model: support GLM4V vision encoder (#18042)

* convert ok

* no deepstack

* less new tensors

* cgraph ok

* add mrope for text model

* faster patch merger

* add GGML_ROPE_TYPE_MRNORM

* add support for metal

* move glm4v do dedicated graph

* convert: add norm_embd

* clip: add debugging fn

* working correctly

* fix style

* use bicubic

* fix mrope metal

* improve cpu

* convert to neox ordering on conversion

* revert backend changes

* force stop if using old weight

* support moe variant

* fix conversion

* fix convert (2)

* Update tools/mtmd/clip-graph.h

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* process mrope_section on TextModel base class

* resolve conflict merge

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

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

@@ -11,7 +11,7 @@ body:
         (i.e. the generated text) are incorrect or llama.cpp crashes during model evaluation.
         If you encountered the issue while using an external UI (e.g. ollama),
         please reproduce your issue using one of the examples/binaries in this repository.
-        The `llama-cli` binary can be used for simple and reproducible model inference.
+        The `llama-completion` binary can be used for simple and reproducible model inference.
   - type: textarea
     id: version
     attributes:
@@ -74,9 +74,12 @@ body:
         Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
         that information would be very much appreciated by us.
+
+        If possible, please try to reproduce the issue using `llama-completion` with `-fit off`.
+        If you can only reproduce the issue with `-fit on`, please provide logs both with and without `--verbose`.
       placeholder: >
-        e.g. when I run llama-cli with -ngl 99 I get garbled outputs.
-        When I use -ngl 0 it works correctly.
+        e.g. when I run llama-completion with `-fa on` I get garbled outputs for very long prompts.
+        With short prompts or `-fa off` it works correctly.
         Here are the exact commands that I used: ...
     validations:
       required: true

.github/workflows/build.yml

@@ -20,7 +20,8 @@ on:
       '**/*.swift',
       '**/*.m',
       '**/*.metal',
-      '**/*.comp'
+      '**/*.comp',
+      '**/*.glsl'
     ]
 
   pull_request:
@@ -40,7 +41,8 @@ on:
       '**/*.swift',
       '**/*.m',
       '**/*.metal',
-      '**/*.comp'
+      '**/*.comp',
+      '**/*.glsl'
     ]
 
 concurrency:

.github/workflows/server-webui.yml

@@ -0,0 +1,295 @@
+# Server WebUI build and tests
+name: Server WebUI
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/server-webui.yml', 'tools/server/webui/**.*', 'tools/server/tests/**.*', 'tools/server/public/**']
+
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  webui-setup:
+    name: WebUI Setup
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+          cache: "npm"
+          cache-dependency-path: "tools/server/webui/package-lock.json"
+
+      - name: Cache node_modules
+        uses: actions/cache@v4
+        id: cache-node-modules
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Install dependencies
+        if: steps.cache-node-modules.outputs.cache-hit != 'true'
+        run: npm ci
+        working-directory: tools/server/webui
+
+  webui-check:
+    needs: webui-setup
+    name: WebUI Check
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Run type checking
+        run: npm run check
+        working-directory: tools/server/webui
+
+      - name: Run linting
+        run: npm run lint
+        working-directory: tools/server/webui
+
+  webui-build:
+    needs: webui-check
+    name: WebUI Build
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Build application
+        run: npm run build
+        working-directory: tools/server/webui
+
+  webui-tests:
+    needs: webui-build
+    name: Run WebUI tests
+    permissions:
+      contents: read
+
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+
+      - name: Restore node_modules cache
+        uses: actions/cache@v4
+        with:
+          path: tools/server/webui/node_modules
+          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
+          restore-keys: |
+            ${{ runner.os }}-node-modules-
+
+      - name: Install Playwright browsers
+        run: npx playwright install --with-deps
+        working-directory: tools/server/webui
+
+      - name: Build Storybook
+        run: npm run build-storybook
+        working-directory: tools/server/webui
+
+      - name: Run Client tests
+        run: npm run test:client
+        working-directory: tools/server/webui
+
+      - name: Run Server tests
+        run: npm run test:server
+        working-directory: tools/server/webui
+
+      - name: Run UI tests
+        run: npm run test:ui -- --testTimeout=60000
+        working-directory: tools/server/webui
+
+      - name: Run E2E tests
+        run: npm run test:e2e
+        working-directory: tools/server/webui
+
+  server-build:
+    needs: [webui-tests]
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        build_type: [RelWithDebInfo]
+        include:
+          - build_type: Release
+            sanitizer: ""
+      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+
+    steps:
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get -y install \
+            build-essential \
+            xxd \
+            git \
+            cmake \
+            curl \
+            wget \
+            language-pack-en \
+            libssl-dev
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Python setup
+        id: setup_python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Tests dependencies
+        id: test_dependencies
+        run: |
+          pip install -r tools/server/tests/requirements.txt
+
+      - name: Setup Node.js for WebUI
+        uses: actions/setup-node@v4
+        with:
+          node-version: "22"
+          cache: "npm"
+          cache-dependency-path: "tools/server/webui/package-lock.json"
+
+      - name: Install WebUI dependencies
+        run: npm ci
+        working-directory: tools/server/webui
+
+      - name: Build WebUI
+        run: npm run build
+        working-directory: tools/server/webui
+
+      - name: Build (no OpenMP)
+        id: cmake_build_no_openmp
+        if: ${{ matrix.sanitizer == 'THREAD' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+              -DGGML_OPENMP=OFF ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Build (sanitizers)
+        id: cmake_build_sanitizers
+        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Build (sanitizers)
+        id: cmake_build
+        if: ${{ matrix.sanitizer == '' }}
+        run: |
+          cmake -B build \
+              -DGGML_NATIVE=OFF \
+              -DLLAMA_CURL=OFF \
+              -DLLAMA_OPENSSL=ON \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ matrix.sanitizer == '' }}
+        env:
+          GITHUB_ACTIONS: "true"
+        run: |
+          cd tools/server/tests
+          ./tests.sh
+
+      - name: Tests (sanitizers)
+        id: server_integration_tests_sanitizers
+        if: ${{ matrix.sanitizer != '' }}
+        run: |
+          cd tools/server/tests
+          LLAMA_SANITIZE=1 ./tests.sh
+
+      - name: Slow tests
+        id: server_integration_tests_slow
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
+        run: |
+          cd tools/server/tests
+          SLOW_TESTS=1 ./tests.sh

.github/workflows/server.yml

@@ -76,270 +76,6 @@ jobs:
         run: |
           pip install -r tools/server/tests/requirements.txt
 
-  webui-setup:
-    name: WebUI Setup
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v4
-        with:
-          node-version: "22"
-          cache: "npm"
-          cache-dependency-path: "tools/server/webui/package-lock.json"
-
-      - name: Cache node_modules
-        uses: actions/cache@v4
-        id: cache-node-modules
-        with:
-          path: tools/server/webui/node_modules
-          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
-          restore-keys: |
-            ${{ runner.os }}-node-modules-
-
-      - name: Install dependencies
-        if: steps.cache-node-modules.outputs.cache-hit != 'true'
-        run: npm ci
-        working-directory: tools/server/webui
-
-  webui-check:
-    needs: webui-setup
-    name: WebUI Check
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v4
-        with:
-          node-version: "22"
-
-      - name: Restore node_modules cache
-        uses: actions/cache@v4
-        with:
-          path: tools/server/webui/node_modules
-          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
-          restore-keys: |
-            ${{ runner.os }}-node-modules-
-
-      - name: Run type checking
-        run: npm run check
-        working-directory: tools/server/webui
-
-      - name: Run linting
-        run: npm run lint
-        working-directory: tools/server/webui
-
-  webui-build:
-    needs: webui-check
-    name: WebUI Build
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v4
-        with:
-          node-version: "22"
-
-      - name: Restore node_modules cache
-        uses: actions/cache@v4
-        with:
-          path: tools/server/webui/node_modules
-          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
-          restore-keys: |
-            ${{ runner.os }}-node-modules-
-
-      - name: Build application
-        run: npm run build
-        working-directory: tools/server/webui
-
-  webui-tests:
-    needs: webui-build
-    name: Run WebUI tests
-    permissions:
-      contents: read
-
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-
-      - name: Setup Node.js
-        uses: actions/setup-node@v4
-        with:
-          node-version: "22"
-
-      - name: Restore node_modules cache
-        uses: actions/cache@v4
-        with:
-          path: tools/server/webui/node_modules
-          key: ${{ runner.os }}-node-modules-${{ hashFiles('tools/server/webui/package-lock.json') }}
-          restore-keys: |
-            ${{ runner.os }}-node-modules-
-
-      - name: Install Playwright browsers
-        run: npx playwright install --with-deps
-        working-directory: tools/server/webui
-
-      - name: Build Storybook
-        run: npm run build-storybook
-        working-directory: tools/server/webui
-
-      - name: Run Client tests
-        run: npm run test:client
-        working-directory: tools/server/webui
-
-      - name: Run Server tests
-        run: npm run test:server
-        working-directory: tools/server/webui
-
-      - name: Run UI tests
-        run: npm run test:ui -- --testTimeout=60000
-        working-directory: tools/server/webui
-
-      - name: Run E2E tests
-        run: npm run test:e2e
-        working-directory: tools/server/webui
-
-  server-build:
-    needs: [webui-tests]
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
-        build_type: [RelWithDebInfo]
-        include:
-          - build_type: Release
-            sanitizer: ""
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
-
-    steps:
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get -y install \
-            build-essential \
-            xxd \
-            git \
-            cmake \
-            curl \
-            wget \
-            language-pack-en \
-            libssl-dev
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Python setup
-        id: setup_python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
-
-      - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v4
-        with:
-          node-version: "22"
-          cache: "npm"
-          cache-dependency-path: "tools/server/webui/package-lock.json"
-
-      - name: Install WebUI dependencies
-        run: npm ci
-        working-directory: tools/server/webui
-
-      - name: Build WebUI
-        run: npm run build
-        working-directory: tools/server/webui
-
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_CURL=OFF \
-              -DLLAMA_OPENSSL=ON \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-              -DGGML_OPENMP=OFF ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build_sanitizers
-        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_CURL=OFF \
-              -DLLAMA_OPENSSL=ON \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build
-        if: ${{ matrix.sanitizer == '' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_CURL=OFF \
-              -DLLAMA_OPENSSL=ON \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ matrix.sanitizer == '' }}
-        env:
-          GITHUB_ACTIONS: "true"
-        run: |
-          cd tools/server/tests
-          ./tests.sh
-
-      - name: Tests (sanitizers)
-        id: server_integration_tests_sanitizers
-        if: ${{ matrix.sanitizer != '' }}
-        run: |
-          cd tools/server/tests
-          LLAMA_SANITIZE=1 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh
-
-
   server-windows:
     runs-on: windows-2022
 

.gitignore

@@ -54,6 +54,7 @@
 /out/
 /tmp/
 /autogen-*.md
+/common/build-info.cpp
 
 # Deprecated
 

CODEOWNERS

@@ -87,7 +87,8 @@
 /tests/                                 @ggerganov
 /tests/test-chat-.*                     @pwilkin
 /tools/batched-bench/                   @ggerganov
-/tools/main/                            @ggerganov
+/tools/cli/                             @ngxson
+/tools/completion/                      @ggerganov
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
 /tools/quantize/                        @ggerganov

README.md

@@ -313,7 +313,7 @@ The Hugging Face platform provides a variety of online tools for converting, qua
 
 To learn more about model quantization, [read this documentation](tools/quantize/README.md)
 
-## [`llama-cli`](tools/main)
+## [`llama-cli`](tools/cli)
 
 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
 
@@ -525,7 +525,8 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
 ## Other documentation
 
-- [main (cli)](tools/main/README.md)
+- [cli](tools/cli/README.md)
+- [completion](tools/completion/README.md)
 - [server](tools/server/README.md)
 - [GBNF grammars](grammars/README.md)
 

SECURITY.md

@@ -68,3 +68,6 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 Please note that using AI to identify vulnerabilities and generate reports is permitted. However, you must (1) explicitly disclose how AI was used and (2) conduct a thorough manual review before submitting the report.
 
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
+
+> [!IMPORTANT]
+> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080

ci/run.sh

@@ -398,6 +398,8 @@ function gg_run_qwen3_0_6b {
     ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
     ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
     (time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -523,6 +525,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
@@ -563,6 +567,8 @@ function gg_run_rerank_tiny {
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
     # for this model, the SEP token is "</s>"
     (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 

common/arg.cpp

@@ -20,6 +20,7 @@
 #include <nlohmann/json.hpp>
 
 #include <algorithm>
+#include <cinttypes>
 #include <climits>
 #include <cstdarg>
 #include <fstream>
@@ -529,7 +530,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.kv_overrides.back().key[0] = 0;
     }
 
-    if (!params.tensor_buft_overrides.empty()) {
+    // pad tensor_buft_overrides for llama_params_fit:
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+    while (params.tensor_buft_overrides.size() < ntbo) {
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
@@ -724,7 +727,7 @@ static void add_rpc_devices(const std::string & servers) {
     }
 }
 
-bool common_params_parse(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map) {
+bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map) {
     common_params dummy_params;
     common_params_context ctx_arg = common_params_parser_init(dummy_params, ex, nullptr);
 
@@ -733,6 +736,9 @@ bool common_params_parse(int argc, char ** argv, llama_example ex, std::map<comm
         for (const auto & arg : opt.args) {
             arg_to_options[arg] = &opt;
         }
+        for (const auto & arg : opt.args_neg) {
+            arg_to_options[arg] = &opt;
+        }
     }
 
     // TODO @ngxson : find a way to deduplicate this code
@@ -1412,7 +1418,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.top_k = value;
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K;
         }
-    ).set_sparam());
+    ).set_sparam().set_env("LLAMA_ARG_TOP_K"));
     add_opt(common_arg(
         {"--top-p"}, "N",
         string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
@@ -2150,6 +2156,34 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_MAIN_GPU"));
+    add_opt(common_arg(
+        { "-fit", "--fit" }, "[on|off]",
+        string_format("whether to adjust unset arguments to fit in device memory ('on' or 'off', default: '%s')", params.fit_params ? "on" : "off"),
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value)) {
+                params.fit_params = true;
+            } else if (is_falsey(value)) {
+                params.fit_params = false;
+            } else {
+                throw std::runtime_error(
+                    string_format("error: unkown value for --fit: '%s'\n", value.c_str()));
+            }
+        }
+    ).set_env("LLAMA_ARG_FIT"));
+    add_opt(common_arg(
+        { "-fitt", "--fit-target" }, "MiB",
+        string_format("target margin per device for --fit option, default: %zu", params.fit_params_target/(1024*1024)),
+        [](common_params & params, int value) {
+            params.fit_params_target = value * size_t(1024*1024);
+        }
+    ).set_env("LLAMA_ARG_FIT_TARGET"));
+    add_opt(common_arg(
+        { "-fitc", "--fit-ctx" }, "N",
+        string_format("minimum ctx size that can be set by --fit option, default: %" PRIu32, params.fit_params_min_ctx),
+        [](common_params & params, int value) {
+            params.fit_params_min_ctx = value;
+        }
+    ).set_env("LLAMA_ARG_FIT_CTX"));
     add_opt(common_arg(
         {"--check-tensors"},
         string_format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"),

common/arg.h

@@ -115,7 +115,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 
 // parse input arguments from CLI into a map
 // TODO: support repeated args in the future
-bool common_params_parse(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
+bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/chat-peg-parser.cpp

@@ -4,9 +4,14 @@
 
 using json = nlohmann::json;
 
-static std::string_view trim_trailing_space(std::string_view sv) {
+static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
+    int count = 0;
     while (!sv.empty() && std::isspace(static_cast<unsigned char>(sv.back()))) {
+        if (max != -1 && count <= max) {
+            break;
+        }
         sv.remove_suffix(1);
+        count++;
     }
     return sv;
 }
@@ -93,7 +98,7 @@ void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
 
     if (is_arg_string && current_tool) {
         // Serialize to JSON, but exclude the end quote
-        std::string dumped = json(node.text).dump();
+        std::string dumped = json(trim_trailing_space(node.text)).dump();
         current_tool->arguments += dumped.substr(0, dumped.size() - 1);
         needs_closing_quote = true;
     }
@@ -101,6 +106,7 @@ void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
     if (is_arg_close && current_tool) {
         if (needs_closing_quote) {
             current_tool->arguments += "\"";
+            needs_closing_quote = false;
         }
     }
 
@@ -109,6 +115,10 @@ void common_chat_peg_constructed_mapper::map(const common_peg_ast_node & node) {
     }
 
     if (is_tool_close && current_tool) {
+        if (needs_closing_quote) {
+            current_tool->arguments += "\"";
+            needs_closing_quote = false;
+        }
         current_tool->arguments += "}";
     }
 }

common/chat.cpp

@@ -711,6 +711,25 @@ static void foreach_function(const json & tools, const std::function<void(const
     }
 }
 
+static void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn) {
+    if (!function.contains("parameters") || !function.at("parameters").is_object()) {
+        return;
+    }
+    const auto & params = function.at("parameters");
+    if (!params.contains("properties") || !params.at("properties").is_object()) {
+        return;
+    }
+    const auto & props = params.at("properties");
+    std::set<std::string> required;
+    if (params.contains("required") && params.at("required").is_array()) {
+        params.at("required").get_to(required);
+    }
+    for (const auto & [name, prop] : props.items()) {
+        bool is_required = (required.find(name) != required.end());
+        fn(name, prop, is_required);
+    }
+}
+
 static std::string apply(
     const common_chat_template & tmpl,
     const struct templates_params & inputs,
@@ -1409,6 +1428,123 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
     return data;
 }
 
+static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    data.prompt = apply(tmpl, inputs);
+    data.format = COMMON_CHAT_FORMAT_PEG_CONSTRUCTED;
+
+    // Handle thinking tags appropriately based on inputs.enable_thinking
+    if (string_ends_with(data.prompt, "<think>\n")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+
+    data.preserved_tokens = {
+        "<think>",
+        "</think>",
+        "<tool_call>",
+        "</tool_call>",
+    };
+
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar = true;
+
+    auto parser = build_chat_peg_constructed_parser([&](auto & p) {
+        auto reasoning = p.eps();
+        if (inputs.enable_thinking && extract_reasoning) {
+            auto reasoning_content = p.reasoning(p.until("</think>")) + ("</think>" | p.end());
+            if (data.thinking_forced_open) {
+                reasoning = reasoning_content;
+            }
+        }
+
+        // Response format parser
+        if (inputs.json_schema.is_object() && !inputs.json_schema.empty()) {
+            return reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema));
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(parameters);
+
+                auto tool_open = "<function=" + p.tool_name(p.literal(name)) + ">\n";
+                auto tool_close = p.literal("</function>\n");
+                auto args = p.sequence();
+                auto arg_string = p.rule("xml-arg-string", p.until_one_of({
+                    "\n</parameter>",
+                    "\n<parameter=",
+                    "\n</function>"
+                }));
+
+                foreach_parameter(function, [&](const auto & param_name, const json & param_schema, bool is_required) {
+                    auto rule_name = "tool-" + name + "-arg-" + param_name;
+
+                    auto arg_open = "<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n";
+                    auto arg_close = p.literal("</parameter>\n");
+                    auto arg_value = p.eps();
+
+                    if (schema_info.resolves_to_string(param_schema)) {
+                        arg_value = p.tool_arg_string_value(arg_string) + "\n";
+                    } else {
+                        arg_value = p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema));
+                    }
+
+                    // Model may or my not close with </parameter>
+                    auto arg_rule = p.rule(rule_name, p.tool_arg_open(arg_open) + arg_value + p.optional(p.tool_arg_close(arg_close)));
+                    args += p.repeat(arg_rule, /* min = */ is_required ? 1 : 0, /* max = */ 1);
+                });
+
+                tool_choice |= p.rule("tool-" + name, p.tool_open(tool_open) + args + p.tool_close(tool_close));
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = inputs.parallel_tool_calls ? -1 : 1;
+            auto tool_call = p.rule("tool-call", "<tool_call>\n" + tool_choice + "</tool_call>" + p.space());
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
+
+            return reasoning << p.content(p.until("<tool_call>")) << tool_calls;
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return reasoning << p.content(p.rest());
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>"}
+        };
+    }
+
+    return data;
+}
+
+
 static common_chat_params common_chat_params_init_apertus(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
@@ -2534,6 +2670,10 @@ static common_chat_params common_chat_templates_apply_jinja(
         src.find("<function=") != std::string::npos &&
         src.find("<parameters>") != std::string::npos &&
         src.find("<parameter=") != std::string::npos) {
+        // Nemotron 3 Nano 30B A3B
+        if (src.find("<think>") != std::string::npos) {
+            return common_chat_params_init_nemotron_v3(tmpl, params);
+        }
         return common_chat_params_init_qwen3_coder_xml(tmpl, params);
     }
 

common/common.cpp

@@ -1013,31 +1013,40 @@ bool tty_can_use_colors() {
 // Model utils
 //
 
-static inline void common_init_sampler_from_model(
+// TODO: move to common/sampling
+static void common_init_sampler_from_model(
     const llama_model * model,
     common_params_sampling & sparams) {
 
     const uint64_t config = sparams.user_sampling_config;
 
     auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[64] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             int32_t v = strtol(buf, &end, 10);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
     auto get_float = [&](const char * key, float & dst, uint64_t user_config) {
-        if (config & user_config) return;
+        if (config & user_config) {
+            return;
+        }
 
         char buf[128] = {0};
         if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) {
             char * end = nullptr;
             float v = strtof(buf, &end);
-            if (end && end != buf) dst = v;
+            if (end && end != buf) {
+                dst = v;
+            }
         }
     };
 
@@ -1065,31 +1074,125 @@ static inline void common_init_sampler_from_model(
     get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA),    sparams.mirostat_eta,    common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA);
 }
 
-struct common_init_result common_init_from_params(common_params & params) {
-    common_init_result iparams;
+struct common_init_result::impl {
+    impl() = default;
+    ~impl() = default;
+
+    llama_model_ptr   model;
+    llama_context_ptr context;
+
+    std::vector<llama_adapter_lora_ptr> lora;
+
+    std::vector<common_sampler_ptr> samplers;
+};
+
+common_init_result::common_init_result(common_params & params) :
+    pimpl(new impl{}) {
     auto mparams = common_model_params_to_llama(params);
+    auto cparams = common_context_params_to_llama(params);
+
+    if (params.fit_params) {
+        LOG_INF("%s: fitting params to device memory, to report bugs during this step use -fit off (or --verbose if you can't)\n", __func__);
+        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
+    }
 
     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', try reducing --n-gpu-layers if you're running out of VRAM\n",
-            __func__, params.model.path.c_str());
-        return iparams;
+        return;
     }
 
-    common_init_sampler_from_model(model, params.sampling);
+    pimpl->model.reset(model);
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    auto cparams = common_context_params_to_llama(params);
+    // updates params.sampling
+    // TODO: fix naming
+    common_init_sampler_from_model(model, params.sampling);
+
+    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;
+    }
+
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(vocab, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
+        }
+    }
+
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
+    //if (params.sampling.penalty_last_n == -1) {
+    //    LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    //if (params.sampling.dry_penalty_last_n == -1) {
+    //    LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
+    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
+    //}
+
+    pimpl->samplers.resize(cparams.n_seq_max);
+
+    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+    }
 
     llama_context * lctx = llama_init_from_model(model, cparams);
     if (lctx == NULL) {
-        LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
-            __func__, params.model.path.c_str());
-        llama_model_free(model);
-        return iparams;
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        return;
     }
 
+    pimpl->context.reset(lctx);
+}
+
+llama_model * common_init_result::model() {
+    return pimpl->model.get();
+}
+
+llama_context * common_init_result::context() {
+    return pimpl->context.get();
+}
+
+common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
+    return pimpl->samplers[seq_id].get();
+}
+
+std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
+    return pimpl->lora;
+}
+
+void common_init_result::free_context() {
+    pimpl->context.reset();
+}
+
+common_init_result_ptr common_init_from_params(common_params & params) {
+    common_init_result_ptr res(new common_init_result(params));
+
+    llama_model * model = res->model();
+    if (model == NULL) {
+        LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    llama_context * lctx = res->context();
+    if (lctx == NULL) {
+        LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
+        return res;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
     if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
@@ -1101,10 +1204,7 @@ struct common_init_result common_init_from_params(common_params & params) {
 
         const auto cvec = common_control_vector_load(params.control_vectors);
         if (cvec.n_embd == -1) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
+            return res;
         }
 
         int err = llama_apply_adapter_cvec(
@@ -1115,10 +1215,7 @@ struct common_init_result common_init_from_params(common_params & params) {
                 params.control_vector_layer_start,
                 params.control_vector_layer_end);
         if (err) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
+            return res;
         }
     }
 
@@ -1142,10 +1239,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (!ok) {
-            llama_free(lctx);
-            llama_model_free(model);
-
-            return iparams;
+            return res;
         }
     }
 
@@ -1155,9 +1249,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         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_model_free(model);
-            return iparams;
+            return res;
         }
 
         char buf[1024];
@@ -1166,43 +1258,13 @@ struct common_init_result common_init_from_params(common_params & params) {
         la.task_name = buf;
         llama_adapter_meta_val_str(la.ptr, "adapter.lora.prompt_prefix", buf, sizeof(buf));
         la.prompt_prefix = buf;
-        iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
+        res->lora().emplace_back(std::move(lora)); // copy to list of loaded adapters
     }
 
     if (!params.lora_init_without_apply) {
         common_set_adapter_lora(lctx, params.lora_adapters);
     }
 
-    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;
-    }
-
-    // initialize once
-    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-        if (llama_vocab_is_eog(vocab, i)) {
-            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
-        }
-    }
-
-    if (params.sampling.ignore_eos) {
-        // add EOG biases to the active set of logit biases
-        params.sampling.logit_bias.insert(
-                params.sampling.logit_bias.end(),
-                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
-    }
-
-    if (params.sampling.penalty_last_n == -1) {
-        LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    }
-
-    if (params.sampling.dry_penalty_last_n == -1) {
-        LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-        params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    }
-
     if (params.warmup) {
         LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
 
@@ -1241,12 +1303,11 @@ struct common_init_result common_init_from_params(common_params & params) {
         llama_set_warmup(lctx, false);
     }
 
-    iparams.model.reset(model);
-    iparams.context.reset(lctx);
-
-    return iparams;
+    return res;
 }
 
+common_init_result::~common_init_result() = default;
+
 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.
@@ -1255,7 +1316,9 @@ std::string get_model_endpoint() {
     std::string model_endpoint = "https://huggingface.co/";
     if (endpoint_env) {
         model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') model_endpoint += '/';
+        if (model_endpoint.back() != '/') {
+            model_endpoint += '/';
+        }
     }
     return model_endpoint;
 }

common/common.h

@@ -99,6 +99,7 @@ enum llama_example {
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
     LLAMA_EXAMPLE_FINETUNE,
+    LLAMA_EXAMPLE_FIT_PARAMS,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -195,7 +196,6 @@ struct common_params_sampling {
 
     std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
 
-
     std::vector<enum common_sampler_type> samplers = {
         COMMON_SAMPLER_TYPE_PENALTIES,
         COMMON_SAMPLER_TYPE_DRY,
@@ -216,6 +216,10 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    bool has_logit_bias() const {
+        return !logit_bias.empty();
+    }
+
     // print the parameters into a string
     std::string print() const;
 };
@@ -303,8 +307,8 @@ struct lr_opt {
 struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
 
 struct common_params {
-    int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =  4096; // context size
+    int32_t n_predict             =    -1; // max. number of new tokens to predict, -1 == no limit
+    int32_t n_ctx                 =     0; // context size, 0 == context the model was trained with
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -325,9 +329,12 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers      = -1;  // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu          = 0;   // the GPU that is used for scratch and small tensors
-    float   tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
+    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
+    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
+    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
+    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -669,15 +676,29 @@ bool tty_can_use_colors();
 // Model utils
 //
 
-// note: defines object's lifetime
-struct common_init_result {
-    llama_model_ptr   model;
-    llama_context_ptr context;
+struct common_sampler;
 
-    std::vector<llama_adapter_lora_ptr> lora;
+// note: defines the model, context, samplers, ets. lifetimes
+struct common_init_result {
+    common_init_result(common_params & params);
+    ~common_init_result();
+
+    llama_model * model();
+    llama_context * context();
+    common_sampler * sampler(llama_seq_id seq_id);
+
+    std::vector<llama_adapter_lora_ptr> & lora();
+
+    void free_context();
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 };
 
-struct common_init_result     common_init_from_params(common_params & params);
+using common_init_result_ptr = std::unique_ptr<common_init_result>;
+
+common_init_result_ptr common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);

common/json-schema-to-grammar.cpp

@@ -305,8 +305,9 @@ static std::string format_literal(const std::string & literal) {
 
 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
 
-class SchemaConverter {
+class common_schema_converter {
 private:
+    friend class common_schema_info;
     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;
@@ -729,7 +730,7 @@ private:
     }
 
 public:
-    SchemaConverter(
+    common_schema_converter(
         const std::function<json(const std::string &)> & fetch_json,
         bool dotall)
           : _fetch_json(fetch_json), _dotall(dotall)
@@ -990,6 +991,134 @@ public:
     }
 };
 
+// common_schema_info implementation (pimpl)
+
+common_schema_info::common_schema_info()
+    : impl_(std::make_unique<common_schema_converter>(
+        [](const std::string &) { return json(); },
+        false)) {}
+
+common_schema_info::~common_schema_info() = default;
+
+common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
+common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
+
+void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
+    impl_->resolve_refs(schema, "");
+}
+
+// Determines if a JSON schema can resolve to a string type through any path.
+// Some models emit raw string values rather than JSON-encoded strings for string parameters.
+// If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
+// true, allowing callers to handle the value as a raw string for simplicity.
+bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
+    std::unordered_set<std::string> visited_refs;
+
+    std::function<bool(const json &)> check = [&](const json & s) -> bool {
+        if (!s.is_object()) {
+            return false;
+        }
+
+        // Handle $ref
+        if (s.contains("$ref")) {
+            const std::string & ref = s["$ref"];
+            if (visited_refs.find(ref) != visited_refs.end()) {
+                // Circular reference, assume not a string to be safe
+                return false;
+            }
+            visited_refs.insert(ref);
+            auto it = impl_->_refs.find(ref);
+            if (it != impl_->_refs.end()) {
+                return check(it->second);
+            }
+            return false;
+        }
+
+        // Check type field
+        if (s.contains("type")) {
+            const json & schema_type = s["type"];
+            if (schema_type.is_string()) {
+                if (schema_type == "string") {
+                    return true;
+                }
+            } else if (schema_type.is_array()) {
+                // Type can be an array like ["string", "null"]
+                for (const auto & t : schema_type) {
+                    if (t == "string") {
+                        return true;
+                    }
+                }
+            }
+        }
+
+        // Check oneOf/anyOf - if any alternative can be a string
+        if (s.contains("oneOf")) {
+            for (const auto & alt : s["oneOf"]) {
+                if (check(alt)) {
+                    return true;
+                }
+            }
+        }
+        if (s.contains("anyOf")) {
+            for (const auto & alt : s["anyOf"]) {
+                if (check(alt)) {
+                    return true;
+                }
+            }
+        }
+
+        // Check allOf - all components must be compatible with string type
+        if (s.contains("allOf")) {
+            bool all_string = true;
+            for (const auto & component : s["allOf"]) {
+                if (!check(component)) {
+                    all_string = false;
+                    break;
+                }
+            }
+            if (all_string) {
+                return true;
+            }
+        }
+
+        // Check const - if the constant value is a string
+        if (s.contains("const")) {
+            if (s["const"].is_string()) {
+                return true;
+            }
+        }
+
+        // Check enum - if any enum value is a string
+        if (s.contains("enum")) {
+            for (const auto & val : s["enum"]) {
+                if (val.is_string()) {
+                    return true;
+                }
+            }
+        }
+
+        // String-specific keywords imply string type
+        if (s.contains("pattern") || s.contains("minLength") || s.contains("maxLength")) {
+            return true;
+        }
+
+        // Check format - many formats imply string
+        if (s.contains("format")) {
+            const std::string & fmt = s["format"];
+            if (fmt == "date" || fmt == "time" || fmt == "date-time" ||
+                fmt == "uri" || fmt == "email" || fmt == "hostname" ||
+                fmt == "ipv4" || fmt == "ipv6" || fmt == "uuid" ||
+                fmt.find("uuid") == 0) {
+                return true;
+            }
+        }
+
+        return false;
+    };
+
+    return check(schema);
+}
+
 std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
@@ -1006,7 +1135,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
 }
 
 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_schema_converter 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);

common/json-schema-to-grammar.h

@@ -3,11 +3,31 @@
 #include <nlohmann/json_fwd.hpp>
 
 #include <functional>
+#include <memory>
 #include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);
 
+class common_schema_converter;
+
+// Probes a JSON schema to extract information about its structure and type constraints.
+class common_schema_info {
+    std::unique_ptr<common_schema_converter> impl_;
+
+  public:
+    common_schema_info();
+    ~common_schema_info();
+
+    common_schema_info(const common_schema_info &) = delete;
+    common_schema_info & operator=(const common_schema_info &) = delete;
+    common_schema_info(common_schema_info &&) noexcept;
+    common_schema_info & operator=(common_schema_info &&) noexcept;
+
+    void resolve_refs(nlohmann::ordered_json & schema);
+    bool resolves_to_string(const nlohmann::ordered_json & schema);
+};
+
 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;

common/peg-parser.cpp

@@ -425,7 +425,7 @@ struct parser_executor {
 
         if (result.need_more_input()) {
             // Propagate - need to know what child would match before negating
-            return result;
+            return common_peg_parse_result(COMMON_PEG_PARSE_RESULT_NEED_MORE_INPUT, start_pos);
         }
 
         // Child failed, so negation succeeds

common/preset.cpp

@@ -157,6 +157,21 @@ static std::map<std::string, common_arg> get_map_key_opt(common_params_context &
     return mapping;
 }
 
+static bool is_bool_arg(const common_arg & arg) {
+    return !arg.args_neg.empty();
+}
+
+static std::string parse_bool_arg(const common_arg & arg, const std::string & key, const std::string & value) {
+    // if this is a negated arg, we need to reverse the value
+    for (const auto & neg_arg : arg.args_neg) {
+        if (rm_leading_dashes(neg_arg) == key) {
+            return common_arg_utils::is_truthy(value) ? "false" : "true";
+        }
+    }
+    // otherwise, not negated
+    return value;
+}
+
 common_presets common_presets_load(const std::string & path, common_params_context & ctx_params) {
     common_presets out;
     auto key_to_opt = get_map_key_opt(ctx_params);
@@ -173,8 +188,13 @@ common_presets common_presets_load(const std::string & path, common_params_conte
         for (const auto & [key, value] : section.second) {
             LOG_DBG("option: %s = %s\n", key.c_str(), value.c_str());
             if (key_to_opt.find(key) != key_to_opt.end()) {
-                preset.options[key_to_opt[key]] = value;
-                LOG_DBG("accepted option: %s = %s\n", key.c_str(), value.c_str());
+                auto & opt = key_to_opt[key];
+                if (is_bool_arg(opt)) {
+                    preset.options[opt] = parse_bool_arg(opt, key, value);
+                } else {
+                    preset.options[opt] = value;
+                }
+                LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
             } else {
                 // TODO: maybe warn about unknown key?
             }

common/sampling.cpp

@@ -104,9 +104,10 @@ struct ring_buffer {
 struct common_sampler {
     common_params_sampling params;
 
-    struct llama_sampler * grmr;
     struct llama_sampler * chain;
 
+    bool grammar;
+
     ring_buffer<llama_token> prev;
 
     std::vector<llama_token_data> cur;
@@ -116,7 +117,6 @@ struct common_sampler {
     void reset() {
         prev.clear();
 
-        llama_sampler_reset(grmr);
         llama_sampler_reset(chain);
     }
 
@@ -167,10 +167,15 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     lparams.no_perf = params.no_perf;
 
-    struct llama_sampler * grmr;
+    llama_sampler * chain = llama_sampler_chain_init(lparams);
+
+    bool grammar = false;
+    std::vector<llama_sampler *> samplers;
+
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
-        grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
+        samplers.push_back(llama_sampler_init_llg(vocab, "lark", params.grammar.c_str()));
+        grammar = true;
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
@@ -217,30 +222,23 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
             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;
+        if (!params.grammar.empty()) {
+             if (params.grammar_lazy) {
+                 samplers.push_back(
+                         llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root",
+                             trigger_patterns_c.data(), trigger_patterns_c.size(),
+                             trigger_tokens.data(),     trigger_tokens.size()));
+             } else {
+                 samplers.push_back(llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"));
+             }
+
+             grammar = true;
         }
     }
 
-    auto * result = new common_sampler {
-        /* .params = */ params,
-        /* .grmr   = */ grmr,
-        /* .chain  = */ llama_sampler_chain_init(lparams),
-        /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
-        /* .cur    = */ {},
-        /* .cur_p  = */ {},
-    };
-
-    llama_sampler_chain_add(result->chain,
-            llama_sampler_init_logit_bias(
-                llama_vocab_n_tokens(vocab),
-                params.logit_bias.size(),
-                params.logit_bias.data()));
+    if (params.has_logit_bias()) {
+        samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), params.logit_bias.size(), params.logit_bias.data()));
+    }
 
     if (params.mirostat == 0) {
         for (const auto & cnstr : params.samplers) {
@@ -253,58 +251,70 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                             c_breakers.push_back(str.c_str());
                         }
 
-                        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()));
+                        samplers.push_back(llama_sampler_init_dry    (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                     }
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k       (params.top_k));
+                    samplers.push_back(llama_sampler_init_top_k      (params.top_k));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p       (params.top_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_top_p      (params.top_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TOP_N_SIGMA:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma));
+                    samplers.push_back(llama_sampler_init_top_n_sigma(params.top_n_sigma));
                     break;
                 case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p       (params.min_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_min_p      (params.min_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc         (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    samplers.push_back(llama_sampler_init_xtc        (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
                     break;
                 case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical     (params.typ_p, params.min_keep));
+                    samplers.push_back(llama_sampler_init_typical    (params.typ_p, params.min_keep));
                     break;
                 case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext    (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    samplers.push_back(llama_sampler_init_temp_ext   (params.temp, params.dynatemp_range, params.dynatemp_exponent));
                     break;
                 case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill      (vocab));
+                    samplers.push_back(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));
+                    samplers.push_back(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));
+
+        samplers.push_back(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_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+        samplers.push_back(llama_sampler_init_temp(params.temp));
+        samplers.push_back(llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
     } 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));
+        samplers.push_back(llama_sampler_init_temp(params.temp));
+        samplers.push_back(llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
     } else {
         GGML_ASSERT(false && "unknown mirostat version");
     }
 
+    for (auto * smpl : samplers) {
+        llama_sampler_chain_add(chain, smpl);
+    }
+
+    auto * result = new common_sampler {
+        /* .params  = */ params,
+        /* .chain   = */ chain,
+        /* .grammar = */ grammar,
+        /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
+        /* .cur     = */ {},
+        /* .cur_p   = */ {},
+    };
+
     return result;
 }
 
 void common_sampler_free(struct common_sampler * gsmpl) {
     if (gsmpl) {
-        llama_sampler_free(gsmpl->grmr);
-
         llama_sampler_free(gsmpl->chain);
 
         delete gsmpl;
@@ -314,11 +324,24 @@ void common_sampler_free(struct common_sampler * gsmpl) {
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
     const auto tm = gsmpl->tm();
 
-    if (accept_grammar) {
-        llama_sampler_accept(gsmpl->grmr, token);
-    }
+    if (gsmpl->grammar) {
+        const int n_smpl = llama_sampler_chain_n(gsmpl->chain);
 
-    llama_sampler_accept(gsmpl->chain, token);
+        for (int i = 0; i < n_smpl; i++) {
+            auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
+
+            // the grammar sampler is always the first one
+            if (i == 0) {
+                if (accept_grammar) {
+                    llama_sampler_accept(smpl, token);
+                }
+            } else {
+                llama_sampler_accept(smpl, token);
+            }
+        }
+    } else {
+        llama_sampler_accept(gsmpl->chain, token);
+    }
 
     gsmpl->prev.push_back(token);
 }
@@ -329,12 +352,12 @@ void common_sampler_reset(struct common_sampler * gsmpl) {
 
 struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
-        /* .params = */ gsmpl->params,
-        /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
-        /* .chain  = */ llama_sampler_clone(gsmpl->chain),
-        /* .prev   = */ gsmpl->prev,
-        /* .cur    = */ gsmpl->cur,
-        /* .cur_p  = */ gsmpl->cur_p,
+        /* .params  = */ gsmpl->params,
+        /* .chain   = */ llama_sampler_clone(gsmpl->chain),
+        /* .grammar = */ gsmpl->grammar,
+        /* .prev    = */ gsmpl->prev,
+        /* .cur     = */ gsmpl->cur,
+        /* .cur_p   = */ gsmpl->cur_p,
     };
 }
 
@@ -383,58 +406,33 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
     }
 }
 
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl) {
+    return gsmpl->chain;
+}
+
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx) {
     llama_synchronize(ctx);
 
     // start measuring sampling time after the llama_context synchronization in order to not measure any ongoing async operations
     const auto tm = gsmpl->tm();
 
-    gsmpl->set_logits(ctx, idx);
+    llama_token id = LLAMA_TOKEN_NULL;
 
-    auto & grmr  = gsmpl->grmr;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
-    if (grammar_first) {
-        llama_sampler_apply(grmr, &cur_p);
-    }
+    gsmpl->set_logits(ctx, idx);
 
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
 
-    const llama_token id = cur_p.data[cur_p.selected].id;
+    id = cur_p.data[cur_p.selected].id;
 
-    if (grammar_first) {
-        return id;
-    }
-
-    // check if it the sampled token fits the grammar
-    {
-        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
-
-        llama_sampler_apply(grmr, &single_token_data_array);
-
-        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-        if (is_valid) {
-            return id;
-        }
-    }
-
-    // resampling:
-    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
-    gsmpl->set_logits(ctx, idx);
-
-    llama_sampler_apply(grmr,  &cur_p);
-    llama_sampler_apply(chain, &cur_p);
-
-    GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
-
-    return cur_p.data[cur_p.selected].id;
+    return id;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft) {
     GGML_ASSERT(idxs.size() == draft.size() + 1 && "idxs.size() must be draft.size() + 1");
 
     std::vector<llama_token> result;
@@ -442,7 +440,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
 
     size_t i = 0;
     for (; i < draft.size(); i++) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -454,7 +452,7 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     }
 
     if (i == draft.size()) {
-        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i], grammar_first);
+        const llama_token id = common_sampler_sample(gsmpl, ctx, idxs[i]);
 
         common_sampler_accept(gsmpl, id, true);
 
@@ -464,13 +462,13 @@ std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sample
     return result;
 }
 
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first) {
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft) {
     std::vector<int> idxs(draft.size() + 1);
     for (size_t i = 0; i < idxs.size(); ++i) {
         idxs[i] = i;
     }
 
-    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft, grammar_first);
+    return common_sampler_sample_and_accept_n(gsmpl, ctx, idxs, draft);
 }
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl) {
@@ -515,7 +513,8 @@ std::string common_sampler_print(const struct common_sampler * gsmpl) {
 
     for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
         const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
+        result += std::string("-> ");
+        result += std::string(llama_sampler_name(smpl)) + " ";
     }
 
     return result;

common/sampling.h

@@ -48,6 +48,8 @@ struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 // arguments can be nullptr to skip printing
 void common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl);
 
+struct llama_sampler * common_sampler_get(const struct common_sampler * gsmpl);
+
 // extended sampling implementation:
 //
 // - set logits
@@ -55,10 +57,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
 // - check if the token fits the grammar (if any)
 // - if not: resample by first applying the grammar constraints and then sampling again (slower path)
 //
-// if grammar_first is true, the grammar is applied before the samplers (slower)
-// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
-//
-llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
+llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx);
 
 // generalized version of common_sampler_sample
 //
@@ -76,10 +75,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 //
 // returns at least 1 token, up to idxs.size()
 //
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft, bool grammar_first = false);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const std::vector<int> & idxs, const llama_tokens & draft);
 
 // assume idxs == [ 0, 1, 2, ..., draft.size() ]
-std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft, bool grammar_first = false);
+std::vector<llama_token> common_sampler_sample_and_accept_n(struct common_sampler * gsmpl, struct llama_context * ctx, const llama_tokens & draft);
 
 uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
 
@@ -107,3 +106,9 @@ std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std:
 
 llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
                 const char * grammar_kind, const char * grammar_data);
+
+struct common_sampler_deleter {
+    void operator()(common_sampler * s) { common_sampler_free(s); }
+};
+
+typedef std::unique_ptr<common_sampler, common_sampler_deleter> common_sampler_ptr;

common/speculative.cpp

@@ -315,7 +315,7 @@ llama_tokens common_speculative_gen_draft(
     for (int i = 0; i < params.n_draft; ++i) {
         common_batch_clear(batch);
 
-        common_sampler_sample(smpl, ctx_dft, 0, true);
+        common_sampler_sample(smpl, ctx_dft, 0);
 
         const auto * cur_p = common_sampler_get_candidates(smpl, true);
 

convert_hf_to_gguf_update.py

@@ -143,6 +143,7 @@ models = [
     {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
     {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
+    {"name": "kormo",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/KORMo-Team/KORMo-tokenizer", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/backend/SYCL.md

@@ -103,6 +103,8 @@ SYCL backend supports Intel GPU Family:
 - Intel Built-in Arc GPU
 - Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
 
+On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the performance is not optimal, and some GPUs may not support OpenCL nor have any GPGPU capabilities.
+
 #### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |

docs/development/HOWTO-add-model.md

@@ -9,7 +9,8 @@ Adding a model requires few steps:
 After following these steps, you can open PR.
 
 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
-- [main](/tools/main/)
+- [cli](/tools/cli/)
+- [completion](/tools/completion/)
 - [imatrix](/tools/imatrix/)
 - [quantize](/tools/quantize/)
 - [server](/tools/server/)

docs/ops.md

@@ -18,12 +18,12 @@ Legend:
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
@@ -31,7 +31,7 @@ Legend:
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -64,7 +64,7 @@ Legend:
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
@@ -98,14 +98,14 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -113,7 +113,7 @@ Legend:
 |                              SUM | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |

examples/batched/batched.cpp

@@ -2,6 +2,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "sampling.h"
 
 #include <algorithm>
 #include <cstdio>
@@ -64,17 +65,23 @@ int main(int argc, char ** argv) {
     ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_predict, n_parallel);
 
-    llama_context * ctx = llama_init_from_model(model, ctx_params);
-
     auto sparams = llama_sampler_chain_default_params();
     sparams.no_perf = false;
 
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+    std::vector<llama_sampler *> samplers;
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sampling.top_k));
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sampling.top_p, params.sampling.min_keep));
-    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sampling.temp));
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sampling.seed));
+    for (int32_t i = 0; i < n_parallel; ++i) {
+        llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+        llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sampling.top_k));
+        llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sampling.top_p, params.sampling.min_keep));
+        llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sampling.temp));
+        llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sampling.seed));
+
+        samplers.push_back(smpl);
+    }
+
+    llama_context * ctx = llama_init_from_model(model, ctx_params);
 
     if (ctx == NULL) {
         LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
@@ -173,7 +180,7 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
+            const llama_token new_token_id = llama_sampler_sample(samplers[i], ctx, i_batch[i]);
 
             // is it an end of generation? -> mark the stream as finished
             if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_predict) {
@@ -229,14 +236,17 @@ int main(int argc, char ** argv) {
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
     LOG("\n");
-    llama_perf_sampler_print(smpl);
+    llama_perf_sampler_print(samplers[0]);
     llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
 
-    llama_sampler_free(smpl);
+    for (auto & sampler_config : samplers) {
+        llama_sampler_free(sampler_config);
+    }
+
     llama_free(ctx);
     llama_model_free(model);
 

examples/embedding/embedding.cpp

@@ -131,10 +131,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx = llama_init->context();
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);

examples/eval-callback/eval-callback.cpp

@@ -202,10 +202,10 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     if (model == nullptr || ctx == nullptr) {
         LOG_ERR("%s : failed to init\n", __func__);

examples/lookahead/lookahead.cpp

@@ -55,10 +55,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the target model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     auto * mem = llama_get_memory(ctx);
 

examples/lookup/lookup-create.cpp

@@ -18,16 +18,16 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model_ptr & model = llama_init.model;
-    llama_context_ptr & ctx = llama_init.context;
+    auto * model = llama_init->model();
+    auto * ctx = llama_init->context();
 
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx.get(), params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
     fprintf(stderr, "%s: tokenization done\n", __func__);
 
     common_ngram_cache ngram_cache;

examples/lookup/lookup-stats.cpp

@@ -28,13 +28,13 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_context_ptr & ctx = llama_init.context;
+    llama_context * ctx = llama_init->context();
 
     // tokenize the prompt
     std::vector<llama_token> inp;
-    inp = common_tokenize(ctx.get(), params.prompt, true, true);
+    inp = common_tokenize(ctx, params.prompt, true, true);
 
     common_ngram_cache ngram_cache_context;
     common_ngram_cache ngram_cache_dynamic;
@@ -65,7 +65,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = llama_n_ctx(ctx.get());
+    const int n_ctx = llama_n_ctx(ctx);
 
     int n_drafted = 0;
     int n_accept  = 0;

examples/lookup/lookup.cpp

@@ -29,10 +29,10 @@ int main(int argc, char ** argv){
     llama_numa_init(params.numa);
 
     // load the model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 

examples/model-conversion/README.md

@@ -10,6 +10,13 @@ and in some cases perplexity checked of the quantized model. And finally the
 model/models need to the ggml-org on Hugging Face. This tool/example tries to
 help with this process.
 
+> 📝 **Note:** When adding a new model from an existing family, verify the
+> previous version passes logits verification first. Existing models can have
+> subtle numerical differences that don't affect generation quality but cause
+> logits mismatches. Identifying these upfront whether they exist in llama.cpp,
+> the conversion script, or in an upstream implementation, can save significant
+> debugging time.
+
 ### Overview
 The idea is that the makefile targets and scripts here can be used in the
 development/conversion process assisting with things like:

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

@@ -200,7 +200,7 @@ with torch.no_grad():
     logits = outputs.logits
 
     # Extract logits for the last token (next token prediction)
-    last_logits = logits[0, -1, :].cpu().numpy()
+    last_logits = logits[0, -1, :].float().cpu().numpy()
 
     print(f"Logits shape: {logits.shape}")
     print(f"Last token logits shape: {last_logits.shape}")

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

@@ -34,8 +34,11 @@ done
 MODEL_PATH="${MODEL_PATH:-"$EMBEDDING_MODEL_PATH"}"
 MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
 
+CONVERTED_MODEL_PATH="${CONVERTED_EMBEDDING_PATH:-"$CONVERTED_EMBEDDING_MODEL"}"
+CONVERTED_MODEL_NAME="${CONVERTED_MODEL_NAME:-$(basename "$CONVERTED_MODEL_PATH" .gguf)}"
+
 if [ -t 0 ]; then
-    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
+    CPP_EMBEDDINGS="data/llamacpp-${CONVERTED_MODEL_NAME}-embeddings.bin"
 else
     # Process piped JSON data and convert to binary (matching logits.cpp format)
     TEMP_FILE=$(mktemp /tmp/tmp.XXXXXX.binn)

examples/parallel/parallel.cpp

@@ -192,10 +192,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the target model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     auto * mem = llama_get_memory(ctx);
 

examples/retrieval/retrieval.cpp

@@ -149,10 +149,10 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);

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

@@ -34,10 +34,10 @@ int main(int argc, char ** argv) {
     std::string result2;
 
     // init
-    common_init_result llama_init = common_init_from_params(params);
+    auto llama_init = common_init_from_params(params);
 
-    llama_model * model = llama_init.model.get();
-    llama_context * ctx = llama_init.context.get();
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);

examples/speculative-simple/speculative-simple.cpp

@@ -40,10 +40,10 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    common_init_result llama_init_tgt = common_init_from_params(params);
+    auto llama_init_tgt = common_init_from_params(params);
 
-    model_tgt = llama_init_tgt.model.get();
-    ctx_tgt   = llama_init_tgt.context.get();
+    model_tgt = llama_init_tgt->model();
+    ctx_tgt   = llama_init_tgt->context();
 
     const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
 
@@ -61,10 +61,10 @@ int main(int argc, char ** argv) {
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
 
-    common_init_result llama_init_dft = common_init_from_params(params);
+    auto llama_init_dft = common_init_from_params(params);
 
-    //model_dft = llama_init_dft.model.get();
-    ctx_dft   = llama_init_dft.context.get();
+    //model_dft = llama_init_dft->model();
+    ctx_dft   = llama_init_dft->context();
 
     if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
         LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str());

examples/speculative/speculative.cpp

@@ -71,10 +71,10 @@ int main(int argc, char ** argv) {
     llama_context * ctx_dft = NULL;
 
     // load the target model
-    common_init_result llama_init_tgt = common_init_from_params(params);
+    auto llama_init_tgt = common_init_from_params(params);
 
-    model_tgt = llama_init_tgt.model.get();
-    ctx_tgt   = llama_init_tgt.context.get();
+    model_tgt = llama_init_tgt->model();
+    ctx_tgt   = llama_init_tgt->context();
 
     // load the draft model
     params.devices = params.speculative.devices;
@@ -87,10 +87,10 @@ int main(int argc, char ** argv) {
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
     params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
 
-    common_init_result llama_init_dft = common_init_from_params(params);
+    auto llama_init_dft = common_init_from_params(params);
 
-    model_dft = llama_init_dft.model.get();
-    ctx_dft   = llama_init_dft.context.get();
+    model_dft = llama_init_dft->model();
+    ctx_dft   = llama_init_dft->context();
 
     const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
     const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
@@ -242,7 +242,7 @@ int main(int argc, char ** argv) {
                 bool accept = false;
                 if (params.sampling.temp > 0) {
                     // stochastic verification
-                    common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
+                    common_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
 
                     auto & dist_tgt = *common_sampler_get_candidates(smpl, true);
 
@@ -491,7 +491,7 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
+                common_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft);
 
                 const auto * cur_p = common_sampler_get_candidates(drafts[s].smpl, true);
 

examples/training/finetune.cpp

@@ -39,9 +39,10 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
     // load the model and apply lora adapter, if any
-    common_init_result   llama_init = common_init_from_params(params);
-    llama_model_ptr    & model      = llama_init.model;
-    llama_context_ptr  & ctx        = llama_init.context;
+    auto llama_init = common_init_from_params(params);
+
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
@@ -54,8 +55,8 @@ int main(int argc, char ** argv) {
         LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
-    std::vector<llama_token> tokens  = common_tokenize(ctx.get(), params.prompt, true);
-    ggml_opt_dataset_t       dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get()) / 2);
+    std::vector<llama_token> tokens  = common_tokenize(ctx, params.prompt, true);
+    ggml_opt_dataset_t       dataset = common_opt_dataset_init(ctx, tokens, llama_n_ctx(ctx) / 2);
 
     struct lr_opt & lr = params.lr;
     LOG_INF("-optimizer %s -lr0 %.2g -wd %.2g -lr-min %.2g -min-epochs %.2g -epochs %d -period %.2g -val %.2g\n",
@@ -70,7 +71,7 @@ int main(int argc, char ** argv) {
         /*get_opt_pars_ud =*/&params.lr,
         /*optimizer_type  =*/params.optimizer,
     };
-    llama_opt_init(ctx.get(), model.get(), lopt_params);
+    llama_opt_init(ctx, model, lopt_params);
 
     const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - params.val_split);
 
@@ -78,7 +79,7 @@ int main(int argc, char ** argv) {
     ggml_opt_result_t result_eval  = ggml_opt_result_init();
 
     for (lr.epoch = 0; lr.epoch < lr.epochs; ++lr.epoch) {
-        llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
+        llama_opt_epoch(ctx, dataset, result_train, result_eval, idata_split,
                         ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
         fprintf(stderr, "\n");
 
@@ -88,7 +89,7 @@ int main(int argc, char ** argv) {
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_eval);
 
-    llama_model_save_to_file(model.get(), params.out_file.c_str());
+    llama_model_save_to_file(model, params.out_file.c_str());
 
     llama_backend_free();
 

ggml/CMakeLists.txt

@@ -54,6 +54,10 @@ if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     # TODO
 else()
     set(GGML_STANDALONE OFF)
+
+    if (NOT CMAKE_RUNTIME_OUTPUT_DIRECTORY)
+        set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+    endif()
 endif()
 
 if (EMSCRIPTEN)

ggml/include/ggml-alloc.h

@@ -53,7 +53,14 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // call with a worst-case graph to avoid buffer reallocations
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
+// ggml_gallocr_resrve_n_size writes the buffer sizes per galloc buffer that would be allocated by ggml_gallocr_reserve_n to sizes
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API void ggml_gallocr_reserve_n_size(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids,
+    size_t * sizes);
 GGML_API bool ggml_gallocr_reserve_n(
     ggml_gallocr_t galloc,
     struct ggml_cgraph * graph,
@@ -68,6 +75,8 @@ GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_i
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
+// ggml_backend_alloc_ctx_tensors_from_buft_size returns the size of the buffer that would be allocated by ggml_backend_alloc_ctx_tensors_from_buft
+GGML_API size_t                       ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
 GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
 

ggml/include/ggml-backend.h

@@ -307,6 +307,7 @@ extern "C" {
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph
+    GGML_API void                 ggml_backend_sched_reserve_size(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph, size_t * sizes);
     GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // returns success
 
     GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);

ggml/include/ggml.h

@@ -2615,7 +2615,8 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
+    GGML_API void ggml_log_get(ggml_log_callback * log_callback, void ** user_data);
+    GGML_API void ggml_log_set(ggml_log_callback   log_callback, void *  user_data);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 

ggml/src/ggml-alloc.c

@@ -594,7 +594,9 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
 }
 
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
+    return t->data != NULL // tensor data already set externally
+        || t->buffer // tensor on external buffer (but not yet allocated)
+        || ggml_gallocr_is_own(galloc, t); // tensor will be allocated by galloc
 }
 
 // free the extra space at the end if the new tensor is smaller
@@ -823,7 +825,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     }
 }
 
-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+static bool ggml_gallocr_reserve_n_impl(
+        ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids, bool no_alloc) {
     size_t min_hash_size = graph->n_nodes + graph->n_leafs;
     // add 25% margin to avoid hash collisions
     min_hash_size += min_hash_size / 4;
@@ -928,16 +931,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
                 size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
                 if (cur_size > 0) {
                     GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n",
-                        __func__, ggml_backend_buft_name(galloc->bufts[i]),
-                        cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+                        __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
                 }
             }
 #endif
             ggml_vbuffer_free(galloc->buffers[i]);
-            galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
-            if (galloc->buffers[i] == NULL) {
-                GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
-                return false;
+            if (no_alloc) {
+                galloc->buffers[i] = NULL;
+            } else {
+                galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+                if (galloc->buffers[i] == NULL) {
+                    GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                    return false;
+                }
             }
         }
     }
@@ -945,6 +951,21 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     return true;
 }
 
+void ggml_gallocr_reserve_n_size(
+        ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids, size_t * sizes) {
+    GGML_ASSERT(ggml_gallocr_reserve_n_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids, /*no_alloc =*/ true));
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        sizes[i] = 0;
+        for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) {
+            sizes[i] += galloc->buf_tallocs[i]->chunks[c]->max_size;
+        }
+    }
+}
+
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    return ggml_gallocr_reserve_n_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids, /*no_alloc =*/ false);
+}
+
 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }
@@ -1147,7 +1168,8 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
     return true;
 }
 
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+static ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft_impl(
+        struct ggml_context * ctx, ggml_backend_buffer_type_t buft, size_t * nbytes_total, bool no_alloc) {
     GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
 
     size_t alignment = ggml_backend_buft_get_alignment(buft);
@@ -1155,6 +1177,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     ggml_backend_buffer_t * buffers = NULL;
     size_t n_buffers = 0;
+    *nbytes_total = 0;
 
     size_t cur_buf_size = 0;
     struct ggml_tensor * first = ggml_get_first_tensor(ctx);
@@ -1166,10 +1189,11 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
         if (cur_buf_size > 0 && (cur_buf_size + this_size) > max_size) {
             // allocate tensors in the current buffer
-            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+            if (!no_alloc && !alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
                 return NULL;
             }
             first = t;
+            *nbytes_total += cur_buf_size;
             cur_buf_size = this_size;
         } else {
             cur_buf_size += this_size;
@@ -1178,15 +1202,21 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
 
     // allocate remaining tensors
     if (cur_buf_size > 0) {
-        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+        *nbytes_total += cur_buf_size;
+        if (!no_alloc && !alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
             return NULL;
         }
     }
 
+    if (no_alloc) {
+        return NULL;
+    }
+
     if (n_buffers == 0) {
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
 #endif
+        GGML_ASSERT(!buffers);
         return NULL;
     }
 
@@ -1196,10 +1226,24 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     } else {
         buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
     }
-    free(buffers);
+    if (buffers) {
+        free(buffers); // can be NULL if context is empty or no_alloc
+    }
     return buffer;
 }
 
+size_t ggml_backend_alloc_ctx_tensors_from_buft_size(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    size_t nbytes_total = 0;
+    ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc=*/ true);
+    GGML_ASSERT(!buf);
+    return nbytes_total;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    size_t nbytes_total = 0;
+    return ggml_backend_alloc_ctx_tensors_from_buft_impl(ctx, buft, &nbytes_total, /*no_alloc =*/ false);
+}
+
 ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
     return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
 }

ggml/src/ggml-backend.cpp

@@ -36,12 +36,11 @@ const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
 }
 
 ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    GGML_ASSERT(buft);
     if (size == 0) {
         // return a dummy buffer for zero-sized allocations
         return ggml_backend_buffer_init(buft, {}, NULL, 0);
     }
-
-    GGML_ASSERT(buft);
     return buft->iface.alloc_buffer(buft, size);
 }
 
@@ -128,6 +127,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
         return NULL;
     }
 
+    // FIXME JG: a multi_buffer has a non-zero size, according to the above comment get_base is not optional,
+    //     I don't know whether the above comment is correct
+    if (!buffer->iface.get_base) {
+        return NULL;
+    }
+
     void * base = buffer->iface.get_base(buffer);
 
     GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
@@ -1727,6 +1732,20 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     sched->is_alloc = false;
 }
 
+void ggml_backend_sched_reserve_size(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph, size_t * sizes) {
+    GGML_ASSERT(sched);
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
+    GGML_ASSERT(sizes);
+
+    ggml_backend_sched_reset(sched);
+
+    ggml_backend_sched_synchronize(sched);
+
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
+    ggml_gallocr_reserve_n_size(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids, sizes);
+}
+
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
     GGML_ASSERT(sched);
     GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -24,6 +24,7 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__aarch64__) && defined(__ARM_NEON) && (defined(__ARM_FEATURE_MATMUL_INT8) || defined(__ARM_FEATURE_DOTPROD))
 static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
                                              int16x8_t *     out_mins,
                                              int8_t *        out_scales) {
@@ -46,6 +47,7 @@ static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
     scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
     memcpy(out_scales, scales_u32, 8);
 }
+#endif
 
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1976,9 +1976,6 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
             break;
 
         case GGML_TYPE_F16:
-            if (!opt_experimental) {
-                return false;
-            }
             break;
 
         default:

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -903,7 +903,7 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
         const float * restrict vy  = (const float * restrict) y;
 
         for (uint32_t i = 0; i < n; i++) {
-            rsum += vx[i] * (__fp16) vy[i];
+            rsum += (float)vx[i] * vy[i];
         }
         *s = rsum;
         return;
@@ -917,7 +917,7 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
 
     // for some reason we need volatile here so that the compiler doesn't try anything funky
     volatile HVX_Vector rsum = Q6_V_vsplat_R(0);
-
+    float r_sum_scalar = 0.0f;
     uint32_t i = 0;
 
     for (i = 0; i < nv0; i++) {
@@ -926,31 +926,42 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
         HVX_Vector     x  = vx[i];
         HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), Q6_Vh_vsplat_R(0x3C00));  // mul by 1.0
 
-        HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_hi_W(xp)), Q6_V_hi_W(yp));
-        HVX_Vector lo = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
+        //NOTE: need volatile here to prevent compiler optimization
+        // Seem compiler cannot guarantee read-after-write??
+        volatile HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_hi_W(xp)), Q6_V_hi_W(yp));
+        volatile HVX_Vector lo = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
 
         HVX_Vector sum = Q6_Vqf32_vadd_Vqf32Vqf32(hi, lo);
         rsum           = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum);
     }
 
     if (nv1) {
-        HVX_VectorPair yp = vy[i];
+        // HVX_VectorPair yp = vy[i];
 
-        HVX_Vector     x  = vx[i];
-        HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), Q6_Vh_vsplat_R(0x3C00));  // mul by 1.0
+        // HVX_Vector     x  = vx[i];
+        // HVX_VectorPair xp = Q6_Wqf32_vmpy_VhfVhf(Q6_Vh_vshuff_Vh(x), Q6_Vh_vsplat_R(0x3C00));  // mul by 1.0
 
-        if (nv1 >= 32) {
-            HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_hi_W(xp)), Q6_V_hi_W(yp));
-            rsum          = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, hi);
-            nv1 -= 32;
-        }
+        // if (nv1 >= 32) {
+        //     volatile HVX_Vector hi = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_hi_W(xp)), Q6_V_hi_W(yp));
+        //     rsum          = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, hi);
+        //     nv1 -= 32;
+        // }
 
+        // rsum = hvx_vec_qf32_reduce_sum(rsum);
+
+        // if (nv1) {
+        //     volatile HVX_Vector lo  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
+        //     HVX_Vector sum = hvx_vec_qf32_reduce_sum_n(lo, nv1);
+        //     rsum           = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum);
+        // }
+
+        //process the remainder using scalar loop
         rsum = hvx_vec_qf32_reduce_sum(rsum);
+        const __fp16 * restrict sx = (const __fp16 * restrict) x;
+        const float * restrict sy  = (const float * restrict) y;
 
-        if (nv1) {
-            HVX_Vector lo  = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(Q6_V_lo_W(xp)), Q6_V_lo_W(yp));
-            HVX_Vector sum = hvx_vec_qf32_reduce_sum_n(lo, nv1);
-            rsum           = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, sum);
+        for (uint32_t i = nv0 * 64; i < n; i++) {
+            r_sum_scalar += (float) sx[i] * sy[i];
         }
 
         // hvx_vec_dump_fp16("X", x);
@@ -961,7 +972,7 @@ static void vec_dot_f16_f32(const int n, float * restrict s, const void * restri
         rsum = hvx_vec_qf32_reduce_sum(rsum);
     }
 
-    *s = hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(rsum));
+    *s = hvx_vec_get_fp32(Q6_Vsf_equals_Vqf32(rsum)) + r_sum_scalar;
 
 #    ifdef HTP_DEBUG
     {
@@ -1498,9 +1509,6 @@ static void matmul_f16_f32(struct htp_tensor * restrict src0,
     uint64_t t1, t2;
     t1 = HAP_perf_get_qtimer_count();
 
-    const size_t src0_row_size = sizeof(__fp16) * ne00;
-    const size_t src1_row_size = sizeof(float) * ne10;
-
     assert(ne12 % ne02 == 0);
     assert(ne13 % ne03 == 0);
 
@@ -1510,8 +1518,6 @@ static void matmul_f16_f32(struct htp_tensor * restrict src0,
     // This is the size of the rest of the dimensions of the result
     const uint32_t nr1 = ne1 * ne2 * ne3;
 
-    uint32_t chunk_size = 64;
-
     // distribute the thread work across the inner or outer loop based on which one is larger
     uint32_t nchunk0 = nr0 > nr1 ? nth : 1;  // parallelize by src0 rows
     uint32_t nchunk1 = nr0 > nr1 ? 1 : nth;  // parallelize by src1 rows
@@ -1544,11 +1550,11 @@ static void matmul_f16_f32(struct htp_tensor * restrict src0,
     const uint32_t blck_0 = 64;
     const uint32_t blck_1 = 64;
 
-    float tmp[32];
+    __attribute__((aligned(128))) float tmp[64];
 
     for (uint32_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) {
         for (uint32_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) {
-            for (uint32_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ir1++) {
+            for (uint32_t ir1 = iir1; ir1 < MIN(iir1 + blck_1, ir1_end); ir1++) {
                 const uint32_t i13 = (ir1 / (ne12 * ne1));
                 const uint32_t i12 = (ir1 - i13 * ne12 * ne1) / ne1;
                 const uint32_t i11 = (ir1 - i13 * ne12 * ne1 - i12 * ne1);
@@ -1561,13 +1567,16 @@ static void matmul_f16_f32(struct htp_tensor * restrict src0,
                 const uint32_t i2 = i12;
                 const uint32_t i3 = i13;
 
-                const uint8_t * restrict src0_row = (const uint8_t *) src0->data + (0 + i02 * nb02 + i03 * nb03);
+                const uint8_t * restrict src0_base = (const uint8_t *) src0->data + (0 + i02 * nb02 + i03 * nb03);
                 const uint8_t * restrict src1_col =
-                    (const uint8_t *) src1->data + (i11 + i12 * ne11 + i13 * ne12 * ne11) * src1_row_size;
+                    (const uint8_t *) src1->data + (i11 * nb11 + i12 * nb12 + i13 * nb13);
                 float * dst_col = (float *) ((uint8_t * restrict) dst->data + (i1 * nb1 + i2 * nb2 + i3 * nb3));
 
-                for (uint32_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ir0++) {
-                    vec_dot_f16_f32(ne00, &tmp[ir0 - iir0], src0_row + ir0 * src0_row_size, src1_col);
+                const uint32_t ir0_block_end = MIN(iir0 + blck_0, ir0_end);
+                for (uint32_t ir0 = iir0; ir0 < ir0_block_end; ir0++) {
+                    // Use nb01 stride for non-contiguous src0 support
+                    const uint8_t * restrict src0_row = src0_base + ir0 * nb01;
+                    vec_dot_f16_f32(ne00, &tmp[ir0 - iir0], src0_row, src1_col);
                 }
 
                 hvx_copy_fp32_ua((uint8_t *) &dst_col[iir0], (uint8_t *) tmp, MIN(iir0 + blck_0, ir0_end) - iir0);

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

@@ -769,9 +769,16 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 #endif
 
             dev->props.use_shared_buffers = dev->props.has_unified_memory;
+#if TARGET_OS_OSX
+            // In case of eGPU, shared memory may be preferable.
+            dev->props.use_shared_buffers |= [dev->mtl_device location] == MTLDeviceLocationExternal;
+#endif
             if (getenv("GGML_METAL_SHARED_BUFFERS_DISABLE") != NULL) {
                 dev->props.use_shared_buffers = false;
             }
+            if (getenv("GGML_METAL_SHARED_BUFFERS_ENABLE") != NULL) {
+                dev->props.use_shared_buffers = true;
+            }
 
             dev->props.supports_gpu_family_apple7 = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
 

ggml/src/ggml-sycl/add-id.cpp

@@ -0,0 +1,77 @@
+#include <sycl/sycl.hpp>
+#include "common.hpp"
+#include "add-id.hpp"
+
+static void add_id_kernel(
+    const float* src0,
+    const float* src1,
+    const int32_t* src2,
+    float* dst,
+    int64_t ne0,
+    int64_t ne1,
+    size_t nb01,
+    size_t nb02,
+    size_t nb11,
+    size_t nb21,
+    sycl::nd_item<3> item_ct1) {
+  const int64_t i1 = item_ct1.get_group(2);
+  const int64_t i2 = item_ct1.get_group(1);
+
+  const int i11 =
+      *(const int32_t*)((const char*)src2 + i1 * sizeof(int32_t) + i2 * nb21);
+
+  const size_t nb1 = ne0 * sizeof(float);
+  const size_t nb2 = ne1 * nb1;
+
+  float* dst_row = (float*)((char*)dst + i1 * nb1 + i2 * nb2);
+  const float* src0_row =
+      (const float*)((const char*)src0 + i1 * nb01 + i2 * nb02);
+  const float* src1_row = (const float*)((const char*)src1 + i11 * nb11);
+
+  for (int64_t i0 = item_ct1.get_local_id(2); i0 < ne0;
+       i0 += item_ct1.get_local_range(2)) {
+    dst_row[i0] = src0_row[i0] + src1_row[i0];
+  }
+}
+
+void ggml_sycl_add_id(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
+  const ggml_tensor* src0 = dst->src[0];
+  const ggml_tensor* src1 = dst->src[1];
+  const ggml_tensor* src2 = dst->src[2];
+
+  GGML_TENSOR_TERNARY_OP_LOCALS
+
+  GGML_ASSERT(dst->type == GGML_TYPE_F32);
+  GGML_ASSERT(src0->type == GGML_TYPE_F32);
+  GGML_ASSERT(src1->type == GGML_TYPE_F32);
+  GGML_ASSERT(src2->type == GGML_TYPE_I32);
+
+  GGML_ASSERT(nb00 == sizeof(float));
+  GGML_ASSERT(nb10 == sizeof(float));
+  GGML_ASSERT(nb20 == sizeof(int32_t));
+
+  const float* src0_d = (const float*)src0->data;
+  const float* src1_d = (const float*)src1->data;
+  const int32_t* src2_d = (const int32_t*)src2->data;
+  float* dst_d = (float*)dst->data;
+
+  int threads = std::min((int)ne00, 768);  // cols
+  ctx.stream()->parallel_for(
+      sycl::nd_range<3>(
+          sycl::range<3>(1, ne02, ne01) * sycl::range<3>(1, 1, threads),
+          sycl::range<3>(1, 1, threads)),
+      [=](sycl::nd_item<3> item_ct1) {
+        add_id_kernel(
+            src0_d,
+            src1_d,
+            src2_d,
+            dst_d,
+            ne0,
+            ne1,
+            nb01,
+            nb02,
+            nb11,
+            nb21,
+            item_ct1);
+      });
+}

ggml/src/ggml-sycl/add-id.hpp

@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_ADD_ID_HPP
+#define GGML_SYCL_ADD_ID_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_add_id(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_ADD_ID_HPP

ggml/src/ggml-sycl/common.hpp

@@ -642,5 +642,22 @@ static __dpct_inline__ sycl::uint2 fast_div_modulo(uint32_t n, const sycl::uint3
     return sycl::uint2(div_val, mod_val);
 }
 
+static __dpct_inline__ int ggml_sycl_dp4a(const int a, const int b, int c) {
+    return dpct::dp4a(a, b, c);
+}
+
+static __dpct_inline__ float ggml_sycl_e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    float result;
+    memcpy(&result, &bits, sizeof(float));
+    return result;
+}
+
 
 #endif // GGML_SYCL_COMMON_HPP

ggml/src/ggml-sycl/convert.cpp

@@ -472,6 +472,16 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k
       }
 }
 
+template <typename dst_t>
+static void dequantize_row_mxfp4_sycl(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+        [=](sycl::nd_item<3> item_ct1) {
+            dequantize_block_mxfp4(vx, y, item_ct1);
+        });
+}
+
 template <typename src_t, typename dst_t>
 static void convert_unary_nc(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01,
                           const int64_t ne02, const int64_t s01, const int64_t s02, const int64_t s03,
@@ -518,6 +528,7 @@ static void convert_unary_sycl(const void * vx, dst_t * y, const int64_t k, dpct
     convert_unary_nc_sycl<src_t>(vx, y, k, 1, 1, 1, k, k, k, queue);
 }
 
+
 to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
     switch (type) {
         case GGML_TYPE_Q4_0:
@@ -571,6 +582,8 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
             return dequantize_row_iq4_xs_sycl;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_sycl;
         case GGML_TYPE_F32:
             return convert_unary_sycl<float>;
 #ifdef GGML_SYCL_HAS_BF16
@@ -636,6 +649,8 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
             return dequantize_row_iq4_xs_sycl;
         case GGML_TYPE_IQ4_NL:
             return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_sycl;
         case GGML_TYPE_F16:
             return convert_unary_sycl<sycl::half>;
 #ifdef GGML_SYCL_HAS_BF16

ggml/src/ggml-sycl/dequantize.hpp

@@ -819,5 +819,23 @@ dequantize_block_iq4_xs(const void *__restrict__ vx, dst_t *__restrict__ yy,
     }
 }
 
+template<typename dst_t>
+static void dequantize_block_mxfp4(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                   const sycl::nd_item<3> &item_ct1) {
+    // auto                item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t       i        = item_ct1.get_group(2);
+    const block_mxfp4 * x = (const block_mxfp4 *) vx + i*(QK_K/QK_MXFP4);
+
+    const int64_t    tid = item_ct1.get_local_id(2);
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = ggml_sycl_e8m0_to_fp32(x[ib].e);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_mxfp4[q4[j] & 0xf]*0.5f;
+        y[j+16] = d * kvalues_mxfp4[q4[j] >>  4]*0.5f;
+    }
+}
 
 #endif // GGML_SYCL_DEQUANTIZE_HPP

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -1860,10 +1860,31 @@ namespace dpct
                                            : id);
     }
 
+    template <typename T1, typename T2>
+    using dot_product_acc_t = std::conditional_t<
+        std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,
+        uint32_t,
+        int32_t>;
+
+    template <typename T>
+    sycl::vec<T, 4> extract_and_sign_or_zero_extend4(T val) {
+      return sycl::vec<T, 1>(val)
+          .template as<sycl::vec<
+              std::conditional_t<std::is_signed_v<T>, int8_t, uint8_t>,
+              4>>()
+          .template convert<T>();
+    }
+
     template <typename T1, typename T2, typename T3>
-    inline auto dp4a(T1 a, T2 b, T3 c)
-    {
-        return syclcompat::dp4a(a, b, c);
+    inline auto dp4a(T1 a, T2 b, T3 c) {
+      dot_product_acc_t<T1, T2> res = c;
+      auto va = extract_and_sign_or_zero_extend4(a);
+      auto vb = extract_and_sign_or_zero_extend4(b);
+      res += va[0] * vb[0];
+      res += va[1] * vb[1];
+      res += va[2] * vb[2];
+      res += va[3] * vb[3];
+      return res;
     }
 
     struct sub_sat
@@ -2972,6 +2993,38 @@ namespace dpct
     atomic_fetch_add<T1, addressSpace>(addr, operand, memoryOrder);
     }
 
+    inline unsigned int byte_level_permute(
+        unsigned int a, unsigned int b, unsigned int s) {
+      unsigned int ret;
+      ret = ((((std::uint64_t)b << 32 | a) >> (s & 0x7) * 8) & 0xff) |
+            (((((std::uint64_t)b << 32 | a) >> ((s >> 4) & 0x7) * 8) & 0xff)
+             << 8) |
+            (((((std::uint64_t)b << 32 | a) >> ((s >> 8) & 0x7) * 8) & 0xff)
+             << 16) |
+            (((((std::uint64_t)b << 32 | a) >> ((s >> 12) & 0x7) * 8) & 0xff)
+             << 24);
+      return ret;
+    }
+
+    inline uint32_t byte_level_permute_custom(
+        uint32_t low32, uint32_t high32, uint32_t sel, int mode = 0) {
+      constexpr uint16_t lookup[6][4] = {
+          {0x3210, 0x4321, 0x5432, 0x6543},  // Forward 4-byte extract
+          {0x5670, 0x6701, 0x7012, 0x0123},  // Backward 4-byte extract
+          {0x0000, 0x1111, 0x2222, 0x3333},  // Replicate 8-bit values
+          {0x3210, 0x3211, 0x3222, 0x3333},  // Edge clamp left
+          {0x0000, 0x1110, 0x2210, 0x3210},  // Edge clamp right
+          {0x1010, 0x3232, 0x1010, 0x3232}   // Replicate 16-bit values
+      };
+
+      if (mode >= 1 && mode <= 6) {
+        return byte_level_permute(low32, high32, lookup[mode - 1][sel & 0x3]);
+      } else if (!mode) {
+        return byte_level_permute(low32, high32, sel);
+      }
+      return 0;
+    }
+
 } // COPY from DPCT head files
 
 #endif // GGML_SYCL_DPCT_HELPER_HPP

ggml/src/ggml-sycl/element_wise.cpp

@@ -911,6 +911,98 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
         });
 }
 
+__dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {
+    x = sycl::fmin(x, limit);
+    g = sycl::fmax(sycl::fmin(g, limit), -limit);
+
+    float out_glu = x / (1.0f + sycl::native::exp(-x * alpha));
+    out_glu = out_glu * (1.0f + g);
+    return out_glu;
+}
+
+
+template <typename T>
+static void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k,
+                              const int64_t n, const int64_t o0, const int64_t o1,
+                              float alpha, float limit, sycl::nd_item<3> item_ct1) {
+    const int64_t i = int64_t(item_ct1.get_local_range(2)) * item_ct1.get_group(2) + item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+
+    const int64_t j0 = (i / n) * o0 + (i % n);
+    const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+
+    float xi = x[j0];
+    float gi = g[j1];
+
+    dst[i] = ggml_sycl_op_swiglu_oai_single(xi, gi, alpha, limit);
+}
+
+template <typename T>
+static void swiglu_oai_sycl(const T *       x,
+                            const T *       g,
+                            T *             dst,
+                            const int64_t   k,
+                            const int64_t   n,
+                            const int64_t   o0,
+                            const int64_t   o1,
+                            const float     alpha,
+                            const float     limit,
+                            dpct::queue_ptr stream) {
+    const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
+                                           sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             swiglu_oai_kernel(x, g, dst, k, n, o0, o1, alpha, limit, item_ct1);
+                         });
+}
+
+void ggml_sycl_op_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    dpct::queue_ptr     stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    //const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
+    float * src0_p = (float *) src0_d;
+    float * src1_p = (float *) src1_d;
+
+    if (!src1) {
+        src0_p += swapped ? nc : 0;
+        src1_p += swapped ? 0 : nc;
+    }
+
+    swiglu_oai_sycl(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), alpha, limit, stream);
+}
+
 static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
         [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
@@ -1070,6 +1162,11 @@ void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_swiglu(ctx, dst);
 }
 
+void ggml_sycl_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_swiglu_oai(ctx, dst);
+}
+
 void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_geglu_erf(ctx, dst);

ggml/src/ggml-sycl/element_wise.hpp

@@ -5,6 +5,8 @@
 #include "ggml.h"
 #include <limits> // For std::numeric_limits
 
+#define SYCL_GLU_BLOCK_SIZE 256
+
 template <typename T>
 T neg_infinity() {
     return -std::numeric_limits<T>::infinity();
@@ -41,6 +43,8 @@ void ggml_sycl_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
+void ggml_sycl_swiglu_oai(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 void ggml_sycl_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

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

@@ -39,6 +39,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 
+#include "ggml-sycl/add-id.hpp"
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
@@ -3313,6 +3314,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
+
     // mmvq and mmq need the __dp4a instruction which is available for gen12+
     // Workaround in https://github.com/ggerganov/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e
     use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS);
@@ -3320,7 +3322,6 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // SYCL_USE_XMX
 
-
     // mmvq path is faster in the CUDA backend.
     if (!g_ggml_sycl_prioritize_dmmv && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda
         // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization
@@ -3711,6 +3712,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_ADD1: // TODO: more efficient implementation
             ggml_sycl_add(ctx, dst);
             break;
+        case GGML_OP_ADD_ID:
+            ggml_sycl_add_id(ctx, dst);
+            break;
         case GGML_OP_SUB:
             ggml_sycl_sub(ctx, dst);
             break;
@@ -3803,6 +3807,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_sycl_swiglu(ctx, dst);
                     break;
+                case GGML_GLU_OP_SWIGLU_OAI:
+                    ggml_sycl_swiglu_oai(ctx, dst);
+                    break;
                 case GGML_GLU_OP_GEGLU_ERF:
                     ggml_sycl_geglu_erf(ctx, dst);
                     break;
@@ -4397,6 +4404,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]);
@@ -4424,15 +4432,18 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     }
                 }
                 ggml_type src0_type = op->src[0]->type;
-                if (src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_MXFP4) {
-                    // TODO: support MXFP4
+                if (src0_type == GGML_TYPE_BF16 ) {
+                    // TODO: support GGML_TYPE_BF16
                     // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
+
                 // TODO: The configuration below needs more work to be supported with oneDNN
-                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) && a->ne[2] > 1 && a->ne[3] > 1) {
-                    return false;
+                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) &&
+                    a->ne[2] > 1 && a->ne[3] > 1 && src0_type == GGML_TYPE_F16) {
+                  return false;
                 }
+
                 // TODO: This specific configuration can fail with oneDNN and needs more debugging
                 if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
                     a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
@@ -4553,9 +4564,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
-            return true;
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
+        case GGML_OP_ADD_ID:
         case GGML_OP_SUB:
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_MUL:

ggml/src/ggml-sycl/mmvq.cpp

@@ -595,6 +595,25 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
+static void mul_mat_vec_mxfp4_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols, const int nrows,
+                                        dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_MXFP4 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+
+    {
+        stream->submit([&](sycl::handler & cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 mul_mat_vec_q<QK_MXFP4, QI_MXFP4, block_mxfp4, VDR_MXFP4_Q8_1_MMVQ, vec_dot_mxfp4_q8_1>(
+                                     vx, vy, dst, ncols, nrows, item_ct1);
+                             });
+        });
+    }
+}
+
+
 static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -1123,6 +1142,9 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
             case GGML_TYPE_IQ4_XS:
                 mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
+            case GGML_TYPE_MXFP4:
+                mul_mat_vec_mxfp4_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                break;
             default:
                 GGML_ABORT("fatal error");
         }

ggml/src/ggml-sycl/pad.cpp

@@ -14,10 +14,10 @@
 #include "pad.hpp"
 
 static void pad_f32(const float * src, float * dst,
-                               const int lp0, const int rp0, const int lp1, const int rp1,
-                               const int lp2, const int rp2, const int lp3, const int rp3,
-                               const int ne0, const int ne1, const int ne2, const int ne3) {
-    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+                    const int lp0, const int rp0, const int lp1, const int rp1,
+                    const int lp2, const int rp2, const int lp3, const int rp3,
+                    const int ne0, const int ne1, const int ne2, const int ne3,
+                    sycl::nd_item<3> item_ct1) {
     int i0 = item_ct1.get_local_id(2) +
              item_ct1.get_group(2) * item_ct1.get_local_range(2);
     int i1 = item_ct1.get_group(1);
@@ -63,7 +63,7 @@ static void pad_f32_sycl(const float *src, float *dst, const int lp0,
                           sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
         [=](sycl::nd_item<3> item_ct1) {
             pad_f32(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1,
-                    ne2, ne3);
+                    ne2, ne3, item_ct1);
         });
 }
 

ggml/src/ggml-sycl/ssm_conv.cpp

@@ -88,7 +88,7 @@ void ggml_sycl_ssm_conv(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
 
-    GGML_ASSERT(src0->nb[1] == src0->ne[0] * static_cast<int>(sizeof(float)));
+    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
 
     const int src_stride_inner = ncs;
     const int src_stride_seq   = ncs * d_inner;

ggml/src/ggml-sycl/vecdotq.hpp

@@ -20,6 +20,18 @@
 typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1,
                                   const int & iqs);
 
+static __dpct_inline__ int get_int_b1(const void * x, const int & i32) {
+    const uint8_t * x8 = (const uint8_t *) x;
+
+    int x32  = x8[4*i32 + 0] <<  0;
+    x32     |= x8[4*i32 + 1] <<  8;
+    x32     |= x8[4*i32 + 2] << 16;
+    x32     |= x8[4*i32 + 3] << 24;
+
+    return x32;
+}
+
+
 static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) {
   const uint16_t* x16 =
       (const uint16_t*)(x8 + sizeof(int) * i32); // assume at least 2 byte
@@ -75,6 +87,28 @@ static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4,
     val2 = v1 | (v2 << 16);
 }
 
+static __dpct_inline__ sycl::int2 get_int_from_table_16(
+    const int& q4, const int8_t* table) {
+  const uint32_t* table32 = (const uint32_t*)table;
+  uint32_t tmp[2];
+  const uint32_t low_high_selection_indices =
+      (0x32103210 | ((q4 & 0x88888888) >> 1));
+#pragma unroll
+  for (uint32_t i = 0; i < 2; ++i) {
+    const uint32_t shift = 16 * i;
+
+    const uint32_t low =
+        dpct::byte_level_permute(table32[0], table32[1], q4 >> shift);
+    const uint32_t high =
+        dpct::byte_level_permute(table32[2], table32[3], q4 >> shift);
+    tmp[i] = dpct::byte_level_permute(
+        low, high, low_high_selection_indices >> shift);
+  }
+  return sycl::int2(
+      dpct::byte_level_permute(tmp[0], tmp[1], 0x6420),
+      dpct::byte_level_permute(tmp[0], tmp[1], 0x7531));
+}
+
 #define VDR_Q2_K_Q8_1_MMVQ 1
 
 // contiguous v/x values
@@ -685,6 +719,30 @@ vec_dot_q4_1_q8_1(const void *__restrict__ vbq,
     return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
 }
 
+#define VDR_MXFP4_Q8_1_MMVQ 2
+#define VDR_MXFP4_Q8_1_MMQ  4
+
+static __dpct_inline__ float vec_dot_mxfp4_q8_1(const void * __restrict__ vbq,
+                                                const block_q8_1 * __restrict__ bq8_1,
+                                                const int & iqs) {
+    const block_mxfp4 * bq4 = (const block_mxfp4 *) vbq;
+
+    const int * q8 = (const int *) bq8_1->qs + iqs;
+
+    int sumi = 0;
+#pragma unroll
+    for (int l = 0; l < VDR_MXFP4_Q8_1_MMVQ; ++l) {
+        const int aux_q4 = get_int_b1(bq4->qs, iqs + l);
+        const sycl::int2 v      = get_int_from_table_16(aux_q4, kvalues_mxfp4);
+        sumi = ggml_sycl_dp4a(v.x(), q8[l + 0], sumi);
+        sumi = ggml_sycl_dp4a(v.y(), q8[l + 4], sumi);
+    }
+
+    const float d = ggml_sycl_e8m0_to_fp32(bq4->e) * 0.5f * (bq8_1->ds)[0];
+    return d * sumi;
+}
+
+
 static __dpct_inline__ float
 vec_dot_q5_0_q8_1(const void *__restrict__ vbq,
                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {

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

@@ -256,6 +256,9 @@ void main() {
         barrier();
     }
 
+    // prevent race on tmpsh
+    barrier();
+
     // reduce across threads
 
     [[unroll]] for (uint32_t r = 0; r < Br; ++r) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -302,6 +302,9 @@ void main() {
         barrier();
     }
 
+    // prevent race on tmpsh
+    barrier();
+
     // reduce across threads
 
     float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -7,34 +7,50 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
-    const uint y_idx = i * QUANT_K + 32 * ib32;
-
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const float d = float(data_a[ibi].d);
-        const uint qh = data_a[ibi].qh[ib32];
-        const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
-        const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
-
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i,
+                     const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx_base = i * QUANT_K + 32 * ib32;
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        const uint base_b_idx = (j * p.batch_stride_b + b_offset + y_idx_base) / 4;
         [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const uint qs = data_a[ibi].qs[4 * ib32 + l];
-            const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
-            const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+            const vec4 b_val_0 = vec4(data_b_v4[base_b_idx + 2 * l]);
+            const vec4 b_val_1 = vec4(data_b_v4[base_b_idx + 2 * l + 1]);
 
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
-                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+            // index for data_a
+            uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                const float d = float(data_a[ibi].d);
+                const uint qh = data_a[ibi].qh[ib32];
+
+                const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+                const uint qs = data_a[ibi].qs[4 * ib32 + l];
+                const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
+                const uint16_t grid = uint16_t(iq1s_grid[qs | (idxhi << 8)]);
+
+                const float delta_val = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+                const vec4 delta_v = vec4(delta_val);
+                const vec4 fbits0 = vec4(
+                    float(bitfieldExtract(grid, 0, 2)),
+                    float(bitfieldExtract(grid, 2, 2)),
+                    float(bitfieldExtract(grid, 4, 2)),
+                    float(bitfieldExtract(grid, 6, 2))
+                );
+                const vec4 fbits1 = vec4(
+                    float(bitfieldExtract(grid, 8, 2)),
+                    float(bitfieldExtract(grid, 10, 2)),
+                    float(bitfieldExtract(grid, 12, 2)),
+                    float(bitfieldExtract(grid, 14, 2))
+                );
+
+                vec4 sum_v = fma(b_val_0, fbits0 + delta_v, vec4(0.0));
+                sum_v      = fma(b_val_1, fbits1 + delta_v, sum_v);
+                FLOAT_TYPE sum = dot(sum_v, vec4(1.0));
 
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int k = 0; k < 4; ++k) {
-                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
-                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
-                }
                 temp[j][n] = fma(dl, sum, temp[j][n]);
+                ibi += num_blocks_per_row;
             }
         }
-        ibi += num_blocks_per_row;
     }
 }
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -244,17 +244,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint iqs = idx % 128;                 // 0..127
 
             const uint n = iqs / 64;                    // 0,1
-            const uint b = (iqs % 64) / 32;             // 0,1
+            const uint b = ((iqs % 64) / 32) * 4;       // 0,4
             const uint is_b = (iqs % 16) / 8;           // 0,1
             const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
             const uint is = 8 * n + qhshift + is_b;     // 0..15
-            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
-            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+            const uint qsi = n * 32 + (iqs % 32);       // 0..63
+            const uint qhi = n * 16 + (iqs % 16);       // 0..31
 
             const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32),
-                                             dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+            const uint ql = (uint(data_a_packed16[ib].ql[qsi]) >> b) & 0x0F0F;
+            const uint qh = (uint(data_a_packed16[ib].qh[qhi]) >> qhshift) & 0x0303;
+            const vec2 q = (vec2(unpack8(ql | (qh << 4)).xy) - 32) * dscale;
+
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(q.x, q.y);
 #elif defined(DATA_A_IQ1_S)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

ggml/src/ggml.c

@@ -7566,6 +7566,11 @@ size_t ggml_quantize_chunk(
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void ggml_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    *log_callback = g_logger_state.log_callback;
+    *user_data    = g_logger_state.log_callback_user_data;
+}
+
 void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
     g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
     g_logger_state.log_callback_user_data = user_data;

gguf-py/gguf/constants.py

@@ -413,6 +413,7 @@ class MODEL_ARCH(IntEnum):
     JAIS             = auto()
     NEMOTRON         = auto()
     NEMOTRON_H       = auto()
+    NEMOTRON_H_MOE   = auto()
     EXAONE           = auto()
     EXAONE4          = auto()
     GRANITE          = auto()
@@ -642,6 +643,7 @@ class MODEL_TENSOR(IntEnum):
     V_MMPROJ_PEG         = auto()
     V_ENC_EMBD_CLS       = auto()
     V_ENC_EMBD_PATCH     = auto()
+    V_ENC_EMBD_NORM      = auto()
     V_ENC_EMBD_POS       = auto()
     V_ENC_INPUT_NORM     = auto()
     V_ENC_ATTN_QKV       = auto()
@@ -660,6 +662,7 @@ class MODEL_TENSOR(IntEnum):
     V_LAYER_SCALE_2      = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_POST_NORM       = auto()
     V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
@@ -786,6 +789,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.JAIS:             "jais",
     MODEL_ARCH.NEMOTRON:         "nemotron",
     MODEL_ARCH.NEMOTRON_H:       "nemotron_h",
+    MODEL_ARCH.NEMOTRON_H_MOE:   "nemotron_h_moe",
     MODEL_ARCH.EXAONE:           "exaone",
     MODEL_ARCH.EXAONE4:          "exaone4",
     MODEL_ARCH.GRANITE:          "granite",
@@ -1014,6 +1018,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_MMPROJ_PEG:              "mm.model.peg.{bid}",
     MODEL_TENSOR.V_ENC_EMBD_CLS:            "v.class_embd",
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
+    MODEL_TENSOR.V_ENC_EMBD_NORM:           "v.norm_embd",
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
     MODEL_TENSOR.V_ENC_ATTN_QKV:            "v.blk.{bid}.attn_qkv",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
@@ -1032,6 +1037,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_LAYER_SCALE_2:           "v.blk.{bid}.ls2",
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
+    MODEL_TENSOR.V_MM_POST_NORM:            "mm.post_norm",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
     MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
@@ -1092,6 +1098,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_MMPROJ_PEG,
         MODEL_TENSOR.V_ENC_EMBD_CLS,
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
+        MODEL_TENSOR.V_ENC_EMBD_NORM,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_INPUT_NORM,
         MODEL_TENSOR.V_ENC_ATTN_QKV,
@@ -1110,6 +1117,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_LAYER_SCALE_2,
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
+        MODEL_TENSOR.V_MM_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
         MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
@@ -2529,6 +2537,33 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.NEMOTRON_H_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_NORM,
+        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        # experts
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        # shared expert
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
     MODEL_ARCH.EXAONE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -3320,6 +3355,7 @@ class VisionProjectorType:
     ULTRAVOX = "ultravox"
     INTERNVL = "internvl"
     QWEN2A = "qwen2a" # audio
+    GLMA = "glma" # audio
     QWEN25O = "qwen2.5o" # omni
     VOXTRAL = "voxtral"
     LFM2 = "lfm2"
@@ -3327,6 +3363,7 @@ class VisionProjectorType:
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"
+    GLM4V = "glm4v"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -154,7 +154,8 @@ class TensorNameMap:
             "model.layers.{bid}.operator_norm",                     # lfm2
             "model.transformer.blocks.{bid}.attn_norm",             # llada
             "layers.{bid}.input_layernorm",                         # qwen3-embedding
-            "model.layers.{bid}.attention_layernorm"                # apertus
+            "model.layers.{bid}.attention_layernorm",               # apertus
+            "model.layers.{bid}.pre_attention_layernorm",           # kormo
         ),
 
         # Attention norm 2
@@ -342,6 +343,7 @@ class TensorNameMap:
             "model.transformer.blocks.{bid}.ff_norm",                        # llada
             "layers.{bid}.post_attention_layernorm",                         # qwen3-embedding
             "model.layers.{bid}.feedforward_layernorm",                      # apertus
+            "model.layers.{bid}.pre_mlp_layernorm",                          # kormo
         ),
 
         # Pre feed-forward norm
@@ -377,6 +379,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.gate",               # lfm2moe
             "model.layers.{bid}.mlp.router.gate",               # afmoe
             "layers.{bid}.gate",                                # mistral-large
+            "backbone.layers.{bid}.mixer.gate",                 # nemotron-h-moe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -390,6 +393,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.expert_bias",                           # afmoe
             "model.layers.{bid}.feed_forward.expert_bias",                  # lfm2moe
             "model.layers.{bid}.block_sparse_moe.e_score_correction",       # minimax-m2
+            "backbone.layers.{bid}.mixer.gate.e_score_correction"           # nemotron-h-moe
         ),
 
         # Feed-forward up
@@ -438,7 +442,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.v1",          # dbrx
-            "model.layers.{bid}.mlp.experts.up_proj",               # qwen2moe olmoe (merged) ernie4.5-moe
+            "model.layers.{bid}.mlp.experts.up_proj",               # qwen2moe olmoe (merged) ernie4.5-moe, nemotron-h-moe (merged)
             "model.layers.{bid}.block_sparse_moe.experts.w3",       # phimoe (merged)
             "model.layers.{bid}.feed_forward.experts.up_proj",      # llama4
             "encoder.layers.{bid}.mlp.experts.mlp.w1",              # nomic-bert-moe
@@ -452,6 +456,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.down_proj",
             "model.layers.{bid}.mlp.shared_mlp.up_proj",             # hunyuan
             "layers.{bid}.shared_experts.w3",                        # mistral-large
+            "backbone.layers.{bid}.mixer.shared_experts.up_proj",    # nemotron-h-moe
         ),
 
         MODEL_TENSOR.FFN_UP_CHEXP: (
@@ -546,7 +551,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
             "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.w2",          # dbrx
-            "model.layers.{bid}.mlp.experts.down_proj",             # qwen2moe olmoe (merged) ernie4.5-moe
+            "model.layers.{bid}.mlp.experts.down_proj",             # qwen2moe olmoe (merged) ernie4.5-moe nemotron-h-moe (merged)
             "model.layers.{bid}.block_sparse_moe.output_linear",    # granitemoe
             "model.layers.{bid}.block_sparse_moe.experts.w2",       # phimoe (merged)
             "model.layers.{bid}.feed_forward.experts.down_proj",    # llama4
@@ -561,6 +566,7 @@ class TensorNameMap:
             "model.layers.{bid}.shared_mlp.output_linear",             # granitemoe
             "model.layers.{bid}.mlp.shared_mlp.down_proj",             # hunyuan
             "layers.{bid}.shared_experts.w2",                          # mistral-large
+            "backbone.layers.{bid}.mixer.shared_experts.down_proj",    # nemotron-h-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_CHEXP: (
@@ -704,6 +710,7 @@ class TensorNameMap:
             "model.layers.{bid}.mamba.dt_proj",         # jamba falcon-h1 granite-hybrid
             "model.layers.layers.{bid}.mixer.dt_proj",  # plamo2
             "model.layers.{bid}.linear_attn.dt_proj",   # qwen3next
+            "backbone.layers.{bid}.mixer.dt",           # nemotron-h-moe
         ),
 
         MODEL_TENSOR.SSM_DT_NORM: (
@@ -1205,6 +1212,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_MMPROJ_FC: (
             "model.connector.modality_projection.proj", # SmolVLM
             "model.vision.linear_proj.linear_proj", # cogvlm
+            "visual.merger.proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MMPROJ_MLP: (
@@ -1238,6 +1246,10 @@ class TensorNameMap:
             "model.vision.patch_embedding.proj", # cogvlm
         ),
 
+        MODEL_TENSOR.V_ENC_EMBD_NORM: (
+            "visual.post_conv_layernorm", # glm4v
+        ),
+
         MODEL_TENSOR.V_ENC_EMBD_POS: (
             "vision_tower.vision_model.embeddings.position_embedding",
             "model.vision_tower.embeddings.position_embeddings", # Intern-S1
@@ -1247,6 +1259,7 @@ class TensorNameMap:
             "vision_tower.patch_embed.pos_emb", # kimi-vl
             "visual.pos_embed", # qwen3vl
             "model.vision.patch_embedding.position_embedding", # cogvlm
+            "visual.embeddings.position_embedding", # glm4v
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
@@ -1402,6 +1415,11 @@ class TensorNameMap:
             "vision_model.layernorm_post", # llama4
             "visual.merger.ln_q", # qwen2vl
             "vision_tower.encoder.final_layernorm", # kimi-vl
+            "visual.post_layernorm", # glm4v
+        ),
+
+        MODEL_TENSOR.V_MM_POST_NORM: (
+            "visual.merger.post_projection_norm", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_INP_PROJ: (
@@ -1471,6 +1489,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_MM_PATCH_MERGER: (
             "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1 - hf
             "patch_merger.merging_layer", # mistral
+            "visual.downsample", # glm4v
         ),
 
         MODEL_TENSOR.V_DS_NORM: (
@@ -1491,14 +1510,17 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MM_UP: (
             "model.vision.linear_proj.dense_h_to_4h", # cogvlm
+            "visual.merger.up_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_DOWN: (
             "model.vision.linear_proj.dense_4h_to_h", # cogvlm
+            "visual.merger.down_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_GATE: (
             "model.vision.linear_proj.gate_proj", # cogvlm
+            "visual.merger.gate_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_TOK_BOI: (

grammars/README.md

@@ -1,6 +1,6 @@
 # GBNF Guide
 
-GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `tools/main` and `tools/server`.
+GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `tools/cli`, `tools/completion` and `tools/server`.
 
 ## Background
 
@@ -135,7 +135,7 @@ While semantically correct, the syntax `x? x? x?.... x?` (with N repetitions) ma
 You can use GBNF grammars:
 
 - In [llama-server](../tools/server)'s completion endpoints, passed as the `grammar` body field
-- In [llama-cli](../tools/main), passed as the `--grammar` & `--grammar-file` flags
+- In [llama-cli](../tools/cli) and [llama-completion](../tools/completion), passed as the `--grammar` & `--grammar-file` flags
 - With [test-gbnf-validator](../tests/test-gbnf-validator.cpp), to test them against strings.
 
 ## JSON Schemas → GBNF
@@ -145,7 +145,7 @@ You can use GBNF grammars:
 - In [llama-server](../tools/server):
     - For any completion endpoints, passed as the `json_schema` body field
     - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}` or `{ type: "json_schema", json_schema: {"schema": ...} }`)
-- In [llama-cli](../tools/main), passed as the `--json` / `-j` flag
+- In [llama-cli](../tools/cli) and [llama-completion](../tools/completion), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
     - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)
     - in JavaScript with [json-schema-to-grammar.mjs](../tools/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../tools/server)'s Web UI)

include/llama.h

@@ -313,6 +313,7 @@ extern "C" {
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
         bool no_host;         // bypass host buffer allowing extra buffers to be used
+        bool no_alloc;        // only load metadata and simulate memory allocations
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
@@ -466,10 +467,24 @@ extern "C" {
     // Frees all allocated memory
     LLAMA_API void llama_free(struct llama_context * ctx);
 
+    // fits mparams and cparams to free device memory (assumes system memory is unlimited)
+    // returns true if the parameters could be successfully modified to fit device memory
+    // this function is NOT thread safe because it modifies the global llama logger state
+    LLAMA_API bool llama_params_fit(
+                                   const char   * path_model,
+                    struct llama_model_params   * mparams,
+                    struct llama_context_params * cparams,
+                                          float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
+        struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
+                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                       uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+                            enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
+
     LLAMA_API int64_t llama_time_us(void);
 
     LLAMA_API size_t llama_max_devices(void);
     LLAMA_API size_t llama_max_parallel_sequences(void);
+    LLAMA_API size_t llama_max_tensor_buft_overrides(void);
 
     LLAMA_API bool llama_supports_mmap       (void);
     LLAMA_API bool llama_supports_mlock      (void);
@@ -1354,7 +1369,9 @@ extern "C" {
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
-    LLAMA_API void llama_log_set(ggml_log_callback log_callback, void * user_data);
+    // The logger state is global so these functions are NOT thread safe.
+    LLAMA_API void llama_log_get(ggml_log_callback * log_callback, void ** user_data);
+    LLAMA_API void llama_log_set(ggml_log_callback   log_callback, void *  user_data);
 
     //
     // Performance utils

models/templates/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.jinja

@@ -0,0 +1,204 @@
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {{- content }}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

scripts/compare-logprobs.py

@@ -0,0 +1,281 @@
+import argparse
+import requests
+import json
+from pathlib import Path
+import logging
+
+logger = logging.getLogger("compare-logprobs")
+logging.basicConfig(level=logging.INFO)
+
+
+DESCRIPTION = """
+Compare logits between llama.cpp and another inference engine using OpenAI-compatible server endpoints.
+
+Unlike compare-logits.py, it allows dumping logits from a hosted API endpoint. Useful when it's not possible to run both models locally.
+
+Example usage:
+    Step 1: Dump logits from two different servers
+        python scripts/compare-logprobs.py dump logits_llama.log http://localhost:8080/v1/completions
+        python scripts/compare-logprobs.py dump logits_other.log http://other-engine:8000/v1/completions
+
+        (optionally, you can add --api-key <key> if the endpoint requires authentication)
+
+    Step 2: Compare the dumped logits
+        python scripts/compare-logprobs.py compare logits_llama.log logits_other.log report.md
+"""
+
+
+def generate_input_prompt(length: int) -> list[str]:
+    CORPUS = """
+    You are an advanced AI assistant capable of using tools to gather information, perform calculations, or execute tasks. Always think step by step before responding. If a user's query requires external data, computation, or actions beyond your internal knowledge, use the appropriate tools via function calls.
+
+    ### Tool Call Format:
+    When you need to use a tool, output the call in this exact XML format. Include the opening and closing tags. Do not escape arguments; they will be parsed as plain text.
+
+    You can make multiple calls in one go by placing them one after another.
+    """
+    words = [w.strip() for w in CORPUS.strip().split(" ")]
+    words = [w for w in words if len(w) > 0]  # filter out empty strings
+    while len(words) < length:
+        words += words
+    return words[:length]
+
+
+def dump_logits(
+    endpoint: str,
+    output_path: Path,
+    input_words: list[str],
+    pattern: list[tuple[bool, int]],
+    api_key=None,
+):
+    logger.info(f"Dumping logits to {output_path} from endpoint {endpoint}...")
+    words = input_words
+    curr_text = ""
+    n_total = sum(n for get, n in pattern if get)
+    n_done = 0
+    i_cur = 0
+    i_total = len(words)
+    with output_path.open("w") as f:
+        for get, n in pattern:
+            if not get:
+                # skip n words
+                for i in range(n):
+                    curr_text += words.pop(0) + " "
+                    i_cur += 1
+                continue
+            # get n words
+            for i in range(n):
+                curr_text += words.pop(0) + " "
+                payload = {
+                    "prompt": curr_text.strip(),
+                    "temperature": 0.0,
+                    "top_k": 1,
+                    "max_tokens": 1,
+                    "logprobs": 1,
+                    "stream": False,
+                }
+                response = requests.post(
+                    endpoint,
+                    json=payload,
+                    headers={"Authorization": f"Bearer {api_key}"} if api_key else {},
+                )
+                response.raise_for_status()
+                data = response.json()
+                data["__index"] = i_cur  # add index for easier debugging later
+                data = json.dumps(data)
+                f.write(f"{data}\n")
+                n_done += 1
+                i_cur += 1
+                logger.info(
+                    f"\n\n{data}\n\n[Step: {n_done}/{n_total} | Word: {i_cur}/{i_total}]"
+                )
+    logger.info(f"Logits dumped to {output_path}")
+
+
+def get_token_logprobs(data: dict):
+    logprobs = data["choices"][0]["logprobs"]
+    if "content" in logprobs:
+        # llama.cpp case
+        top = logprobs["content"][0]["top_logprobs"][0]
+        return top["token"], top["logprob"]
+    else:
+        # vllm case
+        tokens = logprobs["tokens"]
+        token_logprobs = logprobs["token_logprobs"]
+        return tokens[0], token_logprobs[0]
+
+
+def clean_text(text: str) -> str:
+    return (
+        "'"
+        + text.replace("\n", "\\n")
+        .replace("\t", "\\t")
+        .replace("\r", "\\r")
+        .replace("|", "\\|")
+        + "'"
+    )
+
+
+def compare_logits(input1: Path, input2: Path, output_path: Path):
+    with input1.open("r") as f1, input2.open("r") as f2, output_path.open("w") as fout:
+        lines1 = f1.readlines()
+        lines2 = f2.readlines()
+
+        tab_header = [
+            "idx",
+            input1.name,
+            "logprob_1",
+            input2.name,
+            "logprob_2",
+            "diff (abs)",
+        ]
+        tab_entries = []
+        tab_max_widths = [len(h) for h in tab_header]
+
+        assert len(lines1) == len(
+            lines2
+        ), "Input files must have the same number of lines."
+
+        fout.write("# Logits Comparison Report\n\n")
+        for i, (line1, line2) in enumerate(zip(lines1, lines2)):
+            if not line1.strip() or not line2.strip():
+                continue  # skip empty lines
+
+            data1 = json.loads(line1)
+            data2 = json.loads(line2)
+
+            idx1 = data1.get("__index", -1)
+            idx2 = data2.get("__index", -1)
+            if idx1 != idx2:
+                logger.warning(
+                    f"Warning: Mismatched indices at line {i}: {idx1} vs {idx2}"
+                )
+
+            token1, logprob1 = get_token_logprobs(data1)
+            token2, logprob2 = get_token_logprobs(data2)
+
+            token1 = clean_text(token1)
+            token2 = clean_text(token2)
+            abs_diff = abs(logprob1 - logprob2)
+
+            tab_entries.append(
+                (
+                    str(idx1 + 1),
+                    token1,
+                    f"{logprob1:.4f}",
+                    token2,
+                    f"{logprob2:.4f}",
+                    f"{(abs_diff):.4f}",
+                )
+            )
+
+        for i in range(len(tab_entries)):
+            for j in range(len(tab_header)):
+                tab_max_widths[j] = max(tab_max_widths[j], len(tab_entries[i][j]))
+
+        output = ""
+        for j in range(len(tab_header)):
+            output += f"| {tab_header[j]:<{tab_max_widths[j]}} "
+        output += "|\n"
+        for j in range(len(tab_header)):
+            output += f"|{'-' * (tab_max_widths[j] + 2)}"
+        output += "|\n"
+        for entry in tab_entries:
+            for j in range(len(tab_header)):
+                output += f"| {entry[j]:<{tab_max_widths[j]}} "
+            output += "|\n"
+
+        logger.info("\n" + output)
+        fout.write(output)
+        logger.info(f"Report written to {output_path}")
+
+
+def parse_pattern(pattern: str) -> list[tuple[bool, int]]:
+    parts = pattern.split(",")
+    result = []
+    for i, part in enumerate(parts):
+        n = int(part)
+        if i % 2 == 0:
+            result.append((True, n))  # get n words
+        else:
+            result.append((False, n))  # skip n words
+    return result
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description=DESCRIPTION, formatter_class=argparse.RawTextHelpFormatter
+    )
+    subparsers = parser.add_subparsers(
+        dest="verb", required=True, help="action to perform"
+    )
+
+    # dump subcommand
+    parser_dump = subparsers.add_parser("dump", help="dump logits from an endpoint")
+    parser_dump.add_argument(
+        "output", type=Path, help="output path for dumped logits (.log)"
+    )
+    parser_dump.add_argument(
+        "endpoint", type=str, help="OAI-compat /completions endpoint"
+    )
+    parser_dump.add_argument(
+        "--api-key",
+        type=str,
+        default=None,
+        help="API key for authentication (if required)",
+    )
+    parser_dump.add_argument(
+        "--file",
+        type=Path,
+        default=None,
+        help="File containing prompt to use instead of the default",
+    )
+    parser_dump.add_argument(
+        "--pattern",
+        type=str,
+        default="10,1000,10,4000,10",
+        help="Pattern n_get,n_skip,... where n_get is number of words to get and n_skip is number of words to skip (num of words, NOT num of tokens)",
+    )
+
+    # compare subcommand
+    parser_compare = subparsers.add_parser(
+        "compare", help="compare two dumped logits files"
+    )
+    parser_compare.add_argument("input1", type=Path, help="first input file (.log)")
+    parser_compare.add_argument("input2", type=Path, help="second input file (.log)")
+    parser_compare.add_argument(
+        "output", type=Path, help="output path for comparison report (.md)"
+    )
+
+    try:
+        return parser.parse_args()
+    except Exception as e:
+        parser.print_help()
+        raise e
+
+
+def main():
+    args = parse_args()
+
+    if args.verb == "dump":
+        pattern = parse_pattern(args.pattern)
+        input_length = sum(n for _, n in pattern)
+        input_words = generate_input_prompt(input_length)
+        if args.file is not None:
+            with args.file.open("r") as f:
+                input_words = f.read().strip().split(" ")
+                if input_length < sum(n for _, n in pattern):
+                    raise ValueError(
+                        f"Input file has only {input_length} words, but pattern requires at least {input_length} words."
+                    )
+                input_length = len(input_words)
+        logger.info(f"Using {input_length} words")
+        dump_logits(args.endpoint, args.output, input_words, pattern, args.api_key)
+    elif args.verb == "compare":
+        compare_logits(args.input1, args.input2, args.output)
+    else:
+        raise ValueError(f"Unknown verb: {args.verb}")
+
+
+if __name__ == "__main__":
+    main()

scripts/snapdragon/adb/run-mtmd.sh

@@ -0,0 +1,65 @@
+#!/bin/sh
+#
+
+# Basedir on device
+basedir=/data/local/tmp/llama.cpp
+
+cli_opts=
+
+branch=.
+[ "$B" != "" ] && branch=$B
+
+adbserial=
+[ "$S" != "" ] && adbserial="-s $S"
+
+model="gemma-3-4b-it-Q4_0.gguf"
+[ "$M" != "" ] && model="$M"
+
+mmproj="mmproj-F16.gguf"
+[ "$MMPROJ" != "" ] && mmproj="$MMPROJ"
+
+image=
+[ "$IMG" != "" ] && image="$IMG"
+
+device="HTP0"
+[ "$D" != "" ] && device="$D"
+
+verbose=
+[ "$V" != "" ] && verbose="GGML_HEXAGON_VERBOSE=$V"
+
+experimental="GGML_HEXAGON_EXPERIMENTAL=1"
+[ "$E" != "" ] && experimental="GGML_HEXAGON_EXPERIMENTAL=$E"
+
+sched=
+[ "$SCHED" != "" ] && sched="GGML_SCHED_DEBUG=2" cli_opts="$cli_opts -v"
+
+profile=
+[ "$PROF" != "" ] && profile="GGML_HEXAGON_PROFILE=$PROF GGML_HEXAGON_OPSYNC=1"
+
+opmask=
+[ "$OPMASK" != "" ] && opmask="GGML_HEXAGON_OPMASK=$OPMASK"
+
+nhvx=
+[ "$NHVX" != "" ] && nhvx="GGML_HEXAGON_NHVX=$NHVX"
+
+ndev=
+[ "$NDEV" != "" ] && ndev="GGML_HEXAGON_NDEV=$NDEV"
+
+# MTMD backend device for vision model (defaults to CPU if not set)
+mtmd_backend=
+[ "$MTMD_DEVICE" != "" ] && mtmd_backend="MTMD_BACKEND_DEVICE=$MTMD_DEVICE"
+
+set -x
+
+adb $adbserial shell " \
+  cd $basedir; ulimit -c unlimited;        \
+    LD_LIBRARY_PATH=$basedir/$branch/lib   \
+    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
+    $verbose $experimental $sched $opmask $profile $nhvx $ndev $mtmd_backend       \
+      ./$branch/bin/llama-mtmd-cli --no-mmap -m $basedir/../gguf/$model   \
+         --mmproj $basedir/../gguf/$mmproj \
+         --image $basedir/../gguf/$image \
+         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1             \
+         --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on \
+         -ngl 99 --device $device -v $cli_opts $@ \
+"

scripts/sync-ggml.last

@@ -1 +1 @@
-55bc9320a4aae82af18e23eefd5de319a755d7b9
+130bc125a88bb57664b88932c48c38a1cb316fac

src/llama-arch.cpp

@@ -75,6 +75,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
+    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
@@ -1763,6 +1764,39 @@ 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_NEMOTRON_H_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,     "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,    "output_norm" },
+            { LLM_TENSOR_OUTPUT,         "output" },
+            { LLM_TENSOR_ATTN_NORM,      "blk.%d.attn_norm" },
+            // mamba(2) ssm layers
+            { LLM_TENSOR_SSM_IN,         "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,     "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,         "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,          "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,          "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,       "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,        "blk.%d.ssm_out" },
+            // attention layers
+            { 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" },
+            // dense FFN
+            { LLM_TENSOR_FFN_DOWN,       "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,         "blk.%d.ffn_up" },
+            // MoE FFN (for MoE layers)
+            { LLM_TENSOR_FFN_GATE_INP,   "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_UP_EXPS,    "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,  "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,"blk.%d.exp_probs_b" },
+            // MoE shared expert layer
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_EXAONE,
         {
@@ -2817,6 +2851,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_QWEN3NEXT:
             return true;
         default:

src/llama-arch.h

@@ -79,6 +79,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
+    LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,

src/llama-context.cpp

@@ -9,6 +9,7 @@
 #include "llama-model.h"
 
 #include <cinttypes>
+#include <cmath>
 #include <cstring>
 #include <limits>
 #include <stdexcept>
@@ -72,6 +73,43 @@ llama_context::llama_context(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    if (cparams.yarn_ext_factor != 0) {
+        static auto get_mscale = [](float scale, float mscale) {
+            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+        };
+
+        const float factor = 1.0f / cparams.rope_freq_scale;
+
+        // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
+        if (hparams.rope_yarn_log_mul != 0.0f) {
+            // note: here we assume `mscale == 1.0f`
+            // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
+                  float mscale          = 1.0f;
+            const float mscale_all_dims = hparams.rope_yarn_log_mul;
+
+            // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            // special-case DEEPSEEK v2:
+            // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
+            if (model.arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
+                mscale = mscale_all_dims;
+            }
+
+            cparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+
+            LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
+                    __func__, cparams.yarn_attn_factor, mscale, mscale_all_dims);
+        } else {
+            cparams.yarn_attn_factor = get_mscale(factor, 1.0f);
+        }
+
+        // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
+        // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
+        //
+        // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
+        //      https://github.com/ggml-org/llama.cpp/pull/17945
+        cparams.yarn_attn_factor *= 1.0f / (1.0f + 0.1f * logf(factor));
+    }
+
     cparams.yarn_attn_factor *= hparams.rope_attn_factor;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -220,6 +258,7 @@ llama_context::llama_context(
 
         backend_buft.clear();
         backend_ptrs.clear();
+        backend_buf_exp_size.clear();
 
         for (auto & backend : backends) {
             auto * buft = ggml_backend_get_default_buffer_type(backend.get());
@@ -236,6 +275,7 @@ llama_context::llama_context(
 
             backend_buft.push_back(buft);
             backend_ptrs.push_back(backend.get());
+            backend_buf_exp_size.push_back(0);
         }
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
@@ -351,7 +391,8 @@ llama_context::llama_context(
 
         // reserve pp (prompt processing) graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(),
+                model.hparams.no_alloc, model.hparams.no_alloc ? backend_buf_exp_size.data() : nullptr);
             if (!gf) {
                 if (pipeline_parallel) {
                     LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__);
@@ -369,7 +410,7 @@ llama_context::llama_context(
 
         // reserve with tg (token generation) graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -384,7 +425,7 @@ llama_context::llama_context(
             //
             // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
             //
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get(), model.hparams.no_alloc);
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -393,11 +434,13 @@ llama_context::llama_context(
         for (size_t i = 0; i < backend_ptrs.size(); ++i) {
             ggml_backend_t             backend = backend_ptrs[i];
             ggml_backend_buffer_type_t buft    = backend_buft[i];
-            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
-            if (size > 1) {
+            if (!model.hparams.no_alloc) {
+                backend_buf_exp_size[i] = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+            }
+            if (backend_buf_exp_size[i] > 1) {
                 LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                         ggml_backend_buft_name(buft),
-                        size / 1024.0 / 1024.0);
+                        backend_buf_exp_size[i] / 1024.0 / 1024.0);
             }
         }
 
@@ -416,6 +459,23 @@ llama_context::llama_context(
 }
 
 llama_context::~llama_context() {
+    // FIXME this currently results in a use-after-free bug if the model is freed before the context
+    // if (!model.hparams.no_alloc) {
+    //     for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+    //         ggml_backend_t             backend = backend_ptrs[i];
+    //         ggml_backend_buffer_type_t buft    = backend_buft[i];
+
+    //         const size_t size_exp = backend_buf_exp_size[i];
+    //         const size_t size_act = ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    //         if (size_exp == size_act) {
+    //             LLAMA_LOG_DEBUG("%s: %10s compute buffer size is %8.4f MiB, matches expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         } else {
+    //             LLAMA_LOG_WARN("%s: %10s compute buffer size of %8.4f MiB, does not match expectation of %8.4f MiB\n",
+    //                 __func__, ggml_backend_buft_name(buft), size_act / (1024.0*1024.0), size_exp / (1024.0*1024.0));
+    //         }
+    //     }
+    // }
     ggml_opt_free(opt_ctx);
 }
 
@@ -1318,6 +1378,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
             LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+            synchronize();
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;
@@ -1389,7 +1450,8 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
     return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
+ggml_cgraph * llama_context::graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only, size_t * sizes) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
     GGML_ASSERT(n_outputs >= 1);
 
@@ -1426,8 +1488,13 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     // initialize scheduler with the specified graph
     if (split_only) {
-        ggml_backend_sched_split_graph(sched.get(), gf);
+        if (sizes) {
+            ggml_backend_sched_reserve_size(sched.get(), gf, sizes);
+        } else {
+            ggml_backend_sched_split_graph(sched.get(), gf);
+        }
     } else if (!ggml_backend_sched_reserve(sched.get(), gf)) {
+        GGML_ASSERT(!sizes);
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
         return nullptr;
     }
@@ -2049,15 +2116,26 @@ void llama_context::perf_reset() {
 
 std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> llama_context::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> ret;
-    for (const auto & buft_size : model.memory_breakdown()) {
-        ret[buft_size.first].model += buft_size.second;
+    for (const auto & [buft, size] : model.memory_breakdown()) {
+        ret[buft].model += size;
     }
-    for (const auto & buft_size : memory->memory_breakdown()) {
-        ret[buft_size.first].context += buft_size.second;
+    if (memory) {
+        for (const auto & [buft, size] : memory->memory_breakdown()) {
+            ret[buft].context += size;
+        }
     }
-    for (const auto & backend_ptr : backends) {
-        ggml_backend_t backend = backend_ptr.get();
-        ret[ggml_backend_sched_get_buffer_type(sched.get(), backend)].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+    if (model.hparams.no_alloc) {
+        for (size_t i = 0; i < backends.size(); ++i) {
+            ggml_backend_t             backend = backends[i].get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += backend_buf_exp_size[i];
+        }
+    } else {
+        for (const auto & backend_ptr : backends) {
+            ggml_backend_t             backend = backend_ptr.get();
+            ggml_backend_buffer_type_t buft    = ggml_backend_sched_get_buffer_type(sched.get(), backend);
+            ret[buft].compute += ggml_backend_sched_get_buffer_size(sched.get(), backend);
+        }
     }
     return ret;
 }

src/llama-context.h

@@ -26,6 +26,10 @@ struct llama_memory_breakdown_data {
     size_t model   = 0; // memory allocated for the model
     size_t context = 0; // memory allocated for the context
     size_t compute = 0; // memory allocated for temporary compute buffers
+
+    size_t total() const {
+        return model + context + compute;
+    }
 };
 
 struct llama_context {
@@ -206,7 +210,8 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false);
+    ggml_cgraph * graph_reserve(
+        uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only = false, size_t * sizes = nullptr);
 
 private:
     llm_graph_params graph_params(
@@ -281,9 +286,10 @@ private:
 
     std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;
 
-    // buffer types used for the compute buffer of each backend
+    // pointers and 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;
+    std::vector<size_t>                     backend_buf_exp_size; // expected buffer sizes
 
     llm_graph_result_ptr gf_res_prev;
     llm_graph_result_ptr gf_res_reserve;

src/llama-graph.cpp

@@ -78,7 +78,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
         for (int i = 0; i < n_tokens; ++i) {
             const float pos = ubatch->pos[i];
             attn_scale_data[i] = std::log(
-                std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0
+                std::floor((pos + f_attn_temp_offset) / n_attn_temp_floor_scale) + 1.0
             ) * f_attn_temp_scale + 1.0;
         }
 
@@ -254,6 +254,24 @@ void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_rs::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_recurrent_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= s_copy->ne[0] == mctx->get_n_rs();
+
+    res &= s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= s_copy_extra->ne[0] == mctx->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= head == mctx->get_head();
+    res &= rs_z == mctx->get_rs_z();
+
+    return res;
+}
+
 void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
@@ -461,8 +479,46 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    inp_attn->set_input(ubatch);
-    inp_rs->set_input(ubatch);
+    mctx->get_attn()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
+    mctx->get_attn()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+
+    mctx->get_attn()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    if (inp_rs->s_copy) {
+        GGML_ASSERT(ggml_backend_buffer_is_host(inp_rs->s_copy->buffer));
+        int32_t * data = (int32_t *) inp_rs->s_copy->data;
+
+        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_hybrid_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= inp_attn->self_kq_mask->ne[0] == mctx->get_attn()->get_n_kv();
+    res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+
+    res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
+
+    res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= inp_rs->s_copy_extra->ne[0] == mctx->get_recr()->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= inp_rs->head == mctx->get_recr()->get_head();
+    res &= inp_rs->rs_z == mctx->get_recr()->get_rs_z();
+
+    return res;
 }
 
 //
@@ -574,7 +630,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     freq_base        (cparams.rope_freq_base),
     freq_scale       (cparams.rope_freq_scale),
     ext_factor       (cparams.yarn_ext_factor),
-    attn_factor      (llama_hparams::yarn_attn_factor_adjust(cparams.yarn_attn_factor, cparams.rope_freq_scale, cparams.yarn_ext_factor)),
+    attn_factor      (cparams.yarn_attn_factor),
     beta_fast        (cparams.yarn_beta_fast),
     beta_slow        (cparams.yarn_beta_slow),
     norm_eps         (hparams.f_norm_eps),
@@ -1089,6 +1145,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_moe_relu", il);
             } break;
+        case LLM_FFN_RELU_SQR:
+            if (gate_exps) {
+                // TODO: add support for gated squared relu
+                GGML_ABORT("fatal error: gated squared relu not implemented");
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cur = ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_moe_relu_sqr", il);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -1203,7 +1268,7 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
-    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale, hparams.f_attn_temp_offset);
 
     auto & cur = inp->attn_scale;
 
@@ -1841,6 +1906,9 @@ static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
     inp->s_copy_main  = ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0);
     inp->s_copy_extra = ggml_view_1d(ctx0, inp->s_copy, n_rs - n_seqs, n_seqs * inp->s_copy->nb[0]);
 
+    inp->head = mctx_cur->get_head();
+    inp->rs_z = mctx_cur->get_rs_z();
+
     return inp;
 }
 
@@ -1909,10 +1977,10 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
 llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
 
-    auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
+    auto inp_rs   = build_rs_inp_impl     (ctx0, ubatch, mctx_cur->get_recr());
     auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());
 
-    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
 
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }

src/llama-graph.h

@@ -132,8 +132,8 @@ public:
 // temperature tuning, used by llama4
 class llm_graph_input_attn_temp : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
-        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale, float f_attn_temp_offset)
+        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale), f_attn_temp_offset(f_attn_temp_offset) {}
     virtual ~llm_graph_input_attn_temp() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -142,6 +142,7 @@ public:
 
     const uint32_t n_attn_temp_floor_scale;
     const float    f_attn_temp_scale;
+    const float    f_attn_temp_offset;
 };
 
 class llm_graph_input_pos_bucket : public llm_graph_input_i {
@@ -224,6 +225,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * s_copy;  // I32 [n_rs]
 
     // views of s_copy, computed once per graph
@@ -232,6 +235,10 @@ public:
     ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]
 
     const llama_memory_recurrent_context * mctx;
+
+    // used in view offsets, need to match for valid graph reuse
+    uint32_t head;
+    int32_t rs_z;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -364,22 +371,28 @@ public:
 class llm_graph_input_mem_hybrid : public llm_graph_input_i {
 public:
     llm_graph_input_mem_hybrid(
+            const llama_cparams & cparams,
             std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
-            std::unique_ptr<llm_graph_input_rs>              inp_rs,
-            const llama_memory_hybrid_context *              mctx) :
+            std::unique_ptr<llm_graph_input_rs>      inp_rs,
+            const llama_memory_hybrid_context *      mctx) :
         inp_attn(std::move(inp_attn)),
         inp_rs(std::move(inp_rs)),
+        cparams(cparams),
         mctx(mctx) { }
     virtual ~llm_graph_input_mem_hybrid() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
     std::unique_ptr<llm_graph_input_rs>      inp_rs;
 
     llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
     llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }
 
+    const llama_cparams cparams;
+
     const llama_memory_hybrid_context * mctx;
 };
 

src/llama-hparams.cpp

@@ -2,8 +2,8 @@
 
 #include "ggml.h"
 
+#include <algorithm>
 #include <cassert>
-#include <cmath>
 
 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
     if (dense_first) {
@@ -232,12 +232,6 @@ bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama
     return false;
 }
 
-float llama_hparams::yarn_attn_factor_adjust(float attn_factor, float freq_scale, float ext_factor) {
-    GGML_ASSERT(ext_factor >= 0.0f);
-
-    if (ext_factor != 0.0f) {
-        attn_factor *= 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
-    }
-
-    return attn_factor;
+bool llama_hparams::use_mrope() const {
+    return rope_sections[0] > 0 && rope_sections[1] > 0;
 }

src/llama-hparams.h

@@ -34,6 +34,7 @@ struct llama_hparams_convnext {
 
 struct llama_hparams {
     bool vocab_only;
+    bool no_alloc;
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;
@@ -165,6 +166,7 @@ struct llama_hparams {
     uint32_t n_no_rope_layer_step    = 4;
     uint32_t n_attn_temp_floor_scale = 0;
     float    f_attn_temp_scale       = 0.0f;
+    float    f_attn_temp_offset      = 0.0f; // offset position index
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs
@@ -269,12 +271,7 @@ struct llama_hparams {
     // TODO: pack the SWA params in a struct?
     static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
 
-    // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
-    // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
-    //
-    // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
-    //      https://github.com/ggml-org/llama.cpp/pull/17945
-    static float yarn_attn_factor_adjust(float attn_factor, float freq_scale, float ext_factor);
+    bool use_mrope() const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

src/llama-impl.cpp

@@ -25,6 +25,10 @@ time_meas::~time_meas() {
     }
 }
 
+void llama_log_get(ggml_log_callback * log_callback, void ** user_data) {
+    ggml_log_get(log_callback, user_data);
+}
+
 void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     ggml_log_set(log_callback, user_data);
     g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;

src/llama-kv-cache.cpp

@@ -175,7 +175,15 @@ llama_kv_cache::llama_kv_cache(
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (model.hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for kv cache");
         }
@@ -482,9 +490,18 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, buf] : ctxs_bufs) {
+        ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
+
     return ret;
 }
 
@@ -1372,7 +1389,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
     const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
     const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
-    const auto & yarn_attn_factor = llama_hparams::yarn_attn_factor_adjust(cparams.yarn_attn_factor, cparams.rope_freq_scale, cparams.yarn_ext_factor);
+    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
 
     const auto & n_rot     = hparams.n_rot;
     const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
@@ -1544,9 +1561,11 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
         const uint32_t strm = seq_id == -1 ? s : seq_to_stream[seq_id];
 
+        slot_info sinfo;
+
         bool res = true;
-        res = res && state_read_meta(io, strm, cell_count, seq_id);
-        res = res && state_read_data(io, strm, cell_count);
+        res = res && state_read_meta(io, strm, cell_count, sinfo, seq_id);
+        res = res && state_read_data(io, strm, cell_count, sinfo);
 
         if (!res) {
             if (seq_id == -1) {
@@ -1685,7 +1704,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
     }
 }
 
-bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, slot_info & sinfo, llama_seq_id dest_seq_id) {
     auto & cells = v_cells[strm];
     auto & head  = v_heads[strm];
 
@@ -1722,7 +1741,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto sinfo = find_slot(ubatch, true);
+        sinfo = find_slot(ubatch, false);
         if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
@@ -1732,20 +1751,16 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
         //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
         apply_ubatch(sinfo, ubatch);
 
-        const auto head_cur = sinfo.head();
+        LLAMA_LOG_DEBUG("%s: cell_count = %d, dest_seq_id = %d\n", __func__, cell_count, dest_seq_id);
 
-        // keep the head at the old position because we will read the KV data into it in state_read_data()
-        head = head_cur;
-
-        LLAMA_LOG_DEBUG("%s: head_cur = %d, head = %d, cell_count = %d, dest_seq_id = %d\n", __func__, head_cur, head, cell_count, dest_seq_id);
-
-        // DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
-        // Assume that this is one contiguous block of cells
-        GGML_ASSERT(head_cur + cell_count <= cells.size());
-        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
-        GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
-        GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
+        // DEBUG CHECK: verify that all cells were allocated and have correct seq_id and pos values
+        GGML_ASSERT(sinfo.n_stream() == 1);
+        GGML_ASSERT(sinfo.idxs[0].size() == cell_count);
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            const uint32_t idx = sinfo.idxs[0][i];
+            GGML_ASSERT(cells.pos_get(idx) == ubatch.pos[i]);
+            GGML_ASSERT(cells.seq_has(idx, dest_seq_id));
+        }
     } else {
         // whole KV cache restore
 
@@ -1778,15 +1793,24 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             }
         }
 
+        // Create contiguous slot_info for whole cache restore
+        sinfo.s0 = strm;
+        sinfo.s1 = strm;
+        sinfo.resize(1);
+        sinfo.strm[0] = strm;
+        sinfo.idxs[0].resize(cell_count);
+        for (uint32_t i = 0; i < cell_count; ++i) {
+            sinfo.idxs[0][i] = i;
+        }
+
         head = 0;
     }
 
     return true;
 }
 
-bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
+bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo) {
     auto & cells = v_cells[strm];
-    auto & head  = v_heads[strm];
 
     uint32_t v_trans;
     uint32_t n_layer;
@@ -1836,8 +1860,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
         }
 
         if (cell_count) {
-            // Read and set the keys for the whole cell range
-            ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row);
+            if (sinfo.is_contiguous()) {
+                // Fast path: contiguous cells, single memcpy
+                ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), sinfo.head() * k_size_row, cell_count * k_size_row);
+            } else {
+                // Slow path: scatter to non-contiguous positions
+                const void * src = io.read(cell_count * k_size_row);
+                for (uint32_t i = 0; i < cell_count; ++i) {
+                    const size_t dst_offset = sinfo.idxs[0][i] * k_size_row;
+                    ggml_backend_tensor_set(k, (const char*)src + i * k_size_row, dst_offset, k_size_row);
+                }
+            }
         }
     }
 
@@ -1868,8 +1901,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // Read and set the values for the whole cell range
-                ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row);
+                if (sinfo.is_contiguous()) {
+                    // Fast path: contiguous cells, single memcpy
+                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), sinfo.head() * v_size_row, cell_count * v_size_row);
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    const void * src = io.read(cell_count * v_size_row);
+                    for (uint32_t i = 0; i < cell_count; ++i) {
+                        const size_t dst_offset = sinfo.idxs[0][i] * v_size_row;
+                        ggml_backend_tensor_set(v, (const char*)src + i * v_size_row, dst_offset, v_size_row);
+                    }
+                }
             }
         }
     } else {
@@ -1908,10 +1950,22 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             }
 
             if (cell_count) {
-                // For each row in the transposed matrix, read the values for the whole cell range
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    const size_t dst_offset = (head + j * cells.size()) * v_size_el;
-                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                if (sinfo.is_contiguous()) {
+                    // Fast path: contiguous cells
+                    const uint32_t h = sinfo.head();
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const size_t dst_offset = (h + j * cells.size()) * v_size_el;
+                        ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                    }
+                } else {
+                    // Slow path: scatter to non-contiguous positions
+                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                        const void * src = io.read(cell_count * v_size_el);
+                        for (uint32_t i = 0; i < cell_count; ++i) {
+                            const size_t dst_offset = (sinfo.idxs[0][i] + j * cells.size()) * v_size_el;
+                            ggml_backend_tensor_set(v, (const char*)src + i * v_size_el, dst_offset, v_size_el);
+                        }
+                    }
                 }
             }
         }

src/llama-kv-cache.h

@@ -72,6 +72,23 @@ public:
         void clear() {
             idxs.clear();
         }
+
+        // check if indices are contiguous starting from head()
+        bool is_contiguous() const {
+            if (idxs.empty() || idxs[0].empty()) {
+                return true;
+            }
+            if (idxs.size() > 1) {
+                return false;
+            }
+            const uint32_t h = idxs[0][0];
+            for (size_t i = 0; i < idxs[0].size(); ++i) {
+                if (idxs[0][i] != h + i) {
+                    return false;
+                }
+            }
+            return true;
+        }
     };
 
     using slot_info_vec_t = std::vector<slot_info>;
@@ -264,8 +281,8 @@ private:
     void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
 
-    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
-    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
+    bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count,       slot_info & sinfo, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo);
 };
 
 class llama_kv_cache_context : public llama_memory_context_i {

src/llama-memory-hybrid.cpp

@@ -222,7 +222,7 @@ llama_memory_hybrid_context::llama_memory_hybrid_context(
     ubatches(std::move(ubatches)),
     // note: here we copy the ubatches. not sure if this is ideal
     ctx_attn(new llama_kv_cache_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
-    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
     status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
 }
 

src/llama-model-loader.cpp

@@ -473,6 +473,7 @@ llama_model_loader::llama_model_loader(
         std::vector<std::string> & splits,
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) {
     int trace = 0;
@@ -716,6 +717,7 @@ llama_model_loader::llama_model_loader(
 
     this->use_mmap = use_mmap;
     this->check_tensors = check_tensors;
+    this->no_alloc = no_alloc;
 }
 
 std::string llama_model_loader::get_arch_name() const {

src/llama-model-loader.h

@@ -71,6 +71,7 @@ struct llama_model_loader {
 
     bool use_mmap = false;
     bool check_tensors;
+    bool no_alloc;
 
     llama_files files;
     llama_ftype ftype;
@@ -97,6 +98,7 @@ struct llama_model_loader {
         std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
         bool use_mmap,
         bool check_tensors,
+        bool no_alloc,
         const llama_model_kv_override * param_overrides_p,
         const llama_model_tensor_buft_override * param_tensor_buft_overrides_p);
 

src/llama-model.cpp

@@ -120,6 +120,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_16B_A1B:       return "16B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
+        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
         case LLM_TYPE_80B_A3B:       return "80B.A3B";
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
@@ -668,6 +669,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.n_swa                   = 8192;
                     hparams.n_attn_temp_floor_scale = 8192;
                     hparams.f_attn_temp_scale       = 0.1f;
+                    hparams.f_attn_temp_offset      = 1.0f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -1646,6 +1648,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
                 ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
 
+                hparams.f_attn_temp_offset = 0.0f;
+
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
                     case 60: type = LLM_TYPE_236B; break;
@@ -1685,7 +1689,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GLM4:
             {
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
                 switch (hparams.n_layer) {
                     case 40: type = LLM_TYPE_9B; break;
                     case 61: type = LLM_TYPE_32B; break;
@@ -1694,8 +1699,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GLM4_MOE:
             {
-                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
                 // MoE parameters
                 ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
@@ -1794,6 +1800,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
                 ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
@@ -1809,7 +1816,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+
                 switch (hparams.n_layer) {
+                    case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
@@ -2276,6 +2290,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, 0.0f);
 
+                hparams.f_attn_temp_offset = 0.0f;
+
                 // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
                 if (hparams.f_attn_temp_scale != 0.0f) {
                     hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
@@ -2294,32 +2310,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         default: throw std::runtime_error("unsupported model architecture");
     }
 
-    // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
-    if (hparams.rope_yarn_log_mul != 0.0f) {
-        const float factor = 1.0f / hparams.rope_freq_scale_train;
-
-        // note: here we assume `mscale == 1.0f`
-        // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
-              float mscale          = 1.0f;
-        const float mscale_all_dims = hparams.rope_yarn_log_mul;
-
-        // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
-        // special-case DEEPSEEK v2:
-        // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
-        if (arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
-            mscale = mscale_all_dims;
-        }
-
-        static auto get_mscale = [](float scale, float mscale) {
-            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
-        };
-
-        hparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
-
-        LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
-                __func__, hparams.yarn_attn_factor, mscale, mscale_all_dims);
-    }
-
     pimpl->n_bytes = ml.n_bytes;
 
     pimpl->desc_str = arch_name() + " " + type_name() + " " + ml.ftype_name();
@@ -3409,9 +3399,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
                         // optional bias tensors
-                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, 0);
-                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
-                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
@@ -5180,6 +5170,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_NEMOTRON_H:
+            case LLM_ARCH_NEMOTRON_H_MOE:
                 {
                     // mamba2 Mixer SSM params
                     // NOTE: int64_t for tensor dimensions
@@ -5190,6 +5181,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
 
+                    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -5239,12 +5233,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   i), {n_embd_k_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   i), {n_embd_v_gqa_i}, TENSOR_NOT_REQUIRED);
                             layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd},         TENSOR_NOT_REQUIRED);
-                        } else {
-                            // mlp layers
-                            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
-                            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
-                            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
-                            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
+                        }  else {
+                            if (n_expert != 0) {
+                                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                                // MoE branch
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                                // Shared expert branch
+                                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
+                            } else {
+                                // mlp layers
+                                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
+                                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
+                                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+                                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
+                            }
                         }
                     }
                 } break;
@@ -6627,9 +6635,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
         std::vector<ggml_backend_buffer_ptr> bufs;
         if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) {
+            GGML_ASSERT(!ml.no_alloc);
             for (uint32_t idx = 0; idx < ml.files.size(); idx++) {
                 // only the mmap region containing the tensors in the model is mapped to the backend buffer
-                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer, then we could just use metal for all layers
+                // this is important for metal with apple silicon: if the entire model could be mapped to a metal buffer,
+                //     then we could just use metal for all layers
                 // this allows using partial offloading when the model size exceeds the metal buffer size, but not the RAM size
                 void * addr = nullptr;
                 size_t first, last; // NOLINT
@@ -6645,9 +6655,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 bufs.emplace_back(buf);
                 buf_map.emplace(idx, buf);
             }
-        }
-        else {
-            ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft);
+        } else {
+            ggml_backend_buffer_t buf;
+            if (ml.no_alloc) {
+                buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+                for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+                    t->buffer = buf; // set dummy buffer for weights so that the backend scheduler won't try to allocate them
+                }
+            } else {
+                buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); // real buffer
+            }
             if (buf == nullptr) {
                 throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft)));
             }
@@ -6702,6 +6719,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
         }
     }
 
+    if (ml.no_alloc) {
+        return true;
+    }
+
     // load tensor data
     for (auto & [ctx, buf_map] : ctx_buf_maps) {
         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
@@ -6744,9 +6765,18 @@ size_t llama_model::n_devices() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_model::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, bufs] : pimpl->ctxs_bufs) {
-        for (const auto & buf : bufs) {
-            ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, bufs] : pimpl->ctxs_bufs) {
+        if (hparams.no_alloc) {
+            GGML_ASSERT(bufs.size() == 1);
+            ggml_backend_buffer_t buf = bufs[0].get();
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf) == nullptr);
+            ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            for (const auto & buf : bufs) {
+                // GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
+                ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+            }
         }
     }
     return ret;
@@ -6791,6 +6821,7 @@ void llama_model::print_info() const {
     // hparams
     LLAMA_LOG_INFO("%s: arch             = %s\n",     __func__, arch_name().c_str());
     LLAMA_LOG_INFO("%s: vocab_only       = %d\n",     __func__, hparams.vocab_only);
+    LLAMA_LOG_INFO("%s: no_alloc         = %d\n",     __func__, hparams.no_alloc);
 
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
@@ -6848,7 +6879,8 @@ void llama_model::print_info() const {
         arch == LLM_ARCH_PLAMO2 ||
         arch == LLM_ARCH_GRANITE_HYBRID ||
         arch == LLM_ARCH_QWEN3NEXT ||
-        arch == LLM_ARCH_NEMOTRON_H) {
+        arch == LLM_ARCH_NEMOTRON_H ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
@@ -6903,7 +6935,8 @@ void llama_model::print_info() const {
     if (arch == LLM_ARCH_MINICPM ||
         arch == LLM_ARCH_GRANITE ||
         arch == LLM_ARCH_GRANITE_MOE ||
-        arch == LLM_ARCH_GRANITE_HYBRID) {
+        arch == LLM_ARCH_GRANITE_HYBRID ||
+        arch == LLM_ARCH_NEMOTRON_H_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7084,7 +7117,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](int32_t) { return true; };
                         filter_recr = [&](int32_t) { return true; };
-                    } else if (arch == LLM_ARCH_NEMOTRON_H) {
+                    } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
                         filter_attn = [&](int32_t il) {
                             return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0;
                         };
@@ -7455,6 +7488,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
             } break;
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             {
                 llm = std::make_unique<llm_build_nemotron_h>(*this, params);
             } break;
@@ -7639,6 +7673,7 @@ llama_model_params llama_model_default_params() {
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
         /*.no_host                     =*/ false,
+        /*.no_alloc                    =*/ false,
     };
 
     return result;
@@ -7738,6 +7773,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARWKV7:
         case LLM_ARCH_WAVTOKENIZER_DEC:
         case LLM_ARCH_NEMOTRON_H:
+        case LLM_ARCH_NEMOTRON_H_MOE:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
@@ -7758,7 +7794,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
-        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_GRANITE_HYBRID:
@@ -7820,7 +7855,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_SMALLTHINKER:
-        case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_SEED_OSS:
         case LLM_ARCH_GROVEMOE:
         case LLM_ARCH_APERTUS:
@@ -7837,6 +7871,11 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3VLMOE:
             return LLAMA_ROPE_TYPE_IMROPE;
 
+        case LLM_ARCH_GLM4:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NORM;
+        case LLM_ARCH_GLM4_MOE:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NEOX;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");

src/llama-model.h

@@ -113,6 +113,7 @@ enum llm_type {
     LLM_TYPE_16B_A1B,
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
+    LLM_TYPE_31B_A3_5B,
     LLM_TYPE_80B_A3B, // Qwen3 Next
     LLM_TYPE_100B_A6B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air

src/llama-quant.cpp

@@ -596,7 +596,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());

src/llama-vocab.cpp

@@ -1895,7 +1895,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 clean_spaces = false;
             } else if (
                     tokenizer_pre == "qwen2" ||
-                    tokenizer_pre == "deepseek-r1-qwen") {
+                    tokenizer_pre == "deepseek-r1-qwen" ||
+                    tokenizer_pre == "kormo") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
                 clean_spaces = false;
             } else if (

src/llama.cpp

@@ -1,6 +1,9 @@
+#include "llama.h"
+
 #include "llama-impl.h"
 
 #include "llama-chat.h"
+#include "llama-context.h"
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
@@ -11,11 +14,14 @@
 #include "ggml-backend.h"
 
 #include <algorithm>
+#include <cassert>
+#include <cinttypes>
 #include <cstddef>
 #include <cstdint>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <stdexcept>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -37,6 +43,643 @@ const char * llama_flash_attn_type_name(enum llama_flash_attn_type flash_attn_ty
     GGML_ABORT("fatal error");
 }
 
+struct llama_device_memory_data {
+    int64_t total;
+    int64_t free;
+    llama_memory_breakdown_data mb;
+};
+
+static std::vector<llama_device_memory_data> llama_get_device_memory_data(
+        const char * path_model, const llama_model_params * mparams, const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs, uint32_t & hp_ngl, uint32_t & hp_n_ctx_train, uint32_t & hp_n_expert,
+        const ggml_log_level log_level) {
+    struct user_data_t {
+        struct {
+            ggml_log_callback callback;
+            void * user_data;
+        } original_logger;
+        ggml_log_level min_level; // prints below this log level go to debug log
+    };
+    user_data_t ud;
+    llama_log_get(&ud.original_logger.callback, &ud.original_logger.user_data);
+    ud.min_level = log_level;
+
+    llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
+        const user_data_t * ud = (const user_data_t *) user_data;
+        const ggml_log_level level_eff = level >= ud->min_level ? level : GGML_LOG_LEVEL_DEBUG;
+        ud->original_logger.callback(level_eff, text, ud->original_logger.user_data);
+    }, &ud);
+
+    llama_model_params mparams_copy = *mparams;
+    mparams_copy.no_alloc = true;
+    mparams_copy.use_mmap = false;
+
+    llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
+    if (model == nullptr) {
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to load model");
+    }
+
+    llama_context * ctx = llama_init_from_model(model, *cparams);
+    if (ctx == nullptr) {
+        llama_model_free(model);
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to create llama_context from model");
+    }
+
+    std::vector<llama_device_memory_data> ret(model->devices.size());
+
+    std::map<ggml_backend_buffer_type_t, llama_memory_breakdown_data> memory_breakdown = ctx->memory_breakdown();
+
+    for (const auto & [buft, mb] : memory_breakdown) {
+        if (ggml_backend_buft_is_host(buft)) {
+            continue;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            continue;
+        }
+        for (size_t i = 0; i < ret.size(); i++) {
+            if (model->devices[i] == dev) {
+                ret[i].mb.model   += mb.model;
+                ret[i].mb.context += mb.context;
+                ret[i].mb.compute += mb.compute;
+                break;
+            }
+        }
+    }
+    for (size_t i = 0; i < ret.size(); i++) {
+        size_t free, total;
+        ggml_backend_dev_memory(model->devices[i], &free, &total);
+        ret[i].free  = free;
+        ret[i].total = total;
+    }
+
+    devs           = model->devices;
+    hp_ngl         = model->hparams.n_layer;
+    hp_n_ctx_train = model->hparams.n_ctx_train;
+    hp_n_expert    = model->hparams.n_expert;
+
+    llama_memory_breakdown_print(ctx); // goes to debug log
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+    return ret;
+}
+
+// enum to identify part of a layer for distributing its tensors:
+enum layer_fraction_t {
+    LAYER_FRACTION_NONE = 0, // nothing
+    LAYER_FRACTION_ATTN = 1, // attention
+    LAYER_FRACTION_UP   = 2, // attention + up
+    LAYER_FRACTION_GATE = 3, // attention + up + gate
+    LAYER_FRACTION_MOE  = 4, // everything but sparse MoE weights
+};
+// this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
+
+static void llama_params_fit_impl(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    constexpr int64_t MiB = 1024*1024;
+    const int64_t margin = margin_s; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    typedef std::vector<llama_device_memory_data> dmds_t;
+    const llama_model_params default_mparams = llama_model_default_params();
+
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    // step 1: get data for default parameters and check whether any changes are necessary in the first place
+
+    LLAMA_LOG_DEBUG("%s: getting device memory data for initial parameters:\n", __func__);
+    const dmds_t dmds_full = llama_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const size_t nd = devs.size(); // number of devices
+    if (nd == 0) {
+        LLAMA_LOG_INFO("%s: no devices with dedicated memory found\n", __func__);
+        return;
+    }
+
+    std::vector<std::string> dev_names;
+    {
+        dev_names.reserve(nd);
+        size_t max_length = 0;
+        for (ggml_backend_dev_t dev : devs) {
+            std::string name = ggml_backend_dev_name(dev);
+            name += " (";
+            name += ggml_backend_dev_description(dev);
+            name += ")";
+            dev_names.push_back(name);
+            max_length = std::max(max_length, name.length());
+        }
+        for (std::string & dn : dev_names) {
+            dn.insert(dn.end(), max_length - dn.length(), ' ');
+        }
+    }
+
+    int64_t sum_total          = 0;
+    int64_t sum_projected_free = 0;
+    int64_t min_projected_free = INT64_MAX;
+    int64_t sum_projected_used = 0;
+    int64_t sum_projected_ctx  = 0;
+
+    if (nd > 1) {
+        LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
+    }
+    for (size_t id = 0; id < nd; id++) {
+        const llama_device_memory_data & dmd = dmds_full[id];
+
+        const int64_t projected_used = dmd.mb.total();
+        const int64_t projected_free = dmd.free - projected_used;
+
+        sum_total          += dmd.total;
+        sum_projected_used += projected_used;
+        sum_projected_free += projected_free;
+        min_projected_free  = std::min(min_projected_free, projected_free);
+        sum_projected_ctx  += dmd.mb.context;
+
+        if (nd > 1) {
+            LLAMA_LOG_INFO("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
+                __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, std::abs(projected_free)/MiB,
+                projected_free >= 0 ? "surplus" : "deficit");
+        }
+    }
+    assert(sum_total >= 0 && sum_projected_used >= 0 && sum_projected_ctx >= 0);
+    assert(sum_projected_used >= sum_projected_ctx);
+    LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
+        __func__, sum_projected_used/MiB, sum_total/MiB);
+    if (min_projected_free >= margin) {
+        if (nd == 1) {
+            LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
+                __func__, min_projected_free/MiB, margin/MiB);
+            return;
+        }
+        LLAMA_LOG_INFO("%s: will leave at least %" PRId64 " >= %" PRId64 " MiB of free memory on all devices, no changes needed\n",
+            __func__, min_projected_free/MiB, margin/MiB);
+        return;
+    }
+
+    // step 2: try reducing memory use by reducing the context size
+
+    {
+        int64_t global_surplus = sum_projected_free - int64_t(nd)*margin;
+        if (global_surplus < 0) {
+            LLAMA_LOG_INFO(nd == 1 ?
+                "%s: cannot fulfill margin of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n" :
+                "%s: cannot fulfill margin of %" PRId64 " MiB on all devices, need to use %" PRId64 " MiB less in total\n",
+                __func__, margin/MiB, -global_surplus/MiB);
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
+                    const int64_t bytes_per_ctx = sum_projected_ctx / hp_nct;
+                    const uint32_t ctx_reduction = std::min(
+                        uint32_t((-global_surplus + bytes_per_ctx - 1) / bytes_per_ctx), hp_nct - n_ctx_min);
+                    cparams->n_ctx = hp_nct - ctx_reduction;
+                    const int64_t memory_reduction = ctx_reduction * bytes_per_ctx;
+                    global_surplus += memory_reduction;
+                    LLAMA_LOG_INFO("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                        __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                } else {
+                    LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                        __func__, hp_nct, n_ctx_min);
+                }
+            } else {
+                LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
+            }
+        }
+        if (global_surplus >= 0) {
+            LLAMA_LOG_INFO("%s: entire model can be fit across devices by reducing context\n", __func__);
+            return;
+        }
+    }
+
+    if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
+        throw std::runtime_error("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+    }
+    if (nd > 1) {
+        if (!tensor_split) {
+            throw std::runtime_error("did not provide a buffer to write the tensor_split to, abort");
+        }
+        if (mparams->tensor_split) {
+            for (size_t id = 0; id < nd; id++) {
+                if (mparams->tensor_split[id] != 0.0f) {
+                    throw std::runtime_error("model_params::tensor_split already set by user, abort");
+                }
+            }
+        }
+        if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            throw std::runtime_error("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+        }
+        if (hp_ngl < 2*nd) {
+            throw std::runtime_error("model has only " + std::to_string(hp_ngl) + " layers but need at least "
+                + std::to_string(2*nd) + " to fit memory for " + std::to_string(nd) + " devices, abort");
+        }
+    }
+    if (!tensor_buft_overrides) {
+        throw std::runtime_error("did not provide buffer to set tensor_buft_overrides, abort");
+    }
+    if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
+        throw std::runtime_error("model_params::tensor_buft_overrides already set by user, abort");
+    }
+
+    // step 3: iteratively fill the back to front with "dense" layers
+    //   - for a dense model simply fill full layers, giving each device a contiguous slice of the model
+    //   - for a MoE model, same as dense model but with all MoE tensors in system memory
+
+    // utility function that returns a static C string matching the tensors for a specific layer index and layer fraction:
+    auto get_overflow_pattern = [&](const size_t il, const layer_fraction_t lf) -> const char * {
+        constexpr size_t n_strings = 1000;
+        if (il >= n_strings) {
+            throw std::runtime_error("at most " + std::to_string(n_strings) + " model layers are supported");
+        }
+        switch (lf) {
+            case LAYER_FRACTION_ATTN: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_UP: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_GATE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_down.*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_MOE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate)_(ch|)exps";
+                }
+                return patterns[il].c_str();
+            }
+            default:
+                GGML_ABORT("fatal error");
+        }
+    };
+
+    struct ngl_t {
+        uint32_t n_layer = 0; // number of total layers
+        uint32_t n_part  = 0; // number of partial layers, <= n_layer
+
+        // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
+        layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+    };
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+
+    // utility function to set n_gpu_layers and tensor_split
+    auto set_ngl_tensor_split_tbo = [&](
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            llama_model_params & mparams,
+            const bool add_nonrepeating) {
+        mparams.n_gpu_layers = 0;
+        for (size_t id = 0; id < nd; id++) {
+            mparams.n_gpu_layers += ngl_per_device[id].n_layer;
+            if (nd > 1) {
+                tensor_split[id] = ngl_per_device[id].n_layer;
+            }
+        }
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl);
+        uint32_t il0 = hp_ngl - mparams.n_gpu_layers; // start index for tensor buft overrides
+
+        if (add_nonrepeating) {
+            mparams.n_gpu_layers += 1;
+            tensor_split[nd - 1] += 1;
+        }
+        mparams.tensor_split = tensor_split;
+
+        size_t itbo = 0;
+        for (size_t id = 0; id < nd; id++) {
+            il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
+            for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
+                if (itbo + 1 >= ntbo) {
+                    tensor_buft_overrides[itbo].pattern = nullptr;
+                    tensor_buft_overrides[itbo].buft    = nullptr;
+                    itbo++;
+                    mparams.tensor_buft_overrides = tensor_buft_overrides;
+                    throw std::runtime_error("llama_params_fit_n_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model\n");
+                }
+                tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
+                tensor_buft_overrides[itbo].buft = overflow_bufts[id];
+                itbo++;
+            }
+            il0 += ngl_per_device[id].n_part;
+        }
+        tensor_buft_overrides[itbo].pattern = nullptr;
+        tensor_buft_overrides[itbo].buft    = nullptr;
+        itbo++;
+        mparams.tensor_buft_overrides = tensor_buft_overrides;
+    };
+
+    // utility function that returns the memory use per device for given numbers of layers per device
+    auto get_memory_for_layers = [&](
+            const char * func_name,
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            const bool add_nonrepeating) -> std::vector<int64_t> {
+        llama_model_params mparams_copy = *mparams;
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy, add_nonrepeating);
+
+        const dmds_t dmd_nl = llama_get_device_memory_data(
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        LLAMA_LOG_DEBUG("%s: memory for test allocation by device:\n", func_name);
+        for (size_t id = 0; id < nd; id++) {
+            const ngl_t & n = ngl_per_device[id];
+            LLAMA_LOG_DEBUG(
+                "%s: id=%zu, n_layer=%2" PRIu32 ", n_part=%2" PRIu32 ", overflow_type=%d, mem=%6" PRId64 " MiB\n",
+                func_name, id, n.n_layer, n.n_part, int(n.overflow_type), dmd_nl[id].mb.total()/MiB);
+        }
+
+        std::vector<int64_t> ret;
+        ret.reserve(nd);
+        for (const llama_device_memory_data & dmd : dmd_nl) {
+            ret.push_back(dmd.mb.total());
+        }
+        return ret;
+    };
+
+    int64_t global_surplus_cpu_moe = 0;
+    if (hp_nex > 0) {
+        const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate)_(ch|)exps"; // matches all MoE tensors
+        ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
+        tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
+        tensor_buft_overrides[1] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+
+        LLAMA_LOG_DEBUG("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
+        const dmds_t dmds_cpu_moe = llama_get_device_memory_data(
+            path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        for (const llama_device_memory_data & dmd : dmds_cpu_moe) {
+            global_surplus_cpu_moe += dmd.free;
+            global_surplus_cpu_moe -= int64_t(dmd.mb.total()) + margin;
+        }
+
+        if (global_surplus_cpu_moe > 0) {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is a total surplus of %" PRId64 " MiB\n",
+                __func__, global_surplus_cpu_moe/MiB);
+        } else {
+            LLAMA_LOG_INFO("%s: with only dense weights in device memory there is still a total deficit of %" PRId64 " MiB\n",
+                __func__, -global_surplus_cpu_moe/MiB);
+        }
+
+        // reset
+        tensor_buft_overrides[0] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+    }
+
+    std::vector<int64_t> targets; // maximum acceptable memory use per device
+    targets.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        targets.push_back(dmds_full[id].free - margin);
+        LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
+    }
+
+    // whether for the optimal memory use we expect to load at least some MoE tensors:
+    const bool partial_moe = hp_nex > 0 && global_surplus_cpu_moe > 0;
+
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
+    overflow_bufts.reserve(nd);
+    for (size_t id = 0; id < nd - 1; ++id) {
+        overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
+    }
+    overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
+
+    std::vector<ngl_t> ngl_per_device(nd);
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts, partial_moe);
+    if (hp_nex > 0) {
+        for (size_t id = 0; id < nd; id++) {
+            ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
+        }
+    }
+
+    // optimize the number of layers per device using the method of false position:
+    //   - ngl_per_device has 0 layers for each device, lower bound
+    //   - try a "high" configuration where a device is given all unassigned layers
+    //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
+    //   - check memory use of our guess, replace either the low or high bound
+    //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    if (hp_nex == 0) {
+        LLAMA_LOG_INFO("%s: filling dense layers back-to-front:\n", __func__);
+    } else {
+        LLAMA_LOG_INFO("%s: filling dense-only layers back-to-front:\n", __func__);
+    }
+    uint32_t n_unassigned = hp_ngl;
+    for (int id = nd - 1; id >= 0; id--) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        ngl_per_device_high[id].n_layer = n_unassigned;
+        if (hp_nex > 0) {
+            ngl_per_device_high[id].n_part = ngl_per_device_high[id].n_layer;
+        }
+        if (ngl_per_device_high[id].n_layer > 0) {
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+            if (mem_high[id] > targets[id]) {
+                uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                while (delta > 1) {
+                    uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                    step_size = std::max(step_size, uint32_t(1));
+                    step_size = std::min(step_size, delta - 1);
+
+                    std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                    ngl_per_device_test[id].n_layer += step_size;
+                    if (hp_nex) {
+                        ngl_per_device_test[id].n_part += step_size;
+                    }
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                    if (mem_test[id] <= targets[id]) {
+                        ngl_per_device  = ngl_per_device_test;
+                        mem             = mem_test;
+                        n_unassigned   -= ngl_per_device[id].n_layer;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    } else {
+                        ngl_per_device_high = ngl_per_device_test;
+                        mem_high            = mem_test;
+                        LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    }
+                    delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                }
+            } else {
+                ngl_per_device  = ngl_per_device_high;
+                n_unassigned   -= ngl_per_device[id].n_layer;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers, %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
+    }
+    if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+        return;
+    }
+
+    // step 4: for a MoE model where all dense tensors fit,
+    //     convert the dense-only layers in the back to full layers in the front until all devices are full
+    // essentially the same procedure as for the dense-only layers except front-to-back
+    // also, try fitting at least part of one more layer to reduce waste for "small" GPUs with e.g. 24 GiB VRAM
+
+    size_t id_dense_start = nd;
+    for (int id = nd - 1; id >= 0; id--) {
+        if (ngl_per_device[id].n_layer > 0) {
+            id_dense_start = id;
+            continue;
+        }
+        break;
+    }
+    assert(id_dense_start < nd);
+
+    LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
+    for (size_t id = 0; id <= id_dense_start; id++) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        for (size_t jd = id_dense_start; jd < nd; jd++) {
+            const uint32_t n_layer_move = ngl_per_device_high[jd].n_layer;
+            ngl_per_device_high[id].n_layer += n_layer_move;
+            ngl_per_device_high[jd].n_layer -= n_layer_move;
+            ngl_per_device_high[jd].n_part = 0;
+        }
+        size_t id_dense_start_high = nd - 1;
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts, partial_moe);
+
+        if (mem_high[id] > targets[id]) {
+            assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
+            assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
+            assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                   >= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            while (delta > 1) {
+                uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                step_size = std::max(step_size, uint32_t(1));
+                step_size = std::min(step_size, delta - 1);
+
+                std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                size_t id_dense_start_test = id_dense_start;
+                uint32_t n_converted_test = 0;
+                for (;id_dense_start_test < nd; id_dense_start_test++) {
+                    const uint32_t n_convert_jd = std::min(step_size - n_converted_test, ngl_per_device_test[id_dense_start_test].n_part);
+                    ngl_per_device_test[id_dense_start_test].n_layer -= n_convert_jd;
+                    ngl_per_device_test[id_dense_start_test].n_part -= n_convert_jd;
+                    ngl_per_device_test[id].n_layer += n_convert_jd;
+                    n_converted_test += n_convert_jd;
+
+                    if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
+                        break;
+                    }
+                }
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+
+                if (mem_test[id] <= targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                } else {
+                    ngl_per_device_high = ngl_per_device_test;
+                    mem_high            = mem_test;
+                    id_dense_start_high = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
+                        __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
+                }
+                delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
+                    - (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
+            }
+        } else {
+            ngl_per_device = ngl_per_device_high;
+            id_dense_start = id_dense_start_high;
+            LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+        }
+
+        // try to fit at least part of one more layer
+        if (ngl_per_device[id_dense_start].n_layer > 0) {
+            std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+            size_t id_dense_start_test = id_dense_start;
+            ngl_per_device_test[id_dense_start_test].n_layer--;
+            ngl_per_device_test[id_dense_start_test].n_part--;
+            ngl_per_device_test[id].n_layer++;
+            ngl_per_device_test[id].n_part++;
+            if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
+                id_dense_start_test++;
+            }
+            ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+            if (mem_test[id] < targets[id]) {
+                ngl_per_device = ngl_per_device_test;
+                mem            = mem_test;
+                id_dense_start = id_dense_start_test;
+                LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
+                    __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            } else {
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
+                LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts, partial_moe);
+                if (mem_test[id] < targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LLAMA_LOG_INFO(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams, partial_moe);
+}
+
+bool llama_params_fit(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    const int64_t t0_us = llama_time_us();
+    bool ok = true;
+    try {
+        llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
+        LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
+    } catch (const std::runtime_error & e) {
+        LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
+        ok = false;
+    }
+    const int64_t t1_us = llama_time_us();
+    LLAMA_LOG_INFO("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
+    return ok;
+}
+
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
         /*.no_perf                     =*/ true,
@@ -49,6 +692,10 @@ size_t llama_max_devices(void) {
     return 16;
 }
 
+size_t llama_max_tensor_buft_overrides() {
+    return 4096;
+}
+
 bool llama_supports_mmap(void) {
     return llama_mmap::SUPPORTED;
 }
@@ -108,11 +755,12 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
 
         ml.print_info();
 
         model.hparams.vocab_only = params.vocab_only;
+        model.hparams.no_alloc   = params.no_alloc;
 
         try {
             model.load_arch(ml);

src/models/glm4-moe.cpp

@@ -5,11 +5,20 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
     inpL = build_inp_embd(model.tok_embd);
 
+    bool use_mrope = hparams.use_mrope();
+    if (ubatch.embd && !use_mrope) {
+        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
+        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -60,17 +69,25 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
                 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
                 cb(Kcur, "Kcur_normed", il);
             }
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
 
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
+            if (use_mrope) {
+                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                // Normal RoPE
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+            }
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);