;
+    ggml_tensor * build_inp_s_mask() const;
+    ggml_tensor * build_inp_cross_attn_state() const;
+
+    ggml_tensor * build_inp_cross_embd() const;
+    ggml_tensor * build_inp_pos_bucket_enc() const;
+    ggml_tensor * build_inp_pos_bucket_dec() const;
+    ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
+
+    //
+    // attention
+    //
+
+    ggml_tensor * build_attn_mha(
+             ggml_cgraph * gf,
+             ggml_tensor * q,     // [n_embd_head_q, n_tokens, n_head_q]
+             ggml_tensor * k,     // [n_embd_head_k, n_tokens, n_head_k]
+             ggml_tensor * v,     // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false)
+             ggml_tensor * kq_b,
+             ggml_tensor * kq_mask,
+             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                    bool   v_trans,
+                   float   kq_scale) const;
+
+    llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_no_cache * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_kv_unified * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_cross * build_attn_inp_cross() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_cross * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
+
+    //
+    // recurrent
+    //
+
+    ggml_tensor * build_copy_mask_state(
+             ggml_cgraph * gf,
+             ggml_tensor * s,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+                 int32_t   n_state,
+                 int32_t   n_seqs) const;
+
+    ggml_tensor * build_rwkv_token_shift_load(
+             ggml_cgraph * gf,
+             ggml_tensor * state_copy,
+             ggml_tensor * state_mask,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    ggml_tensor * build_rwkv_token_shift_store(
+             ggml_tensor * token_shift,
+      const llama_ubatch & ubatch,
+                     int   il) const;
+
+    //
+    // pooling
+    //
+
+    void build_pooling(
+            ggml_cgraph * gf,
+            ggml_tensor * cls,
+            ggml_tensor * cls_b,
+            ggml_tensor * cls_out,
+            ggml_tensor * cls_out_b) const;
+};

llama/llama.cpp/src/llama-hparams.cpp

@@ -2,8 +2,6 @@
 
 #include "ggml.h"
 
-#include <algorithm>
-
 uint32_t llama_hparams::n_head(uint32_t il) const {
     if (il < n_layer) {
         return n_head_arr[il];
@@ -54,7 +52,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
 uint32_t llama_hparams::n_embd_k_s() const {
     if (wkv_head_size != 0) {
         // for RWKV models
-        return 2 * n_embd;
+        return token_shift_count * n_embd;
     }
 
     // TODO: maybe support other convolution strides than 1
@@ -80,6 +78,14 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_swa(uint32_t il) const {
+    if (il < n_layer) {
+        return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1);
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 bool llama_hparams::cross_attention_layers(uint32_t il) const {
     return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
 }

llama/llama.cpp/src/llama-hparams.h

@@ -2,6 +2,8 @@
 
 #include "llama.h"
 
+#include <algorithm>
+
 #include <array>
 
 // bump if necessary
@@ -30,19 +32,23 @@ struct llama_hparams {
     bool use_par_res;
     bool swin_norm;
 
-    uint32_t n_vocab = 0;
     uint32_t n_ctx_train; // context size the model was trained on
     uint32_t n_embd;
     uint32_t n_embd_features = 0;
     uint32_t n_layer;
     uint32_t n_rot;
     uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
     uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
-    uint32_t n_vocab_type = 0; // for BERT-style token types
     uint32_t n_rel_attn_bkts = 0;
+    uint32_t n_vocab = 0;
+
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 0;
 
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
@@ -79,10 +85,17 @@ struct llama_hparams {
     uint32_t time_mix_extra_dim     = 0;
     uint32_t time_decay_extra_dim   = 0;
     uint32_t wkv_head_size          = 0;
+    uint32_t token_shift_count      = 2;
+    uint32_t n_lora_decay           = 0;
+    uint32_t n_lora_iclr            = 0;
+    uint32_t n_lora_value_res_mix   = 0;
+    uint32_t n_lora_gate            = 0;
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
     float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
     uint32_t n_ctx_orig_yarn;
     float    rope_yarn_log_mul;
 
@@ -109,6 +122,14 @@ struct llama_hparams {
     bool use_alibi     = false;
     bool attn_soft_cap = false;
 
+    uint32_t n_moe_layer_step        = 0;
+    bool     use_kq_norm             = true;
+    uint32_t n_attn_chunk            = 0;
+    // values below seems to be fixed on llama4
+    uint32_t n_no_rope_layer_step    = 4;
+    uint32_t n_attn_temp_floor_scale = 8192;
+    float    f_attn_temp_scale       = 0.1;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -141,8 +162,10 @@ struct llama_hparams {
     // Block skip connection
     bool n_bskcn(uint32_t n, uint32_t il) const;
 
-    // cross attention layers   
+    // cross attention layers
     bool cross_attention_layers(uint32_t il) const;
+
+    bool is_swa(uint32_t il) const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

llama/llama.cpp/src/llama-impl.cpp

@@ -1,5 +1,6 @@
 #include "llama-impl.h"
 
+#include "gguf.h"
 #include "llama.h"
 
 #include <cinttypes>
@@ -138,7 +139,7 @@ std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             {
                 const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
                 int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {

llama/llama.cpp/src/llama-impl.h

@@ -6,13 +6,13 @@
 #include <vector>
 
 #ifdef __GNUC__
-#ifdef __MINGW32__
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
 #else
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-#else
-#define LLAMA_ATTRIBUTE_FORMAT(...)
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
 #endif
 
 //