From ea9c32be71b91b42ecc538bd902e93cbb5fb36cb Mon Sep 17 00:00:00 2001
From: Xuan Son Nguyen <thichthat@gmail.com>
Date: Wed, 25 Sep 2024 17:26:01 +0200
Subject: [PATCH 01/11] ci : fix docker build number and tag name (#9638)

* ci : fix docker build number and tag name

* fine-grant permissions
---
 .dockerignore                |  2 +-
 .github/workflows/docker.yml | 58 +++++++++++++++++++++++-------------
 2 files changed, 39 insertions(+), 21 deletions(-)

diff --git a/.dockerignore b/.dockerignore
index 8916e2a66..064b7c7be 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,7 +1,7 @@
 *.o
 *.a
 .cache/
-.git/
+# Do not ignore .git directory, otherwise the reported build number will always be 0
 .github/
 .gitignore
 .vs/
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index 9044cd78b..a4ac9b217 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -15,11 +15,17 @@ on:
     branches:
       - master
     paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  workflow_dispatch: # allows manual triggering, useful for debugging
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  packages: write
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -46,6 +52,8 @@ jobs:
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -60,6 +68,34 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+
+          # determine tag name postfix (build number, commit hash)
+          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
+            TAG_POSTFIX="b${BUILD_NUMBER}"
+          else
+            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
+            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
+          fi
+
+          # list all tags possible
+          TAGS=""
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
+
+          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
+          echo "output_tags=$TAGS"  # print out for debugging
+        env:
+          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main
@@ -77,25 +113,6 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Downcase github.repository_owner
-        run: |
-          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
       - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v6
@@ -103,5 +120,6 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          # tag list is generated from step above
+          tags: ${{ steps.tag.outputs.output_tags }}
           file: ${{ matrix.config.dockerfile }}

From 7691654c68aa5316108f881136c59f815ccb6809 Mon Sep 17 00:00:00 2001
From: R0CKSTAR <xiaodong.ye@mthreads.com>
Date: Thu, 26 Sep 2024 09:27:40 +0800
Subject: [PATCH 02/11] mtgpu: enable VMM (#9597)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
---
 ggml/src/ggml-cuda.cu | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index 0bb7f2d99..6efdab14c 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -187,7 +187,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -199,7 +199,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -335,7 +335,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -429,14 +429,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
     }
 };
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 

From 95bc82fbc0df6d48cf66c857a4dda3d044f45ca2 Mon Sep 17 00:00:00 2001
From: Neo Zhang Jianyu <jianyu.zhang@intel.com>
Date: Thu, 26 Sep 2024 17:38:31 +0800
Subject: [PATCH 03/11] [SYCL] add missed dll file in package (#9577)

* update oneapi to 2024.2

* use 2024.1

---------

Co-authored-by: arthw <14088817+arthw@users.noreply.github.com>
---
 .github/workflows/build.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index a54c5de99..e6a977b60 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -956,6 +956,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 

From 44f59b4301c51f071daa2e951301bb17c14acc9b Mon Sep 17 00:00:00 2001
From: Borislav Stanimirov <b.stanimirov@abv.bg>
Date: Fri, 27 Sep 2024 10:42:06 +0300
Subject: [PATCH 04/11] cmake : add option for common library (#9661)

---
 CMakeLists.txt | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 973907819..415743c2a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -62,6 +62,9 @@ option(LLAMA_SANITIZE_THREAD    "llama: enable thread sanitizer"    OFF)
 option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
 option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
+# utils
+option(LLAMA_BUILD_COMMON "llama: build common utils library" ON)
+
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
@@ -191,15 +194,17 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
 
 #
-# programs, examples and tests
+# utils, programs, examples and tests
 #
 
-add_subdirectory(common)
+if (LLAMA_BUILD_COMMON)
+    add_subdirectory(common)
+endif()
 
 if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     include(CTest)
     add_subdirectory(tests)
-endif ()
+endif()
 
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)

From b5de3b74a595cbfefab7eeb5a567425c6a9690cf Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Fri, 27 Sep 2024 20:57:51 +0300
Subject: [PATCH 05/11] readme : update hot topics

---
 README.md | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index ce954f713..93225b63c 100644
--- a/README.md
+++ b/README.md
@@ -17,7 +17,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
+- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
 

From 89f9944981010d195e411a9fbfbb19959412f710 Mon Sep 17 00:00:00 2001
From: Markus Tavenrath <mtavenrath@users.noreply.github.com>
Date: Sat, 28 Sep 2024 12:05:05 +0200
Subject: [PATCH 06/11] Enable use to the rebar feature to upload buffers to
 the device. (#9251)

---
 ggml/src/ggml-vulkan.cpp | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan.cpp
index f9da45881..a877145e8 100644
--- a/ggml/src/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan.cpp
@@ -1079,7 +1079,8 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
             // Fall back to host memory type
             buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
         } else {
-            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal);
+            // use rebar if available, otherwise fallback to device only visible memory
+            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
         }
     } catch (const vk::SystemError& e) {
         std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
@@ -2806,7 +2807,11 @@ static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t
 
 static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) {
     VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")");
-    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+
+    // If the device is not an UMA device the memory is host-accessible through rebar. While writing
+    // through PCIe is sufficient fast reading back data from PCIe is slower than going through
+    // the HW device to host copy path.
+    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
         GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
         memcpy(dst, (uint8_t *) src->ptr + offset, size);

From 6a0f7794847244fb3b99a983e03137d7e832b585 Mon Sep 17 00:00:00 2001
From: Dan Johansson <164997844+eddnjjn@users.noreply.github.com>
Date: Sat, 28 Sep 2024 14:06:16 +0200
Subject: [PATCH 07/11] ggml : add run-time detection of neon, i8mm and sve
 (#9331)

* ggml: Added run-time detection of neon, i8mm and sve

Adds run-time detection of the Arm instructions set features
neon, i8mm and sve for Linux and Apple build targets.

* ggml: Extend feature detection to include non aarch64 Arm arch

* ggml: Move definition of ggml_arm_arch_features to the global data section
---
 ggml/include/ggml.h     |   3 ++
 ggml/src/ggml-aarch64.c |  13 +-----
 ggml/src/ggml-quants.c  |   4 +-
 ggml/src/ggml-quants.h  |   4 --
 ggml/src/ggml.c         | 101 ++++++++++++++++++++++++++++++++++------
 5 files changed, 93 insertions(+), 32 deletions(-)

diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index e24b8a319..9f96e0c48 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -2507,6 +2507,9 @@ extern "C" {
     GGML_API int ggml_cpu_has_cann       (void);
     GGML_API int ggml_cpu_has_llamafile  (void);
 
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
+
     //
     // Internal types and functions exposed for tests and benchmarks
     //
diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c
index 8912de63d..b27f41147 100644
--- a/ggml/src/ggml-aarch64.c
+++ b/ggml/src/ggml-aarch64.c
@@ -598,15 +598,6 @@ size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_
     return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
 }
 
-// Return the number of byte lanes in the SVE vector if SVE is supported; otherwise, returns 0 if SVE is not supported.
-static int sve_lane_count(void) {
-#if defined(__ARM_FEATURE_SVE)
-    return ggml_sve_cnt_b;
-#else
-    return 0;
-#endif
-}
-
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -843,7 +834,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE)
-    if (ggml_cpu_has_sve() && sve_lane_count() == QK8_0) {
+    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -2020,7 +2011,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
 #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && sve_lane_count() == QK8_0) {
+    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 8bffce860..7aa6dce89 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -4013,7 +4013,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     svfloat32_t sumv0 = svdup_n_f32(0.0f);
     svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    const int vector_length = ggml_sve_cnt_b*8;
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
 
     // VLA Implementation using switch case
     switch (vector_length) {
@@ -5597,7 +5597,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
     svfloat32_t sumv0 = svdup_n_f32(0.0f);
     svfloat32_t sumv1 = svdup_n_f32(0.0f);
 
-    const int vector_length = ggml_sve_cnt_b*8;
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
 
     //VLA Implemenation for SVE
     switch (vector_length) {
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h
index e96ce2b5e..df9c4b24a 100644
--- a/ggml/src/ggml-quants.h
+++ b/ggml/src/ggml-quants.h
@@ -142,10 +142,6 @@ void iq2xs_free_impl(enum ggml_type type);
 void iq3xs_init_impl(int grid_size);
 void iq3xs_free_impl(int grid_size);
 
-#if defined(__ARM_FEATURE_SVE)
-extern int ggml_sve_cnt_b;
-#endif
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 4b782b0c1..fac4466e3 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -39,9 +39,6 @@
 #include <unistd.h>
 #endif
 
-#if defined(__ARM_FEATURE_SVE)
-int ggml_sve_cnt_b = 0;
-#endif
 #if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
 #undef GGML_USE_LLAMAFILE
 #endif
@@ -455,6 +452,15 @@ static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];
 
+#if defined(__ARM_ARCH)
+struct ggml_arm_arch_features_type {
+    int has_neon;
+    int has_i8mm;
+    int has_sve;
+    int sve_cnt;
+} ggml_arm_arch_features = {-1, -1, -1, 0};
+#endif
+
 GGML_CALL const char * ggml_status_to_string(enum ggml_status status) {
     switch (status) {
         case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
@@ -3673,6 +3679,66 @@ static inline int ggml_up(int n, int m) {
 
 ////////////////////////////////////////////////////////////////////////////////
 
+#if defined(__ARM_ARCH)
+
+#if defined(__linux__) && defined(__aarch64__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+static void ggml_init_arm_arch_features(void) {
+#if defined(__linux__) && defined(__aarch64__)
+    uint32_t hwcap = getauxval(AT_HWCAP);
+    uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+    ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#endif
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_neon = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_i8mm = oldp;
+
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#else
+// Run-time CPU feature detection not implemented for this platform, fallback to compile time
+#if defined(__ARM_NEON)
+    ggml_arm_arch_features.has_neon = 1;
+#else
+    ggml_arm_arch_features.has_neon = 0;
+#endif
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_arm_arch_features.has_i8mm = 1;
+#else
+    ggml_arm_arch_features.has_i8mm = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.has_sve = 1;
+    ggml_arm_arch_features.sve_cnt = 16;
+#else
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#endif
+#endif
+}
+#endif
+
 struct ggml_context * ggml_init(struct ggml_init_params params) {
     // make this function thread safe
     ggml_critical_section_start();
@@ -3723,6 +3789,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
+#if defined(__ARM_ARCH)
+        ggml_init_arm_arch_features();
+#endif
+
         is_first_call = false;
     }
 
@@ -3771,12 +3841,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_ASSERT_ALIGNED(ctx->mem_buffer);
 
-#if defined(__ARM_FEATURE_SVE)
-    if (!ggml_sve_cnt_b) {
-        ggml_sve_cnt_b = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-    }
-#endif
-
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
     ggml_critical_section_end();
@@ -23578,16 +23642,16 @@ int ggml_cpu_has_fma(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_NEON)
-    return 1;
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
 }
 
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_FEATURE_SVE)
-    return 1;
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_sve;
 #else
     return 0;
 #endif
@@ -23734,11 +23798,18 @@ int ggml_cpu_has_vsx(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    return 1;
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
 #endif
 }
 
+int ggml_cpu_get_sve_cnt(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.sve_cnt;
+#else
+    return 0;
+#endif
+}
 ////////////////////////////////////////////////////////////////////////////////

From 43bcdd9703ec19af7d2a519640b5ed6f4aac3d53 Mon Sep 17 00:00:00 2001
From: Aarni Koskela <akx@iki.fi>
Date: Sat, 28 Sep 2024 15:07:14 +0300
Subject: [PATCH 08/11] readme : add tool (#9655)

---
 README.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/README.md b/README.md
index 93225b63c..a452a6d78 100644
--- a/README.md
+++ b/README.md
@@ -174,6 +174,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
 - [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 - [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)

From 9a913110cf471a8287ac06c43cbe307d3cf6df99 Mon Sep 17 00:00:00 2001
From: nopperl <54780682+nopperl@users.noreply.github.com>
Date: Sat, 28 Sep 2024 12:08:43 +0000
Subject: [PATCH 09/11] llama : add support for Chameleon (#8543)

* convert chameleon hf to gguf

* add chameleon tokenizer tests

* fix lint

* implement chameleon graph

* add swin norm param

* return qk norm weights and biases to original format

* implement swin norm

* suppress image token output

* rem tabs

* add comment to conversion

* fix ci

* check for k norm separately

* adapt to new lora implementation

* fix layer input for swin norm

* move swin_norm in gguf writer

* add comment regarding special token regex in chameleon pre-tokenizer

* Update src/llama.cpp

Co-authored-by: compilade <git@compilade.net>

* fix punctuation regex in chameleon pre-tokenizer (@compilade)

Co-authored-by: compilade <git@compilade.net>

* fix lint

* trigger ci

---------

Co-authored-by: compilade <git@compilade.net>
---
 convert_hf_to_gguf.py                |  44 +++++
 convert_hf_to_gguf_update.py         |   1 +
 gguf-py/gguf/constants.py            |  19 ++
 gguf-py/gguf/gguf_writer.py          |   3 +
 gguf-py/gguf/tensor_mapping.py       |   4 +-
 include/llama.h                      |   1 +
 models/ggml-vocab-chameleon.gguf.inp | 112 ++++++++++++
 models/ggml-vocab-chameleon.gguf.out |  46 +++++
 src/llama-vocab.cpp                  |  14 ++
 src/llama.cpp                        | 263 +++++++++++++++++++++++++++
 10 files changed, 505 insertions(+), 2 deletions(-)
 create mode 100644 models/ggml-vocab-chameleon.gguf.inp
 create mode 100644 models/ggml-vocab-chameleon.gguf.out

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 7be609054..2cd5a8c11 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -640,6 +640,9 @@ class Model:
         if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
             # ref: https://huggingface.co/microsoft/phi-2
             res = "phi-2"
+        if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
+            # ref: https://huggingface.co/facebook/chameleon-7b
+            res = "chameleon"
 
         if res is None:
             logger.warning("\n")
@@ -4138,6 +4141,47 @@ class GraniteMoeModel(GraniteModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@Model.register("ChameleonForCausalLM")
+class ChameleonModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHAMELEON
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # ignore image tokenizer for now
+        # TODO: remove this once image support is implemented for Chameleon
+        if name.startswith("model.vqmodel"):
+            return []
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        hidden_dim = self.hparams.get("hidden_size")
+
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        if name.endswith(("q_norm.weight", "q_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
+        if name.endswith(("k_norm.weight", "k_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
+    @staticmethod
+    def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
+        head_dim = hidden_dim // n_heads
+        data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
+        data_torch = data_torch.repeat_interleave(n_heads, 0)
+        return data_torch
+
+
 ###### CONVERSION LOGIC ######
 
 
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index 021f65abd..4d11059f3 100755
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -99,6 +99,7 @@ models = [
     {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
     {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
     {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
+    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
 ]
 
 
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index 560eee916..2fd2e9d2b 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -94,6 +94,7 @@ class Keys:
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
         ATTN_LOGIT_SOFTCAPPING            = "{arch}.attn_logit_softcapping"
         FINAL_LOGIT_SOFTCAPPING           = "{arch}.final_logit_softcapping"
+        SWIN_NORM                         = "{arch}.swin_norm"
         RESCALE_EVERY_N_LAYERS            = "{arch}.rescale_every_n_layers"
         TIME_MIX_EXTRA_DIM                = "{arch}.time_mix_extra_dim"
         TIME_DECAY_EXTRA_DIM              = "{arch}.time_decay_extra_dim"
@@ -236,6 +237,7 @@ class MODEL_ARCH(IntEnum):
     EXAONE       = auto()
     GRANITE      = auto()
     GRANITE_MOE  = auto()
+    CHAMELEON    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -394,6 +396,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.EXAONE:         "exaone",
     MODEL_ARCH.GRANITE:        "granite",
     MODEL_ARCH.GRANITE_MOE:    "granitemoe",
+    MODEL_ARCH.CHAMELEON:      "chameleon",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -1260,6 +1263,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],
+    MODEL_ARCH.CHAMELEON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index bd059b45c..5c460ef1b 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -670,6 +670,9 @@ class GGUFWriter:
     def add_expert_weights_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.EXPERT_WEIGHTS_SCALE.format(arch=self.arch), value)
 
+    def add_swin_norm(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
+
     def add_rescale_every_n_layers(self, count: int) -> None:
         self.add_uint32(Keys.LLM.RESCALE_EVERY_N_LAYERS.format(arch=self.arch), count)
 
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 4e850726e..5ef91f11d 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -380,7 +380,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -389,7 +389,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
diff --git a/include/llama.h b/include/llama.h
index 132937a07..caef0bfff 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -102,6 +102,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BLOOM          = 23,
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
+        LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
     };
 
     enum llama_rope_type {
diff --git a/models/ggml-vocab-chameleon.gguf.inp b/models/ggml-vocab-chameleon.gguf.inp
new file mode 100644
index 000000000..9baf7d77a
--- /dev/null
+++ b/models/ggml-vocab-chameleon.gguf.inp
@@ -0,0 +1,112 @@
+ied 4 ½ months
+__ggml_vocab_test__
+Führer
+__ggml_vocab_test__
+
+__ggml_vocab_test__
+ 
+__ggml_vocab_test__
+  
+__ggml_vocab_test__
+   
+__ggml_vocab_test__
+	
+__ggml_vocab_test__
+
+
+__ggml_vocab_test__
+
+
+
+__ggml_vocab_test__
+
+
+
+
+__ggml_vocab_test__
+	
+
+__ggml_vocab_test__
+Hello world
+__ggml_vocab_test__
+ Hello world
+__ggml_vocab_test__
+Hello World
+__ggml_vocab_test__
+ Hello World
+__ggml_vocab_test__
+ Hello World!
+__ggml_vocab_test__
+Hello, world!
+__ggml_vocab_test__
+ Hello, world!
+__ggml_vocab_test__
+ this is 🦙.cpp
+__ggml_vocab_test__
+w048 7tuijk dsdfhu
+__ggml_vocab_test__
+нещо на Български
+__ggml_vocab_test__
+កាន់តែពិសេសអាចខលចេញ
+__ggml_vocab_test__
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
+__ggml_vocab_test__
+Hello
+__ggml_vocab_test__
+ Hello
+__ggml_vocab_test__
+  Hello
+__ggml_vocab_test__
+   Hello
+__ggml_vocab_test__
+    Hello
+__ggml_vocab_test__
+    Hello
+    Hello
+__ggml_vocab_test__
+ (
+__ggml_vocab_test__
+
+ =
+__ggml_vocab_test__
+' era
+__ggml_vocab_test__
+Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
+__ggml_vocab_test__
+!!!!!!
+__ggml_vocab_test__
+3
+__ggml_vocab_test__
+33
+__ggml_vocab_test__
+333
+__ggml_vocab_test__
+3333
+__ggml_vocab_test__
+33333
+__ggml_vocab_test__
+333333
+__ggml_vocab_test__
+3333333
+__ggml_vocab_test__
+33333333
+__ggml_vocab_test__
+333333333
+__ggml_vocab_test__
+Cửa Việt
+__ggml_vocab_test__
+ discards
+__ggml_vocab_test__
+
+ 
+
+ 
+
+
+ 	 		 	
+  
+   
+    
+     
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
+__ggml_vocab_test__
diff --git a/models/ggml-vocab-chameleon.gguf.out b/models/ggml-vocab-chameleon.gguf.out
new file mode 100644
index 000000000..7c5413fee
--- /dev/null
+++ b/models/ggml-vocab-chameleon.gguf.out
@@ -0,0 +1,46 @@
+ 17245 16604 16403 16604 33583 18355
+ 16421 51153
+
+ 16604
+ 16650
+ 16650 16604
+ 16581
+ 16582
+ 16582 16582
+ 16582 16582 16582
+ 16581 16582
+ 31596 17394
+ 34926 17394
+ 31596 18671
+ 34926 18671
+ 34926 18671 16384
+ 31596 16395 17394 16384
+ 34926 16395 17394 16384
+ 16811 16704 20410 16483 16631 16397 52854
+ 16470 16399 16403 16407 16604 16406 35764 38185 51595 22592 26639
+ 29479 23955 17012 20103 25527 27670 17408 19005 21473 24774
+ 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 21954 16607 21954 16633 21954 16611 29409 16607 21954 16615
+ 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 16604 16391 24664 17153 57169 16721 16872 17073 17304 28729 16392
+ 31596
+ 34926
+ 16650 31596
+ 16650 34926
+ 16696 31596
+ 16696 31596 16582 16696 31596
+ 16604 16391
+ 16582 16604 16412
+ 16390 22623
+ 31596 16395 16712 16390 16828 16384 17674 16769 16732 23686 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636
+ 16384 16384 16384 16384 16384 16384
+ 16402
+ 16402 16402
+ 16402 16402 16402
+ 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402 16402
+ 16402 16402 16402 16402 16402 16402 16402 16402 16402
+ 16418 19038 16639 16448 24315 33727 16467
+ 18765 17981
+ 16582 16604 16582 16582 16604 16582 16582 16582 16604 16581 16604 16581 16581 16604 16581 16582 16650 16582 16650 16604 16582 16696 16582 16696 16604 16582 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 20410 16483 16631 18885 16483 16631 16604 16402 16604 16402 16402 16604 16402 16402 16402 16604 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16402 16604 16402 16397 16402 16604 16402 16397 16397 16402 16604 16402 16397 16397 16397 16402 16604 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 27683 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636 16604 16396 16396 16396 16396 16396 16396 16412 16412 16412 16412 16412 16412 16412 27268 23955 17012 20103 25527 27670 17408 19005 21473 24774 16604 16390 16390 16390 16390 16390 16390 16447 16447 16447 16447 16447 16447 16447 16385 16385 16385 16385 16397 16397 16397 16397 16397 16397 16384 16384 16384 16384 16384 16384 16414 16414 16414 16414 16414 16414 16687 16390 16690 16992 16604 16390 61797 16733 16390 16466 16986 16395 16604 16390 17879 16732 17811 16414 16604 16390 16428 16804 17811 16687 16390 16683 17190 16728 16395 16604 16390 16419 16732 16945 16991 25251 16414 17119 16390 38127 16641 16390 16459 16427
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index a771eccda..146d416f7 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -450,6 +450,20 @@ struct llm_tokenizer_bpe {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_CHAMELEON:
+                // Note: in theory, the special token (sentinel and image token) regex_exprs below
+                // are unnecessary, as they are split in `tokenizer_st_partition` anyway.
+                // However, since the upstream pre-tokenizer uses them, they are also
+                // included here (see https://huggingface.co/facebook/chameleon-7b).
+                regex_exprs = {
+                    "<sentinel:[0-9]+>",  // Sentinel tokens
+                    "(IMGIMG)((A|B|C|D|E|F|G|H|I){1,4})Z",  // Image tokens
+                    "([\\t\\n]|    |  )",  // directly from tokenizer.json
+                    "\\p{N}", // Individual digits
+                    "[\\p{P}!-/:-@\\[-`{-~]",  // Punctuation, Isolated
+                    "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
diff --git a/src/llama.cpp b/src/llama.cpp
index 0accb1492..f450eaf9d 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -216,6 +216,7 @@ enum llm_arch {
     LLM_ARCH_RWKV6,
     LLM_ARCH_GRANITE,
     LLM_ARCH_GRANITE_MOE,
+    LLM_ARCH_CHAMELEON,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -268,6 +269,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RWKV6,           "rwkv6"        },
     { LLM_ARCH_GRANITE,         "granite"      },
     { LLM_ARCH_GRANITE_MOE,     "granitemoe"   },
+    { LLM_ARCH_CHAMELEON,       "chameleon"    },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -304,6 +306,7 @@ enum llm_kv {
     LLM_KV_DECODER_START_TOKEN_ID,
     LLM_KV_ATTN_LOGIT_SOFTCAPPING,
     LLM_KV_FINAL_LOGIT_SOFTCAPPING,
+    LLM_KV_SWIN_NORM,
     LLM_KV_RESCALE_EVERY_N_LAYERS,
     LLM_KV_TIME_MIX_EXTRA_DIM,
     LLM_KV_TIME_DECAY_EXTRA_DIM,
@@ -411,6 +414,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
     { LLM_KV_ATTN_LOGIT_SOFTCAPPING,            "%s.attn_logit_softcapping"            },
     { LLM_KV_FINAL_LOGIT_SOFTCAPPING,           "%s.final_logit_softcapping"           },
+    { LLM_KV_SWIN_NORM,                         "%s.swin_norm"                         },
     { LLM_KV_RESCALE_EVERY_N_LAYERS,            "%s.rescale_every_n_layers"            },
     { LLM_KV_TIME_MIX_EXTRA_DIM,                "%s.time_mix_extra_dim"                },
     { LLM_KV_TIME_DECAY_EXTRA_DIM,              "%s.time_decay_extra_dim"              },
@@ -1499,6 +1503,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_CHAMELEON,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2362,6 +2385,7 @@ struct llama_hparams {
     bool vocab_only;
     bool rope_finetuned;
     bool use_par_res;
+    bool swin_norm;
 
     uint32_t n_vocab;
     uint32_t n_ctx_train; // context size the model was trained on
@@ -6084,6 +6108,18 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_CHAMELEON:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                hparams.f_norm_eps = 1e-5;  // eps for qk-norm, torch default
+                ml.get_key(LLM_KV_SWIN_NORM, hparams.swin_norm);
+
+                switch (hparams.n_layer) {
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 48: model.type = e_model::MODEL_34B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+               }
+            } break;
         default: (void)0;
     }
 
@@ -6341,6 +6377,11 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "exaone") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
+            } else if (
+                tokenizer_pre == "chameleon") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
+                vocab.tokenizer_add_bos = true;
+                vocab.tokenizer_clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -8728,6 +8769,45 @@ static bool llm_load_tensors(
                     }
 
                 } break;
+            case LLM_ARCH_CHAMELEON:
+                {
+                 model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                 // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output      = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_q_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head});
+                        layer.attn_k_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv});
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i),  {n_embd_head_k, n_head}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i),  {n_embd_head_k, n_head_kv}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -15872,6 +15952,184 @@ struct llm_build_context {
 
         return gf;
     }
+
+    // ref: https://github.com/facebookresearch/chameleon
+    // based on the original build_llama() function, changes:
+    //   * qk-norm
+    //   * swin-norm
+    //   * removed bias
+    //   * removed MoE
+    struct ggml_cgraph * build_chameleon() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            if (hparams.swin_norm) {
+                cur = inpL;
+            } else {
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+                cb(cur, "attn_norm", il);
+            }
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
+                                ggml_element_size(Qcur) * n_embd_head,
+                                ggml_element_size(Qcur) * n_embd_head * n_head,
+                                0);
+                    cb(Qcur, "Qcur", il);
+
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                                model.layers[il].attn_q_norm,
+                                model.layers[il].attn_q_norm_b,
+                                LLM_NORM, cb, il);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
+                                ggml_element_size(Kcur) * n_embd_head,
+                                ggml_element_size(Kcur) * n_embd_head * n_head_kv,
+                                0);
+                    cb(Kcur, "Kcur", il);
+
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                               model.layers[il].attn_k_norm,
+                               model.layers[il].attn_k_norm_b,
+                               LLM_NORM, cb, il);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, nullptr,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+                if (hparams.swin_norm) {
+                    cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].attn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                }
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            if (!hparams.swin_norm) {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+            }
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            if (hparams.swin_norm) {
+                cur = llm_build_norm(ctx0, cur, hparams,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output_with_img_logits", -1);
+
+        // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
+        // Needs to be removed once image outputs are supported.
+        int img_token_end_idx = 8196;
+        int img_token_start_idx = 4;
+        int num_img_tokens = img_token_end_idx - img_token_start_idx;
+        // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
+        // which ensures that text token values are always at least larger than image token values
+        struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
+        img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
+        cb(img_logits, "img_logits", -1);
+        cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -16132,6 +16390,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_rwkv6();
             } break;
+        case LLM_ARCH_CHAMELEON:
+            {
+                result = llm.build_chameleon();
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -19257,6 +19519,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
+        case LLM_ARCH_CHAMELEON:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

From 6102037bbb55521880ae78a6ee6c2a0c00c901df Mon Sep 17 00:00:00 2001
From: Zhenwei Jin <109658203+kylo5aby@users.noreply.github.com>
Date: Sat, 28 Sep 2024 20:10:58 +0800
Subject: [PATCH 10/11] vocab : refactor tokenizer to reduce init overhead
 (#9449)

* refactor tokenizer

* llama : make llm_tokenizer more private

ggml-ci

* refactor tokenizer

* refactor tokenizer

* llama : make llm_tokenizer more private

ggml-ci

* remove unused files

* remove unused fileds to avoid unused filed build error

* avoid symbol link error

* Update src/llama.cpp

* Update src/llama.cpp

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
---
 .../convert-llama2c-to-ggml.cpp               |  14 +-
 src/llama-vocab.cpp                           | 266 +++++++++++-------
 src/llama-vocab.h                             |   9 +
 src/llama.cpp                                 |   2 +
 tests/test-tokenizer-0.cpp                    |  88 +++---
 5 files changed, 238 insertions(+), 141 deletions(-)

diff --git a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
index ecff95f9a..c140daed3 100644
--- a/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
+++ b/examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
@@ -201,7 +201,7 @@ static void print_sample_weights(TransformerWeights *w){
 
 //////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
 
-struct llama_vocab {
+struct my_llama_vocab {
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -525,7 +525,7 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
+static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
         LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
@@ -583,13 +583,13 @@ static void load_vocab(const char * filename, const Config * config, struct llam
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (llama_vocab::id id=0; id<n_vocab; ++id) {
+        for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
 
             unsigned char byte_val;
-            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
             if (id == UNKNOWN_TOKEN_ID) {
                 text = "<unk>";
                 type = LLAMA_TOKEN_TYPE_UNKNOWN;
@@ -631,7 +631,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
 }
 
 static void save_as_llama_model(
-    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+    struct my_llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
 ) {
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
@@ -671,7 +671,7 @@ static void save_as_llama_model(
     std::vector<const char*> tokens;
     std::vector<float> scores;
     std::vector<llama_token_type> token_types;
-    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
+    for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
         tokens.push_back(token_data.text.c_str());
         scores.push_back(token_data.score);
         token_types.push_back(token_data.type);
@@ -905,7 +905,7 @@ int main(int argc, char ** argv) {
         fclose(file);
     }
 
-    struct llama_vocab vocab;
+    struct my_llama_vocab vocab;
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index 146d416f7..e4d844a73 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -50,7 +50,7 @@ struct naive_trie {
             res.first->second.insert(key + 1, len - 1, value);
         }
     }
-    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) {
+    std::pair<const char *, size_t> get_longest_prefix(const char * key, size_t len, size_t offset = 0) const {
         if (len == 0 || offset == len) {
             return std::make_pair(key, offset);
         }
@@ -79,6 +79,15 @@ struct naive_trie {
 // impl
 //
 
+struct llm_tokenizer {
+   llm_tokenizer() {}
+   virtual ~llm_tokenizer() = default;
+};
+
+llama_vocab::~llama_vocab() {
+    delete tokenizer;
+}
+
 int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string & token_right) const {
     GGML_ASSERT(token_left.find(' ')   == std::string::npos);
     GGML_ASSERT(token_left.find('\n')  == std::string::npos);
@@ -187,10 +196,15 @@ struct llm_bigram_spm {
     size_t size;
 };
 
-struct llm_tokenizer_spm {
-    llm_tokenizer_spm(const llama_vocab & vocab) : vocab(vocab) {}
+struct llm_tokenizer_spm : llm_tokenizer {
+    llm_tokenizer_spm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
+
+struct llm_tokenizer_spm_session {
+    llm_tokenizer_spm_session(const llama_vocab & vocab) : vocab(vocab) {}
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
+
         // split string into utf8 chars
         int index = 0;
         size_t offs = 0;
@@ -271,7 +285,7 @@ private:
             return;
         }
 
-        resegment(symbols[p->second.first],  output);
+        resegment(symbols[p->second.first], output);
         resegment(symbols[p->second.second], output);
     }
 
@@ -279,7 +293,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         const std::string text = std::string(symbols[left].text, symbols[left].n + symbols[right].n);
         auto token = vocab.token_to_id.find(text);
 
@@ -306,10 +319,11 @@ private:
     }
 
     const llama_vocab & vocab;
+    // currently unused
+    // const llm_tokenizer_spm * spm_tokenizer;
 
     std::vector<llm_symbol> symbols;
     llm_bigram_spm::queue work_queue;
-
     std::map<std::string, std::pair<int, int>> rev_merge;
 };
 
@@ -352,8 +366,8 @@ struct llm_bigram_bpe {
     size_t size;
 };
 
-struct llm_tokenizer_bpe {
-    llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {
+struct llm_tokenizer_bpe : llm_tokenizer {
+    llm_tokenizer_bpe(const llama_vocab & vocab) : llm_tokenizer() {
         GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE);
         switch (vocab.type_pre) {
             case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
@@ -476,7 +490,14 @@ struct llm_tokenizer_bpe {
         }
     }
 
-    void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output) const {
+    std::vector<std::string> regex_exprs;
+};
+
+struct llm_tokenizer_bpe_session {
+    llm_tokenizer_bpe_session(const llama_vocab & vocab) : vocab(vocab),
+        bpe_tokenizer(static_cast<const llm_tokenizer_bpe *>(vocab.tokenizer)) {}
+
+    static void append(const llama_vocab::id token_id, std::vector<llama_vocab::id> & output)  {
         output.push_back(token_id);
     }
 
@@ -515,12 +536,11 @@ struct llm_tokenizer_bpe {
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         int final_prev_index = -1;
-
-        const auto word_collection = unicode_regex_split(text, regex_exprs);
+        const auto word_collection = unicode_regex_split(text, bpe_tokenizer->regex_exprs);
 
         symbols_final.clear();
 
-        for (auto & word : word_collection) {
+        for (const auto & word : word_collection) {
             work_queue = llm_bigram_bpe::queue();
             symbols.clear();
 
@@ -623,7 +643,6 @@ private:
         if (left == -1 || right == -1) {
             return;
         }
-
         std::string left_token  = std::string(symbols[left].text,  symbols[left].n);
         std::string right_token = std::string(symbols[right].text, symbols[right].n);
 
@@ -647,12 +666,10 @@ private:
     }
 
     const llama_vocab & vocab;
-
-    std::vector<std::string> regex_exprs;
+    const llm_tokenizer_bpe * bpe_tokenizer;
 
     std::vector<llm_symbol> symbols;
     std::vector<llm_symbol> symbols_final;
-
     llm_bigram_bpe::queue work_queue;
 };
 
@@ -660,15 +677,17 @@ private:
 // WPM tokenizer
 //
 
-struct llm_tokenizer_wpm {
-    llm_tokenizer_wpm(const llama_vocab & vocab): vocab(vocab) {}
+struct llm_tokenizer_wpm : llm_tokenizer {
+    llm_tokenizer_wpm(const llama_vocab & /*vocab*/) : llm_tokenizer() {}
+};
 
-    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) const {
+struct llm_tokenizer_wpm_session {
+    llm_tokenizer_wpm_session(const llama_vocab & vocab) : vocab(vocab) {}
+
+    void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         const auto & token_map = vocab.token_to_id;
-
         // normalize and split by whitespace
         std::vector<std::string> words = preprocess(text);
-
         // bos token prepended already
 
         // find the longest tokens that form the words
@@ -713,7 +732,7 @@ struct llm_tokenizer_wpm {
     }
 
     // TODO: reduce string copies by using cpts_offs array
-    std::vector<std::string> preprocess(const std::string & text) const {
+    static std::vector<std::string> preprocess(const std::string & text)  {
         const std::vector<uint32_t> cpts_nfd = unicode_cpts_normalize_nfd(unicode_cpts_from_utf8(text));
         std::vector<std::string> words(1, "");
 
@@ -765,15 +784,18 @@ struct llm_tokenizer_wpm {
             //(cpt >= 0xFF00  && cpt <= 0xFFEF);
     }
 
+private:
     const llama_vocab & vocab;
+    // currently unused
+    // const llm_tokenizer_wpm * wpm_tokenizer;
 };
 
 //
 // UGM tokenizer
 //
 
-struct llm_tokenizer_ugm {
-    llm_tokenizer_ugm(const llama_vocab & vocab) : vocab(vocab) {
+struct llm_tokenizer_ugm : llm_tokenizer {
+    llm_tokenizer_ugm(const llama_vocab & vocab) : llm_tokenizer() {
         if (vocab.precompiled_charsmap.size() > 0) {
             size_t charsmap_offset = 0;
 
@@ -819,6 +841,30 @@ struct llm_tokenizer_ugm {
         unknown_token_score = min_score - unknown_token_score_penalty;
     }
 
+    // escaped space symbol - U+2581 (Lower One Eighth Block)
+    const std::string escaped_space = "\xE2\x96\x81";
+
+    const char * prefix_replacements = NULL;
+    size_t prefix_replacements_size = 0;
+
+    const uint32_t * xcda_array = NULL;
+    size_t xcda_array_size = 0;
+
+    struct naive_trie user_defined_token_matcher;
+
+    float min_score = FLT_MAX;
+    float max_score = -FLT_MAX;
+
+    float unknown_token_score_penalty = 10.0;
+    float unknown_token_score;
+
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_ugm_session {
+    llm_tokenizer_ugm_session(const llama_vocab & vocab) : vocab(vocab),
+        ugm_tokenizer(static_cast<const llm_tokenizer_ugm *>(vocab.tokenizer)) {}
+
     /* This implementation is based on SentencePiece optimized Viterbi algorithm for
      * unigram language models. The general idea is to:
      * - move along the input sequence in steps of one UTF code point,
@@ -857,7 +903,7 @@ struct llm_tokenizer_ugm {
             // traverse the token matcher trie to find a matching token
             bool single_codepoint_token_found = false;
             const struct best_tokenization & current_best = tokenization_results[input_offset];
-            const struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]);
+            const struct naive_trie * node = ugm_tokenizer->token_matcher.traverse(normalized[prefix_offset++]);
 
             while (prefix_offset <= input_len && node != NULL) {
                 // check if we found valid token in prefix
@@ -887,7 +933,7 @@ struct llm_tokenizer_ugm {
             // if we didn't find a valid token corresponding to the whole UTF code point
             // then use unknown token as the tokenization of this UTF code point
             if (!single_codepoint_token_found) {
-                const double challenger_score = current_best.score_sum + unknown_token_score;
+                const double challenger_score = current_best.score_sum + ugm_tokenizer->unknown_token_score;
                 prefix_offset = input_offset + n_utf8_code_units;
                 struct best_tokenization & current_champ = tokenization_results[prefix_offset];
                 if (challenger_score > current_champ.score_sum) {
@@ -919,7 +965,6 @@ struct llm_tokenizer_ugm {
     }
 
 private:
-    const llama_vocab & vocab;
 
     // helper structure for returning normalization results
     struct normalization_result {
@@ -932,7 +977,7 @@ private:
         normalized->clear();
         normalized->reserve(input.size() * 3);
 
-        const std::string space = vocab.tokenizer_escape_whitespaces ? escaped_space : " ";
+        const std::string space = vocab.tokenizer_escape_whitespaces ? ugm_tokenizer->escaped_space : " ";
 
         bool shall_prepend_space = !vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
         bool shall_append_space = vocab.tokenizer_treat_whitespace_as_suffix && vocab.tokenizer_add_space_prefix;
@@ -1014,13 +1059,21 @@ private:
         size_t xcda_array_size;
     };
 
+    // this structure stores the best tokenization so far at input_offset
+    struct best_tokenization {
+        llama_token token_id;
+        size_t input_offset;
+        float score_sum;
+    };
+
     struct normalization_result normalize_prefix(const std::string & input, size_t input_offset) {
         if (input_offset == input.size()) {
             return { &input[input_offset], 0, 0 };
         }
 
         // if input prefix matches some user-defined token return this token as normalization result
-        auto user_defined_token_match = user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
+        auto user_defined_token_match =
+           ugm_tokenizer->user_defined_token_matcher.get_longest_prefix(&input[input_offset], input.size() - input_offset);
         if (user_defined_token_match.second > 0) {
             return { &input[input_offset], user_defined_token_match.second, user_defined_token_match.second };
         }
@@ -1028,8 +1081,8 @@ private:
         size_t longest_prefix_length = 0;
         size_t longest_prefix_offset = 0;
 
-        if (xcda_array_size > 0) {
-            struct xcda_array_view xcda_view(xcda_array, xcda_array_size);
+        if (ugm_tokenizer->xcda_array_size > 0) {
+            struct xcda_array_view xcda_view(ugm_tokenizer->xcda_array, ugm_tokenizer->xcda_array_size);
 
             // Find the longest normalized sequence matching the input prefix by walking
             // the XOR-compressed compact double array (XCDA) starting from the root node
@@ -1065,50 +1118,27 @@ private:
 
         if (longest_prefix_length > 0) {
             // we have a match, so return the replacement sequence
-            if (longest_prefix_offset >= prefix_replacements_size) {
+            if (longest_prefix_offset >= ugm_tokenizer->prefix_replacements_size) {
                 throw std::runtime_error("Index out of array bounds in precompiled charsmap!");
             }
-            const char * prefix_replacement = &prefix_replacements[longest_prefix_offset];
+            const char * prefix_replacement = &(ugm_tokenizer->prefix_replacements)[longest_prefix_offset];
             return { prefix_replacement, strlen(prefix_replacement), longest_prefix_length };
-        } else {
-            // check if the input prefix contains a valid sequence of UTF-8 code units
-            try {
-                // if yes, return this sequence unmodified
-                size_t prefix_offset = input_offset;
-                unicode_cpt_from_utf8(input, prefix_offset);
-                return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
-            } catch (std::invalid_argument & /*ex*/) {
-                // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
-                return { "\xEF\xBF\xBD", 3, 1 };
-            }
+        }
+
+        // check if the input prefix contains a valid sequence of UTF-8 code units
+        try {
+            // if yes, return this sequence unmodified
+            size_t prefix_offset = input_offset;
+            unicode_cpt_from_utf8(input, prefix_offset);
+            return { &input[input_offset], prefix_offset - input_offset, prefix_offset - input_offset };
+        } catch (std::invalid_argument & /*ex*/) {
+            // if no, consume 1 byte and return U+FFFD - REPLACEMENT CHARACTER
+            return { "\xEF\xBF\xBD", 3, 1 };
         }
     }
 
-    // escaped space symbol - U+2581 (Lower One Eighth Block)
-    const std::string escaped_space = "\xE2\x96\x81";
-
-    const char * prefix_replacements = NULL;
-    size_t prefix_replacements_size = 0;
-
-    const uint32_t * xcda_array = NULL;
-    size_t xcda_array_size = 0;
-
-    struct naive_trie user_defined_token_matcher;
-
-    // this structure stores the best tokenization so far at input_offset
-    struct best_tokenization {
-        llama_token token_id;
-        size_t input_offset;
-        float score_sum;
-    };
-
-    float min_score = FLT_MAX;
-    float max_score = -FLT_MAX;
-
-    float unknown_token_score_penalty = 10.0;
-    float unknown_token_score;
-
-    struct naive_trie token_matcher;
+    const llama_vocab & vocab;
+    const llm_tokenizer_ugm * ugm_tokenizer;
 };
 
 //
@@ -1169,8 +1199,8 @@ static std::vector<uint8_t> llama_unescape_rwkv_token(const std::string & escape
     return output;
 }
 
-struct llm_tokenizer_rwkv {
-    llm_tokenizer_rwkv(const llama_vocab & vocab): vocab(vocab) {
+struct llm_tokenizer_rwkv : llm_tokenizer {
+    llm_tokenizer_rwkv(const llama_vocab & vocab) : llm_tokenizer() {
         // RWKV supports arbitrary byte tokens, but the vocab struct only supports string tokens.
         // For now, we decode the vocab here into the lookup we'll use for tokenization.
 
@@ -1182,11 +1212,17 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+    struct naive_trie token_matcher;
+};
+
+struct llm_tokenizer_rwkv_session {
+    llm_tokenizer_rwkv_session(const llama_vocab & vocab) : vocab(vocab),
+        rwkv_tokenizer(static_cast<const llm_tokenizer_rwkv &>(*vocab.tokenizer)) {}
+
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         uint32_t position = 0;
-
         while (position < text.size()) {
-            const struct naive_trie * node = token_matcher.traverse(text[position]);
+            const struct naive_trie * node = rwkv_tokenizer.token_matcher.traverse(text[position]);
             if (node == NULL) {
                 // no matching token found, add unknown token
                 output.push_back(vocab.special_unk_id);
@@ -1211,11 +1247,33 @@ struct llm_tokenizer_rwkv {
         }
     }
 
+private:
     const llama_vocab & vocab;
-
-    struct naive_trie token_matcher;
+    const llm_tokenizer_rwkv & rwkv_tokenizer;
 };
 
+void llama_vocab::init_tokenizer() {
+    switch (type) {
+        case LLAMA_VOCAB_TYPE_SPM:
+            tokenizer = new llm_tokenizer_spm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_BPE:
+            tokenizer = new llm_tokenizer_bpe(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_WPM:
+            tokenizer = new llm_tokenizer_wpm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_UGM:
+            tokenizer = new llm_tokenizer_ugm(*this);
+            break;
+        case LLAMA_VOCAB_TYPE_RWKV:
+            tokenizer = new llm_tokenizer_rwkv(*this);
+            break;
+        default:
+            GGML_ABORT("unsupported vocab type");
+    }
+}
+
 //
 // (de-) tokenize
 //
@@ -1277,7 +1335,7 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
 
             // if a fragment is text ( not yet processed )
             if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
-                auto & raw_text = fragment.raw_text;
+                const auto & raw_text = fragment.raw_text;
 
                 auto raw_text_base_offset = fragment.offset;
                 auto raw_text_base_length = fragment.length;
@@ -1376,7 +1434,13 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
     }
 }
 
-std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool add_special, bool parse_special) {
+std::vector<llama_vocab::id> llama_tokenize_internal(
+        const llama_vocab & vocab,
+        std::string raw_text,
+        bool add_special,
+        bool parse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     std::vector<llama_vocab::id> output;
     std::forward_list<fragment_buffer_variant> fragment_buffer;
 
@@ -1413,9 +1477,9 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        llm_tokenizer_spm tokenizer(vocab);
                         llama_escape_whitespace(raw_text);
-                        tokenizer.tokenize(raw_text, output);
+                        llm_tokenizer_spm_session session(vocab);
+                        session.tokenize(raw_text, output);
                         is_prev_special = false;
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
@@ -1437,10 +1501,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                llm_tokenizer_bpe tokenizer(vocab);
-
+                llm_tokenizer_bpe_session session(vocab);
+                // it calls some other methods that are not exist in llm_tokenizer,
+                // here just cast it to bpe tokenizer object
                 if (add_special) {
-                    tokenizer.append_bos(output);
+                    session.append_bos(output);
                 }
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1449,15 +1514,15 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
-                        tokenizer.append(fragment.token, output);
+                        session.append(fragment.token, output);
                     }
                 }
 
                 if (add_special) {
-                    tokenizer.append_eos(output);
-                    tokenizer.check_double_bos_eos(output);
+                    session.append_eos(output);
+                    session.check_double_bos_eos(output);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
@@ -1467,7 +1532,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                     output.push_back(vocab.special_cls_id);
                 }
 
-                llm_tokenizer_wpm tokenizer(vocab);
+                llm_tokenizer_wpm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1476,7 +1541,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1489,12 +1554,11 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_UGM:
             {
-                llm_tokenizer_ugm tokenizer(vocab);
-
                 if (add_special && vocab.tokenizer_add_bos != 0) {
                     GGML_ASSERT(vocab.special_bos_id != -1);
                     output.push_back(vocab.special_bos_id);
                 }
+                llm_tokenizer_ugm_session session(vocab);
 
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -1502,7 +1566,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1522,6 +1586,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
             } break;
         case LLAMA_VOCAB_TYPE_RWKV:
             {
+                llm_tokenizer_rwkv_session session(vocab);
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
                         auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@@ -1530,8 +1595,7 @@ std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab,
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
 #endif
 
-                        llm_tokenizer_rwkv tokenizer(vocab);
-                        tokenizer.tokenize(raw_text, output);
+                        session.tokenize(raw_text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
                         output.push_back(fragment.token);
                     }
@@ -1644,13 +1708,13 @@ llama_token llama_token_eom_impl(const struct llama_vocab & vocab) {
 }
 
 int32_t llama_tokenize_impl(
-    const struct llama_vocab & vocab,
-                  const char * text,
-                     int32_t   text_len,
-                 llama_token * tokens,
-                     int32_t   n_tokens_max,
-                        bool   add_special,
-                        bool   parse_special) {
+        const struct llama_vocab & vocab,
+                      const char * text,
+                         int32_t   text_len,
+                     llama_token * tokens,
+                         int32_t   n_tokens_max,
+                            bool   add_special,
+                            bool   parse_special) {
     auto res = llama_tokenize_internal(vocab, std::string(text, text_len), add_special, parse_special);
     if (n_tokens_max < (int) res.size()) {
         // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
@@ -1775,6 +1839,8 @@ int32_t llama_detokenize_impl(
                          int32_t   text_len_max,
                             bool   remove_special,
                             bool   unparse_special) {
+    GGML_ASSERT(vocab.tokenizer && "Tokenizer not initialized. Call llama_vocab::init_tokenizer() first.");
+
     int32_t avail = text_len_max;
     int32_t total = 0;
 
diff --git a/src/llama-vocab.h b/src/llama-vocab.h
index cc46f642b..069bdc423 100644
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -8,6 +8,8 @@
 #include <map>
 #include <set>
 
+struct llm_tokenizer;
+
 struct llama_vocab {
     using id    = llama_token;
     using token = std::string;
@@ -65,7 +67,14 @@ struct llama_vocab {
 
     std::vector<char> precompiled_charsmap;
 
+    llm_tokenizer * tokenizer = nullptr;
+
+    llama_vocab() = default;
+    ~llama_vocab();
+
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+
+    void init_tokenizer();
 };
 
 //
diff --git a/src/llama.cpp b/src/llama.cpp
index f450eaf9d..44afb31d7 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -6464,6 +6464,8 @@ static void llm_load_vocab(
     }
     GGML_ASSERT(vocab.id_to_token.size() == vocab.token_to_id.size());
 
+    vocab.init_tokenizer();
+
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         // For Fill-In-the-Middle (FIM)/infill models which where converted
diff --git a/tests/test-tokenizer-0.cpp b/tests/test-tokenizer-0.cpp
index d3d21331b..4d49850c9 100644
--- a/tests/test-tokenizer-0.cpp
+++ b/tests/test-tokenizer-0.cpp
@@ -7,6 +7,7 @@
 #include <map>
 #include <vector>
 #include <fstream>
+#include <thread>
 
 //static const std::map<std::string, std::vector<llama_token>> & k_tests() {
 //    static std::map<std::string, std::vector<llama_token>> _k_tests = {
@@ -194,45 +195,64 @@ int main(int argc, char **argv) {
 
     const bool add_special = false;
 
-    for (const auto & test_kv : k_tests) {
-        const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
+    // multi-threaded tokenization
+    const int nthread = std::thread::hardware_concurrency();
+    std::vector<std::thread> threads(nthread);
 
-        printf("\n");
-        printf("src: '%s'\n", test_kv.first.c_str());
-        printf("res: '%s'\n", llama_detokenize(ctx, res).c_str());
-        printf("tok: ");
-        for (const auto & tok : res) {
-            printf("%d ", tok);
-        }
-        printf("\n");
+    for (int i = 0; i < nthread; i++) {
+        threads[i] = std::thread([&, i]() {
+            for (const auto & test_kv : k_tests) {
+                const std::vector<llama_token> res = llama_tokenize(ctx, test_kv.first, add_special, false);
 
-        bool correct = res.size() == test_kv.second.size();
-        for (int i = 0; i < (int) res.size() && correct; ++i) {
-            if (test_kv.second[i] != res[i]) {
-                correct = false;
+                // here only print the result of the first thread
+                // because the other threads are running the same tests
+                if (i != 0) {
+                    continue;
+                }
+
+                printf("\n");
+                printf("src: '%s'\n", test_kv.first.c_str());
+                printf("res: '%s'\n", llama_detokenize(ctx, res).c_str());
+                printf("tok: ");
+                for (const auto & tok : res) {
+                    printf("%d ", tok);
+                }
+                printf("\n");
+
+                bool correct = res.size() == test_kv.second.size();
+                for (int i = 0; i < (int) res.size() && correct; ++i) {
+                    if (test_kv.second[i] != res[i]) {
+                        correct = false;
+                    }
+                }
+
+                if (!correct) {
+                    fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
+                    fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
+                        llama_detokenize(ctx, res).c_str(),
+                        llama_detokenize(ctx, test_kv.second).c_str());
+                    fprintf(stderr, "%s : expected tokens: ", __func__);
+                    for (const auto & t : test_kv.second) {
+                        fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+                    fprintf(stderr, "%s : got tokens:      ", __func__);
+                    for (const auto & t : res) {
+                        fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
+                    }
+                    fprintf(stderr, "\n");
+
+                    success = false;
+                }
             }
-        }
-
-        if (!correct) {
-            fprintf(stderr, "%s : failed test:    '%s'\n", __func__, test_kv.first.c_str());
-            fprintf(stderr, "%s : detokenized to: '%s' instead of '%s'\n", __func__,
-                llama_detokenize(ctx, res).c_str(),
-                llama_detokenize(ctx, test_kv.second).c_str());
-            fprintf(stderr, "%s : expected tokens: ", __func__);
-            for (const auto & t : test_kv.second) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
-            fprintf(stderr, "%s : got tokens:      ", __func__);
-            for (const auto & t : res) {
-                fprintf(stderr, "%6d '%s', ", t, llama_token_to_piece(ctx, t).c_str());
-            }
-            fprintf(stderr, "\n");
-
-            success = false;
-        }
+        });
     }
 
+    for (int i = 0; i < nthread; i++) {
+        threads[i].join();
+    }
+
+    // single threaded tokenization
     if (!fname_text.empty()) {
         fprintf(stderr, "%s : tokenizing: '%s'\n", __func__, fname_text.c_str());
 

From 739842703e32cd43443c45e0b4f6647cc4e6b3d6 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Sat, 28 Sep 2024 15:13:21 +0300
Subject: [PATCH 11/11] llama : add comment about thread-safety [no ci] (#9449)

---
 include/llama.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/include/llama.h b/include/llama.h
index caef0bfff..4ea8a2c2b 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -911,6 +911,8 @@ extern "C" {
     //
     // Tokenization
     //
+    // The API is thread-safe.
+    //
 
     /// @details Convert the provided text into tokens.
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.