minor : spacing

* Add LLMFarm (ui for iOS) to list

.github/workflows/build.yml

@@ -76,6 +76,17 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
 
@@ -669,8 +680,7 @@ jobs:
         run: |
           cd examples/llama.android
 
-          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
-          ./gradlew build --no-daemon -Pskip-armeabi-v7a
+          ./gradlew build --no-daemon
 
 #  freeBSD-latest:
 #    runs-on: macos-12

.github/workflows/server.yml

@@ -6,11 +6,10 @@ on:
   push:
     branches:
       - master
-      - test/server-add-ci-test # FIXME remove
-    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
 
 jobs:
   server:
@@ -18,45 +17,21 @@ jobs:
 
     strategy:
       matrix:
-        build: [noavx, avx2, avx, avx512, cublas, clblast, openblas, kompute, vulkan]
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
         build_type: [Debug, Release]
         include:
-          - build: 'noavx'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF'
-            image: ubuntu:latest
-          - build: 'avx2'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
-            image: ubuntu:latest
-          - build: 'avx'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF'
-            image: ubuntu:latest
-          - build: 'avx512'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON'
-            image: ubuntu:latest
-            experimental: true
-          - build: 'cublas'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON'
-            image: nvidia/cuda:12.3.1-devel-ubuntu22.04
-            arch_not_available: true # require nvidia docker engine
-          - build: 'clblast'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON'
-            image: ubuntu:latest
-            arch_not_available: true
-          - build: 'openblas'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS'
-            image: ubuntu:latest
-          - build: 'kompute'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
-            image: ubuntu:latest
-            arch_not_available: true
-          - build: 'vulkan'
-            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON'
-            image: ubuntu:latest
-            arch_not_available: true
+          - build_type: Release
+            sanitizer: ""
+        exclude:
+          - build_type: Release
+            sanitizer: ADDRESS
+          - build_type: Release
+            sanitizer: THREAD
+          - build_type: Release
+            sanitizer: UNDEFINED
 
     container:
-      image: ${{ matrix.image }}
+      image: ubuntu:latest
       ports:
         - 8888
       options: --cpus 4
@@ -72,40 +47,22 @@ jobs:
           apt-get update
           apt-get -y install \
             build-essential \
-            pkg-config \
             git \
             cmake \
             python3-pip \
             wget \
             psmisc
 
-      - name: Download CLBlast
-        id: get_clblast
-        if: ${{ matrix.build == 'clblast' }}
-        run: |
-          apt install -y libclblast-dev
-
-      - name: Download OpenBLAS
-        id: get_openblas
-        if: ${{ matrix.build == 'openblas' }}
-        run: |
-          apt-get -y install libopenblas-dev
-
-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
-        run: |
-          wget -qO- https://packages.lunarg.com/lunarg-signing-key-pub.asc | tee /etc/apt/trusted.gpg.d/lunarg.asc
-          wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list http://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          apt-get update
-          apt-get -y install vulkan-sdk
-
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ${{ matrix.defines }}
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
 
       - name: Tests dependencies
@@ -121,7 +78,6 @@ jobs:
 
       - name: Tests
         id: server_integration_test
-        continue-on-error: ${{ matrix.experimental || matrix.arch_not_available }}
         run: |
           cd examples/server/tests
           PORT=8888 ./tests.sh

Makefile

@@ -597,7 +597,7 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH

README.md

@@ -114,6 +114,9 @@ Typically finetunes of the base models below are supported as well.
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 
+**HTTP server**
+
+[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
@@ -155,6 +158,8 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
 - [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
+- [Msty](https://msty.app) (proprietary)
+- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 
 ---
 

common/sampling.cpp

@@ -266,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
             //    }
             //}
 
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 

examples/convert-llama2c-to-ggml/README.md

@@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface.
 
 `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
+Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
+
 Now you can use the model with a command like:
 
 `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "log.h"
 
 #include <unordered_map>
 #include <vector>
@@ -78,111 +79,101 @@ typedef struct {
 
 struct TransformerWeights {
     // token embedding table
-    float* token_embedding_table;    // (vocab_size, dim)
+    std::vector<float> token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
-    float* rms_att_weight; // (layer, dim) rmsnorm weights
-    float* rms_ffn_weight; // (layer, dim)
+    std::vector<float> rms_att_weight; // (layer, dim) rmsnorm weights
+    std::vector<float> rms_ffn_weight; // (layer, dim)
     // weights for matmuls
-    float* wq; // (layer, dim, dim)
-    float* wk; // (layer, dim, dim)
-    float* wv; // (layer, dim, dim)
-    float* wo; // (layer, dim, dim)
+    std::vector<float> wq; // (layer, dim, dim)
+    std::vector<float> wk; // (layer, dim, dim)
+    std::vector<float> wv; // (layer, dim, dim)
+    std::vector<float> wo; // (layer, dim, dim)
     // weights for ffn
-    float* w1; // (layer, hidden_dim, dim)
-    float* w2; // (layer, dim, hidden_dim)
-    float* w3; // (layer, hidden_dim, dim)
+    std::vector<float> w1; // (layer, hidden_dim, dim)
+    std::vector<float> w2; // (layer, dim, hidden_dim)
+    std::vector<float> w3; // (layer, hidden_dim, dim)
     // final rmsnorm
-    float* rms_final_weight; // (dim,)
+    std::vector<float> rms_final_weight; // (dim,)
     // freq_cis for RoPE relatively positional embeddings
-    // float* freq_cis_real; // (seq_len, dim/2)
-    // float* freq_cis_imag; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_real; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    float* wcls;
-
-    ~TransformerWeights() {
-        delete[] token_embedding_table;
-        delete[] rms_att_weight;
-        delete[] rms_ffn_weight;
-        delete[] wq;
-        delete[] wk;
-        delete[] wv;
-        delete[] wo;
-        delete[] w1;
-        delete[] w2;
-        delete[] w3;
-        delete[] rms_final_weight;
-        delete[] wcls;
-    }
+    std::vector<float> wcls;
 };
 
-static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
-    // we calloc instead of malloc to keep valgrind happy
-    w->token_embedding_table = new float[p->vocab_size * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) {
+    const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
+    try {
+        w->token_embedding_table.resize(p->vocab_size * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
-    w->rms_att_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        w->rms_att_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
-    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        w->rms_ffn_weight.resize(p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
-    w->wq = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wq.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->wk = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wv = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
-    w->wo = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        w->wo.resize(p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
-    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
-    w->rms_final_weight = new float[p->dim]();
-    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        w->rms_final_weight.resize(p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
-    if (shared_weights) {
-        w->wcls = NULL;
-    } else {
-        w->wcls = new float[p->vocab_size * p->dim]();
-        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        if (shared_weights) {
+            w->wcls = {};
+        } else {
+            w->wcls.resize(p->vocab_size * p->dim);
+            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        }
+    }
+    catch (std::length_error &) {
+        die("Invalid configuration. Failed to allocate memory for weights");
     }
 }
 
-static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
-    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
-    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
-    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) {
+    if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1;
+    if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1;
+    if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1;
+    if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1;
+    if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1;
+    if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1;
+    if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1;
+    if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1;
+    if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1;
+    if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1;
+    if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1;
 
     // Skip freq_cis_real & freq_cis_imag
     int head_size = p->dim / p->n_heads;
     fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
 
-    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1;
 
     // Check we didn't forget to read anything
     auto curr = ftell(f);
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    printf("----- Quick print of first of the weight vales of all the variables\n");
-    printf("%f\n", w->token_embedding_table[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    printf("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    printf("%f\n", w->wq[0]);
-    printf("%f\n", w->wk[0]);
-    printf("%f\n", w->wv[0]);
-    printf("%f\n", w->wo[0]);
-    printf("%f\n", w->w1[0]);
-    printf("%f\n", w->w2[0]);
-    printf("%f\n", w->w3[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    if (w->wcls) printf("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -225,14 +216,16 @@ struct llama_vocab {
 };
 
 struct my_llama_hparams {
-    uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
-    uint32_t n_ff    = 11008;
-    uint32_t n_mult  = 4;
-    uint32_t n_head  = 32;
-    uint32_t n_layer = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_vocab   = 32000;
+    uint32_t n_ctx     = 512;   // this is provided as user input?
+    uint32_t n_embd    = 4096;
+    uint32_t n_ff      = 11008;
+    uint32_t n_mult    = 4;
+    uint32_t n_head    = 32;
+    uint32_t n_head_kv = 32;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
     bool operator!=(const my_llama_hparams& other) const {
         return memcmp(this, &other, sizeof(my_llama_hparams));
     }
@@ -325,14 +318,30 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
-    printf("%s: n_head:  %u\n", __func__, params->n_head);
-    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
-    printf("%s: n_layer: %u\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+}
+
+static void print_tensor_info(const struct ggml_context * ctx) {
+    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        LOG("%s: Allocating ", __func__);
+        int64_t total = 1;
+        int i = 0;
+        for (; i < ggml_n_dims(t); ++i) {
+            if (i > 0) LOG("x ");
+            LOG("[%" PRId64 "] ", t->ne[i]);
+            total *= t->ne[i];
+        }
+        if (i > 1) LOG("= [%" PRId64 "] ", total);
+        LOG("float space for %s\n", ggml_get_name(t));
+    }
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
+    const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv;
+
     const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
@@ -350,25 +361,8 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
-
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
-
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
-
-    // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");
@@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) {
         layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
         layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
         layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
@@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) {
         ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
         ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
     }
+
+    print_tensor_info(ctx);
 }
 
 static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
@@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
 static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
-        printf(" %f", p);
+        LOG(" %f", p);
     }
-    printf("\n");
+    LOG("\n");
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
@@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
+            LOG(" %.2f", p);
         }
-        printf("\n");
+        LOG("\n");
     }
 }
 
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
@@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
 
         const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+        if (n_vocab != static_cast<uint32_t>(config->vocab_size)) {
+            die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size);
+        }
 
         vocab->id_to_token.resize(n_vocab);
 
@@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 }
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
-    int ct;
-    switch (ggml_n_dims(gg_weights)) {
-        case 1:
-            ct = 0;
-            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
-                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
-                *ptr = karpathy_weights[ct];
-                ct++;
-            }
-            break;
-        case 2:
-            ct = 0;
-            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
-                    *ptr = karpathy_weights[ct];
-                    ct++;
-                }
-            }
-            break;
-        case 3:
-            ct = 0;
-            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
-                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
-                        *ptr = karpathy_weights[ct];
-                        ct++;
-                    }
-                }
-            }
-            break;
+    int size = 1;
+    for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) {
+        size *= gg_weights->ne[dim];
+    }
+    for (int ct = 0; ct < size; ++ct) {
+        int64_t i0 = 0; int64_t i1 = 0;
+        int64_t i2 = 0; int64_t i3 = 0;
+        ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3);
+        ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]);
     }
 }
 
@@ -679,16 +635,18 @@ static void save_as_llama_model(
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
-    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data());
+    convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data());
 
-    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data());
     //print_row(model->norm, 0);
 
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     int n_ff = model->hparams.n_ff;
 
+    const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv;
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
@@ -697,9 +655,10 @@ static void save_as_llama_model(
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
         convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
         convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length/n_multiqueries]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length/n_multiqueries]);
 
         convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
         convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
@@ -736,8 +695,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
-    // n_head_kv is optional, default to n_head
-    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv);
     gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
     gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
     gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
@@ -789,12 +748,12 @@ static void save_as_llama_model(
 
 static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_vocab_model          = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
-    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+    params.fn_train_data           = "shakespeare.txt";
+    params.fn_checkpoint_in        = "checkpoint.bin";
+    params.fn_checkpoint_out       = "checkpoint.bin";
+    params.fn_model_out            = "ggml-checkpoint-f32.bin";
 
     params.seed       =   -1;
 
@@ -829,8 +788,8 @@ static struct train_params get_default_train_params() {
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
 
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
+    params.mem_model_gb    = 2;
+    params.mem_compute_gb  = 24;
     params.mem_compute0_gb = 8;
     params.mem_compute1_gb = 2;
 
@@ -916,19 +875,30 @@ int main(int argc, char ** argv) {
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        FILE *file = fopen(params.fn_llama2c_model, "rb");
-        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        FILE *file = fopen(params.fn_llama2c_model, "r");
+        if (!file) {
+            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            return 1;
+        }
         // read in the config header
-        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        if (fread(&config, sizeof(Config), 1, file) != 1) {
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         auto shared_weights = config.vocab_size > 0;
         config.vocab_size = abs(config.vocab_size);
 
         // read in the Transformer weights
-        malloc_weights(&weights, &config, shared_weights);
-        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        alloc_weights(&weights, &config, shared_weights);
+        if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
+            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         fclose(file);
     }
 
@@ -936,15 +906,18 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
-    model.hparams.n_embd  = config.dim; //params.n_embd;
-    model.hparams.n_ff    = config.hidden_dim;
-    model.hparams.n_mult  = 32;//params.n_mult;
-    model.hparams.n_head  = config.n_heads; //params.n_head;
-    model.hparams.n_layer = config.n_layers; //params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx     = params.n_ctx;
+    model.hparams.n_embd    = config.dim; //params.n_embd;
+    model.hparams.n_ff      = config.hidden_dim;
+    model.hparams.n_mult    = 32;//params.n_mult;
+    model.hparams.n_head    = config.n_heads; //params.n_head;
+    model.hparams.n_head_kv = config.n_kv_heads;
+    model.hparams.n_layer   = config.n_layers; //params.n_layer;
+    model.hparams.n_rot     = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+
     print_params(&model.hparams);
+
     struct ggml_init_params lcparams;
     lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
     lcparams.mem_buffer = NULL;
@@ -956,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/infill/infill.cpp

@@ -447,8 +447,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/llama.android/app/build.gradle.kts

@@ -21,12 +21,8 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // Workaround for https://github.com/llvm/llvm-project/issues/65820
-            // affecting armeabi-v7a. Skip armeabi-v7a when invoked with
-            // -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
-            if (project.hasProperty("skip-armeabi-v7a")) {
-                abiFilters += listOf("arm64-v8a", "x86_64", "x86")
-            }
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
         }
         externalNativeBuild {
             cmake {

examples/main/main.cpp

@@ -548,8 +548,8 @@ int main(int argc, char ** argv) {
                     LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -576,9 +576,9 @@ int main(int argc, char ** argv) {
                     LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 

examples/passkey/passkey.cpp

@@ -126,7 +126,7 @@ int main(int argc, char ** argv) {
     const int n_batch     = ctx_params.n_batch;
     const int n_batch_grp = ctx_params.n_batch/n_grp;
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
 
     // print the prompt token-by-token
 
@@ -146,10 +146,11 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+            llama_kv_cache_update  (ctx);
 
-            n_past -= bd;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
 
         llama_batch_clear(batch);
@@ -179,10 +180,12 @@ int main(int argc, char ** argv) {
 
         LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+        llama_kv_cache_defrag (ctx);
+        llama_kv_cache_update (ctx);
 
-        n_past -= n_discard;
+        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
 
         llama_batch_clear(batch);
 
@@ -208,10 +211,12 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+            llama_kv_cache_defrag (ctx);
+            llama_kv_cache_update (ctx);
 
-            n_past -= n_discard;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/server/README.md

@@ -1,8 +1,20 @@
-# llama.cpp/example/server
+# LLaMA.cpp HTTP Server
 
-This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
+Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/yhirose/cpp-httplib), [nlohmann::json](https://github.com/nlohmann/json) and **llama.cpp**.
 
-Command line options:
+Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
+
+**Features:**
+ * LLM inference of F16 and quantum models on GPU and CPU
+ * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
+ * Parallel decoding with multi-user support
+ * Continuous batching
+ * Multimodal (wip)
+ * Monitoring endpoints
+
+The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
+
+**Command line options:**
 
 - `--threads N`, `-t N`: Set the number of threads to use during generation.
 - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
@@ -39,9 +51,12 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 - `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
+- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
+- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
 - `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+- `--log-disable`: Output logs to stdout only, default: enabled.
+- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
 
 ## Build
 
@@ -457,6 +472,18 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
+- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
+
+Available metrics:
+- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
+- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
+- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
+- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
+- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
+- `llamacpp:kv_cache_tokens`: KV-cache tokens.
+- `llamacpp:requests_processing`: Number of request processing.
+- `llamacpp:requests_deferred`: Number of request deferred.
+
 ## More examples
 
 ### Change system prompt on runtime

examples/server/server.cpp

@@ -43,9 +43,11 @@ struct server_params
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
     bool slots_endpoint = true;
+    bool metrics_endpoint = false;
 };
 
 bool server_verbose = false;
+bool server_log_json = true;
 
 static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
 {
@@ -301,12 +303,76 @@ struct llama_client_slot
     }
 
     void print_timings() const {
-        LOG_TEE("\n");
-        LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed);
-        LOG_TEE("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, t_token_generation, n_decoded,t_token_generation / n_decoded, 1e3 / t_token_generation * n_decoded);
-        LOG_TEE("%s:       total time = %10.2f ms\n", __func__, t_prompt_processing + t_token_generation);
+       char buffer[512];
+        double t_token = t_prompt_processing / num_prompt_tokens_processed;
+        double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
+        sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
+                t_prompt_processing, num_prompt_tokens_processed,
+                t_token, n_tokens_second);
+        LOG_INFO(buffer, {
+            {"slot_id",                     id},
+            {"task_id",                     task_id},
+            {"t_prompt_processing",         t_prompt_processing},
+            {"num_prompt_tokens_processed", num_prompt_tokens_processed},
+            {"t_token",                     t_token},
+            {"n_tokens_second",             n_tokens_second},
+        });
+
+        t_token = t_token_generation / n_decoded;
+        n_tokens_second = 1e3 / t_token_generation * n_decoded;
+        sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
+                t_token_generation, n_decoded,
+                t_token, n_tokens_second);
+        LOG_INFO(buffer, {
+            {"slot_id",            id},
+            {"task_id",            task_id},
+            {"t_token_generation", t_token_generation},
+            {"n_decoded",          n_decoded},
+            {"t_token",            t_token},
+            {"n_tokens_second",    n_tokens_second},
+        });
+
+        sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
+        LOG_INFO(buffer, {
+            {"slot_id",             id},
+            {"task_id",             task_id},
+            {"t_prompt_processing", t_prompt_processing},
+            {"t_token_generation",  t_token_generation},
+            {"t_total",             t_prompt_processing + t_token_generation},
+        });
+    }
+};
+
+struct llama_metrics {
+    uint64_t n_prompt_tokens_processed_total = 0;
+    uint64_t n_tokens_predicted_total        = 0;
+
+    uint64_t n_prompt_tokens_processed = 0;
+    uint64_t t_prompt_processing       = 0;
+
+    uint64_t n_tokens_predicted       = 0;
+    uint64_t t_tokens_generation      = 0;
+
+
+    void on_prompt_eval(const llama_client_slot &slot) {
+        n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
+
+        n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
+        t_prompt_processing       += slot.t_prompt_processing;
+    }
+
+    void on_prediction(const llama_client_slot &slot) {
+        n_tokens_predicted_total += slot.n_decoded;
+
+        n_tokens_predicted  += slot.n_decoded;
+        t_tokens_generation += slot.t_token_generation;
+    }
+
+    void reset_bucket() {
+        n_prompt_tokens_processed = 0;
+        t_prompt_processing       = 0;
+        n_tokens_predicted        = 0;
+        t_tokens_generation       = 0;
     }
 };
 
@@ -344,6 +410,8 @@ struct llama_server_context
     llama_server_queue queue_tasks;
     llama_server_response queue_results;
 
+    llama_metrics metrics;
+
     ~llama_server_context()
     {
         if (ctx)
@@ -363,7 +431,7 @@ struct llama_server_context
         params = params_;
         if (!params.mmproj.empty()) {
             multimodal = true;
-            LOG_TEE("Multi Modal Mode Enabled");
+            LOG_INFO("Multi Modal Mode Enabled", {});
             clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
             if(clp_ctx == nullptr) {
                 LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
@@ -416,7 +484,7 @@ struct llama_server_context
 
         const int32_t n_ctx_slot = n_ctx / params.n_parallel;
 
-        LOG_TEE("Available slots:\n");
+        LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
         for (int i = 0; i < params.n_parallel; i++)
         {
             llama_client_slot slot;
@@ -425,7 +493,10 @@ struct llama_server_context
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
 
-            LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
+            LOG_INFO("new slot", {
+                {"slot_id",    slot.id},
+                {"n_ctx_slot", slot.n_ctx}
+            });
 
             const int ga_n = params.grp_attn_n;
             const int ga_w = params.grp_attn_w;
@@ -435,7 +506,12 @@ struct llama_server_context
                 GGML_ASSERT(ga_w % ga_n == 0            && "ga_w must be a multiple of ga_n");     // NOLINT
                 //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of ga_w");    // NOLINT
                 //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
-                LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w);
+
+                LOG_INFO("slot self-extend", {
+                    {"slot_id",   slot.id},
+                    {"ga_n",      ga_n},
+                    {"ga_w",      ga_w}
+                });
             }
 
             slot.ga_i = 0;
@@ -729,10 +805,16 @@ struct llama_server_context
                     img_sl.img_data = clip_image_u8_init();
                     if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
                     {
-                        LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
+                        LOG_ERROR("failed to load image", {
+                            {"slot_id",   slot->id},
+                            {"img_sl_id", img_sl.id}
+                        });
                         return false;
                     }
-                    LOG_TEE("slot %i - loaded image\n", slot->id);
+                    LOG_VERBOSE("image loaded", {
+                        {"slot_id",   slot->id},
+                        {"img_sl_id", img_sl.id}
+                    });
                     img_sl.request_encode_image = true;
                     slot->images.push_back(img_sl);
                 }
@@ -792,7 +874,10 @@ struct llama_server_context
 
         all_slots_are_idle = false;
 
-        LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id);
+        LOG_INFO("slot is processing task", {
+            {"slot_id", slot->id},
+            {"task_id", slot->task_id},
+        });
 
         return true;
     }
@@ -817,10 +902,24 @@ struct llama_server_context
                 llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
             }
 
-            if (llama_decode(ctx, batch) != 0)
+            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
             {
-                LOG_TEE("%s: llama_decode() failed\n", __func__);
-                return;
+                const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
+                llama_batch batch_view = {
+                    n_tokens,
+                    batch.token    + i,
+                    nullptr,
+                    batch.pos      + i,
+                    batch.n_seq_id + i,
+                    batch.seq_id   + i,
+                    batch.logits   + i,
+                    0, 0, 0, // unused
+                };
+                if (llama_decode(ctx, batch_view) != 0)
+                {
+                    LOG_TEE("%s: llama_decode() failed\n", __func__);
+                    return;
+                }
             }
 
             // assign the system KV cache to all parallel sequences
@@ -1355,7 +1454,7 @@ struct llama_server_context
                 if (slot == nullptr)
                 {
                     // if no slot is available, we defer this task for processing later
-                    LOG_VERBOSE("no slot is available", {});
+                    LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
                     queue_tasks.defer(task);
                     break;
                 }
@@ -1404,7 +1503,7 @@ struct llama_server_context
             case TASK_TYPE_NEXT_RESPONSE: {
                 // do nothing
             } break;
-            case TASK_TYPE_SLOTS_DATA: {
+            case TASK_TYPE_METRICS: {
                 json slots_data        = json::array();
                 int n_idle_slots       = 0;
                 int n_processing_slots = 0;
@@ -1431,17 +1530,41 @@ struct llama_server_context
                     }
                     slots_data.push_back(slot_data);
                 }
-                LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
+                LOG_INFO("slot data", {
+                    {"task_id",            task.id},
+                    {"n_idle_slots",       n_idle_slots},
+                    {"n_processing_slots", n_processing_slots}
+                });
+                LOG_VERBOSE("slot data", {
+                    {"task_id",            task.id},
+                    {"n_idle_slots",       n_idle_slots},
+                    {"n_processing_slots", n_processing_slots},
+                    {"slots",              slots_data}
+                });
                 task_result res;
                 res.id = task.id;
                 res.multitask_id = task.multitask_id;
                 res.stop = true;
                 res.error = false;
                 res.result_json = {
-                        { "idle",       n_idle_slots       },
-                        { "processing", n_processing_slots },
-                        { "slots",      slots_data         }
+                        { "idle",                            n_idle_slots       },
+                        { "processing",                      n_processing_slots },
+                        { "deferred",                        queue_tasks.queue_tasks_deferred.size() },
+
+                        { "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total},
+                        { "n_tokens_predicted_total",        metrics.n_tokens_predicted_total},
+
+                        { "n_prompt_tokens_processed",       metrics.n_prompt_tokens_processed},
+                        { "t_prompt_processing",             metrics.t_prompt_processing},
+                        { "n_tokens_predicted",              metrics.n_tokens_predicted},
+                        { "t_tokens_generation",             metrics.t_tokens_generation},
+
+                        { "kv_cache_tokens_count",          llama_get_kv_cache_token_count(ctx)},
+                        { "kv_cache_used_cells",            llama_get_kv_cache_used_cells(ctx)},
+
+                        { "slots",                          slots_data },
                 };
+                metrics.reset_bucket();
                 queue_results.send(res);
             } break;
         }
@@ -1469,7 +1592,7 @@ struct llama_server_context
     bool update_slots() {
         if (system_need_update)
         {
-            LOG_TEE("updating system prompt\n");
+            LOG_INFO("updating system prompt", {});
             update_system_prompt();
         }
 
@@ -1479,12 +1602,13 @@ struct llama_server_context
         {
             if (system_prompt.empty() && clean_kv_cache)
             {
-                LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
+                LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
                 kv_cache_clear();
             }
             return true;
         }
 
+        LOG_VERBOSE("posting NEXT_RESPONSE", {});
         task_server task;
         task.type = TASK_TYPE_NEXT_RESPONSE;
         task.target_id = -1;
@@ -1498,12 +1622,22 @@ struct llama_server_context
                 {
                     // Shift context
                     const int n_keep    = slot.params.n_keep + add_bos_token;
-                    const int n_left    = system_tokens.size() + slot.n_past - n_keep;
+                    const int n_left    = (int) system_tokens.size() + slot.n_past - n_keep;
                     const int n_discard = n_left / 2;
 
-                    LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
-                    llama_kv_cache_seq_rm   (ctx, slot.id, n_keep            , n_keep + n_discard);
-                    llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
+                    LOG_INFO("slot context shift", {
+                        {"slot_id",         slot.id},
+                        {"task_id",         slot.task_id},
+                        {"n_keep",          n_keep},
+                        {"n_left",          n_left},
+                        {"n_discard",       n_discard},
+                        {"n_ctx",           n_ctx},
+                        {"n_past",          slot.n_past},
+                        {"n_system_tokens", system_tokens.size()},
+                        {"n_cache_tokens",  slot.cache_tokens.size()}
+                    });
+                    llama_kv_cache_seq_rm (ctx, slot.id, n_keep            , n_keep + n_discard);
+                    llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
 
                     for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
                     {
@@ -1515,17 +1649,12 @@ struct llama_server_context
                     slot.n_past -= n_discard;
 
                     slot.truncated = true;
-
-                    LOG_VERBOSE("context shift", {
-                        { "n_ctx", n_ctx },
-                        { "n_keep", n_keep },
-                        { "n_left", n_left },
-                    });
                 }
             }
         }
 
         // decode any currently ongoing sequences
+        LOG_VERBOSE("decoding ongoing sequences", {});
         for (auto & slot : slots)
         {
             // release the slot
@@ -1535,7 +1664,15 @@ struct llama_server_context
                 slot.command = NONE;
                 slot.t_last_used = ggml_time_us();
 
-                LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
+                LOG_INFO("slot released", {
+                    {"slot_id",         slot.id},
+                    {"task_id",         slot.task_id},
+                    {"n_ctx",           n_ctx},
+                    {"n_past",          slot.n_past},
+                    {"n_system_tokens", system_tokens.size()},
+                    {"n_cache_tokens",  slot.cache_tokens.size()},
+                    {"truncated",       slot.truncated}
+                });
                 queue_tasks.notify_slot_changed();
 
                 continue;
@@ -1662,6 +1799,14 @@ struct llama_server_context
                         }
 
                         slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
+
+                        // the last token of the cache is not in the KV cache until the next call to llama_decode
+                        // (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
+                        if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
+                        {
+                            slot.n_past -= 1;
+                        }
+
                         slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
 
                         if (slot.ga_n != 1)
@@ -1683,7 +1828,12 @@ struct llama_server_context
                             slot.ga_i = ga_i;
                         }
 
-                        LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
+                        LOG_INFO("slot progression", {
+                            { "slot_id", slot.id },
+                            { "task_id", slot.task_id },
+                            { "n_past",  slot.n_past },
+                            { "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
+                        });
                     }
 
                     slot.cache_tokens = prompt_tokens;
@@ -1691,7 +1841,10 @@ struct llama_server_context
                     if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
                     {
                         // we have to evaluate at least 1 token to generate logits.
-                        LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);
+                        LOG_INFO("we have to evaluate at least 1 token to generate logits", {
+                            { "slot_id", slot.id },
+                            { "task_id", slot.task_id }
+                        });
                         slot.n_past--;
                         if (slot.ga_i > 0)
                         {
@@ -1699,9 +1852,13 @@ struct llama_server_context
                         }
                     }
 
-                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
-
-                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
+                    int p0 = (int) system_tokens.size() + slot.n_past;
+                    LOG_INFO("kv cache rm [p0, end)", {
+                        { "slot_id", slot.id },
+                        { "task_id", slot.task_id },
+                        { "p0",      p0 }
+                    });
+                    llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
 
                     LOG_VERBOSE("prompt ingested", {
                                                     {"n_past",  slot.n_past},
@@ -1736,7 +1893,13 @@ struct llama_server_context
 
                     if (has_images && !ingest_images(slot, n_batch))
                     {
-                        LOG_TEE("failed processing images\n");
+                        LOG_ERROR("failed processing images", {
+                            "slot_id", slot.id,
+                            "task_id", slot.task_id,
+                        });
+                        // FIXME @phymbert: to be properly tested
+                        //  early returning without changing the slot state will block the slot for ever
+                        // no one at the moment is checking the return value
                         return false;
                     }
 
@@ -1778,9 +1941,9 @@ struct llama_server_context
                         LOG_TEE("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
                         LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
 
-                        llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
+                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
                         llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n);
-                        llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
+                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
 
                         slot.n_past_se -= bd;
 
@@ -1849,6 +2012,7 @@ struct llama_server_context
                 {
                     slot.t_start_genereration = ggml_time_us();
                     slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
+                    metrics.on_prompt_eval(slot);
                 }
 
                 llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
@@ -1871,11 +2035,14 @@ struct llama_server_context
                     slot.release();
                     slot.print_timings();
                     send_final_response(slot);
+                    metrics.on_prediction(slot);
                 }
 
                 slot.i_batch = -1;
             }
         }
+
+        LOG_VERBOSE("slots updated", {});
         return true;
     }
 
@@ -1953,8 +2120,10 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  -ctv TYPE, --cache-type-v TYPE\n");
     printf("                            KV cache data type for V (default: f16)\n");
     printf("  --mmproj MMPROJ_FILE      path to a multimodal projector file for LLaVA.\n");
+    printf("  --log-format              log output format: json or text (default: json)\n");
     printf("  --log-disable             disables logging to a file.\n");
     printf("  --slots-endpoint-disable  disables slots monitoring endpoint.\n");
+    printf("  --metrics                 enable prometheus compatible metrics endpoint (default: %s).\n", sparams.metrics_endpoint ? "enabled" : "disabled");
     printf("\n");
     printf("  -n, --n-predict           maximum tokens to predict (default: %d)\n", params.n_predict);
     printf("  --override-kv KEY=TYPE:VALUE\n");
@@ -2405,6 +2574,27 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.mmproj = argv[i];
         }
+        else if (arg == "--log-format")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            if (std::strcmp(argv[i], "json") == 0)
+            {
+                server_log_json = true;
+            }
+            else if (std::strcmp(argv[i], "text") == 0)
+            {
+                server_log_json = false;
+            }
+            else
+            {
+                invalid_param = true;
+                break;
+            }
+        }
         else if (arg == "--log-disable")
         {
             log_set_target(stdout);
@@ -2414,6 +2604,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
         {
             sparams.slots_endpoint = false;
         }
+        else if (arg == "--metrics")
+        {
+            sparams.metrics_endpoint = true;
+        }
         else if (arg == "--chat-template")
         {
             if (++i >= argc)
@@ -2514,32 +2708,40 @@ static json format_partial_response(
 
 static json format_tokenizer_response(const std::vector<llama_token> &tokens)
 {
-    return json{
-        {"tokens", tokens}};
+    return json {
+        {"tokens", tokens}
+    };
 }
 
 static json format_detokenized_response(std::string content)
 {
-    return json{
-        {"content", content}};
+    return json {
+        {"content", content}
+    };
 }
 
 
 static void log_server_request(const httplib::Request &req, const httplib::Response &res)
 {
+    // skip GH copilot requests when using default port
+    if (req.path == "/v1/health" || req.path == "/v1/completions")
+    {
+        return;
+    }
+
     LOG_INFO("request", {
-                            {"remote_addr", req.remote_addr},
-                            {"remote_port", req.remote_port},
-                            {"status", res.status},
-                            {"method", req.method},
-                            {"path", req.path},
-                            {"params", req.params},
-                        });
+        {"remote_addr", req.remote_addr},
+        {"remote_port", req.remote_port},
+        {"status",      res.status},
+        {"method",      req.method},
+        {"path",        req.path},
+        {"params",      req.params},
+    });
 
     LOG_VERBOSE("request", {
-                               {"request", req.body},
-                               {"response", res.body},
-                           });
+        {"request",  req.body},
+        {"response", res.body},
+    });
 }
 
 struct token_translator
@@ -2621,7 +2823,7 @@ int main(int argc, char **argv)
                 // request slots data using task queue
                 task_server task;
                 task.id   = llama.queue_tasks.get_new_id();
-                task.type = TASK_TYPE_SLOTS_DATA;
+                task.type = TASK_TYPE_METRICS;
                 task.target_id = -1;
 
                 llama.queue_results.add_waiting_task_id(task.id);
@@ -2668,7 +2870,7 @@ int main(int argc, char **argv)
             // request slots data using task queue
             task_server task;
             task.id = llama.queue_tasks.get_new_id();
-            task.type = TASK_TYPE_SLOTS_DATA;
+            task.type = TASK_TYPE_METRICS;
             task.target_id = -1;
 
             llama.queue_results.add_waiting_task_id(task.id);
@@ -2683,6 +2885,87 @@ int main(int argc, char **argv)
         });
     }
 
+    if (sparams.metrics_endpoint) {
+        svr.Get("/metrics", [&](const httplib::Request&, httplib::Response& res) {
+            // request slots data using task queue
+            task_server task;
+            task.id = llama.queue_tasks.get_new_id();
+            task.type = TASK_TYPE_METRICS;
+            task.target_id = -1;
+
+            llama.queue_results.add_waiting_task_id(task.id);
+            llama.queue_tasks.post(task);
+
+            // get the result
+            task_result result = llama.queue_results.recv(task.id);
+            llama.queue_results.remove_waiting_task_id(task.id);
+
+            json data = result.result_json;
+
+            uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"];
+            uint64_t t_prompt_processing       = data["t_prompt_processing"];
+
+            uint64_t n_tokens_predicted       = data["n_tokens_predicted"];
+            uint64_t t_tokens_generation      = data["t_tokens_generation"];
+
+            int32_t kv_cache_used_cells = data["kv_cache_used_cells"];
+
+            // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
+            json all_metrics_def = json {
+                    {"counter", {{
+                            {"name",  "prompt_tokens_total"},
+                            {"help",  "Number of prompt tokens processed."},
+                            {"value",  data["n_prompt_tokens_processed_total"]}
+                    }, {
+                            {"name",  "tokens_predicted_total"},
+                            {"help",  "Number of generation tokens processed."},
+                            {"value",  data["n_tokens_predicted_total"]}
+                    }}},
+                    {"gauge", {{
+                            {"name",  "prompt_tokens_seconds"},
+                            {"help",  "Average prompt throughput in tokens/s."},
+                            {"value",  n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0}
+                    },{
+                            {"name",  "predicted_tokens_seconds"},
+                            {"help",  "Average generation throughput in tokens/s."},
+                            {"value",  n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0}
+                     },{
+                            {"name",  "kv_cache_usage_ratio"},
+                            {"help",  "KV-cache usage. 1 means 100 percent usage."},
+                            {"value",  1. * kv_cache_used_cells / params.n_ctx}
+                     },{
+                            {"name",  "kv_cache_tokens"},
+                            {"help",  "KV-cache tokens."},
+                            {"value",  data["kv_cache_tokens_count"]}
+                    },{
+                            {"name",  "requests_processing"},
+                            {"help",  "Number of request processing."},
+                            {"value",  data["processing"]}
+                  },{
+                            {"name",  "requests_deferred"},
+                            {"help",  "Number of request deferred."},
+                            {"value",  data["deferred"]}
+                  }}}
+            };
+
+            std::stringstream prometheus;
+            for (const auto& el : all_metrics_def.items()) {
+                const auto& type = el.key();
+                const auto& metrics_def = el.value();
+                for (const auto& metric_def : metrics_def) {
+                    std::string name = metric_def["name"];
+                    std::string help = metric_def["help"];
+                    prometheus << "# HELP llamacpp:" << name << " " << help                << "\n"
+                               << "# TYPE llamacpp:" << name << " " << type                << "\n"
+                               << "llamacpp:"        << name << " " << metric_def["value"] << "\n";
+                }
+            }
+
+            res.set_content(prometheus.str(), "text/plain; version=0.0.4");
+            res.status = 200; // HTTP OK
+        });
+    }
+
     svr.set_logger(log_server_request);
 
     svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@@ -2735,9 +3018,6 @@ int main(int argc, char **argv)
     // Set the base directory for serving static files
     svr.set_base_dir(sparams.public_path);
 
-    // to make it ctrl+clickable:
-    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
-
     std::unordered_map<std::string, std::string> log_data;
     log_data["hostname"] = sparams.hostname;
     log_data["port"] = std::to_string(sparams.port);

examples/server/tests/README.md

@@ -32,6 +32,7 @@ It's possible to override some scenario steps values with environment variables:
  - `PORT` -> `context.server_port` to set the listening port of the server during scenario, default: `8080`
  - `LLAMA_SERVER_BIN_PATH` -> to change the server binary path, default: `../../../build/bin/server`
  - `DEBUG` -> "ON" to enable steps and server verbose mode `--verbose`
+ - `SERVER_LOG_FORMAT_JSON` -> if set switch server logs to json format
 
 ### Run @bug, @wip or @wrong_usage annotated scenario
 

examples/server/tests/features/environment.py

@@ -16,6 +16,8 @@ def before_scenario(context, scenario):
 
 
 def after_scenario(context, scenario):
+    if context.server_process is None:
+        return
     if scenario.status == "failed":
         if 'GITHUB_ACTIONS' in os.environ:
             print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")

examples/server/tests/features/server.feature

@@ -13,6 +13,7 @@ Feature: llama.cpp server
     And   1 slots
     And   embeddings extraction
     And   32 server max tokens to predict
+    And   prometheus compatible metrics exposed
     Then  the server is starting
     Then  the server is healthy
 
@@ -25,11 +26,12 @@ Feature: llama.cpp server
     And   <n_predict> max tokens to predict
     And   a completion request with no api error
     Then  <n_predicted> tokens are predicted matching <re_content>
+    And   prometheus metrics are exposed
 
     Examples: Prompts
-      | prompt                           | n_predict | re_content                   | n_predicted |
-      | I believe the meaning of life is | 8         | read                         | 8           |
-      | Write a joke about AI            | 64        | (park<or>friends<or>scared)+ | 32          |
+      | prompt                           | n_predict | re_content                             | n_predicted |
+      | I believe the meaning of life is | 8         | (read<or>going)+                       | 8           |
+      | Write a joke about AI            | 64        | (park<or>friends<or>scared<or>always)+ | 32          |
 
   Scenario Outline: OAI Compatibility
     Given a model <model>

examples/server/tests/features/steps/steps.py

@@ -13,6 +13,7 @@ import aiohttp
 import openai
 from behave import step
 from behave.api.async_step import async_run_until_complete
+from prometheus_client import parser
 
 
 @step(u"a server listening on {server_fqdn}:{server_port}")
@@ -34,6 +35,8 @@ def step_server_config(context, server_fqdn, server_port):
     context.server_api_key = None
     context.server_continuous_batching = False
     context.server_embeddings = False
+    context.server_metrics = False
+    context.server_process = None
     context.server_seed = None
     context.user_api_key = None
 
@@ -82,6 +85,11 @@ def step_server_embeddings(context):
     context.server_embeddings = True
 
 
+@step(u'prometheus compatible metrics exposed')
+def step_server_metrics(context):
+    context.server_metrics = True
+
+
 @step(u"the server is starting")
 def step_start_server(context):
     start_server_background(context)
@@ -424,6 +432,23 @@ def step_check_options_header_value(context, cors_header, cors_header_value):
     assert context.options_response.headers[cors_header] == cors_header_value
 
 
+@step(u'prometheus metrics are exposed')
+@async_run_until_complete
+async def step_prometheus_metrics_exported(context):
+    async with aiohttp.ClientSession() as session:
+        async with await session.get(f'{context.base_url}/metrics') as metrics_response:
+            assert metrics_response.status == 200
+            assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
+            metrics_raw = await metrics_response.text()
+            metric_exported = False
+            for metric in parser.text_string_to_metric_families(metrics_raw):
+                match metric.name:
+                    case "llamacpp:kv_cache_usage_ratio":
+                        assert len(metric.samples) > 0
+                        metric_exported = True
+            assert metric_exported, "No metrics exported"
+
+
 async def concurrent_requests(context, f_completion, *args, **kwargs):
     n_prompts = len(context.prompts)
     if context.debug:
@@ -674,6 +699,8 @@ async def wait_for_health_status(context,
     if context.debug:
         print(f"Starting checking for health for expected_health_status={expected_health_status}")
     timeout = 3  # seconds
+    if expected_health_status == 'ok':
+        timeout = 10 # CI slow inference
     interval = 0.5
     counter = 0
     async with aiohttp.ClientSession() as session:
@@ -711,7 +738,7 @@ async def wait_for_health_status(context,
                         if n_completions > 0:
                             return
 
-                assert False, 'timeout exceeded'
+                assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
 
 
 def assert_embeddings(embeddings):
@@ -753,6 +780,8 @@ def start_server_background(context):
         server_args.append('--cont-batching')
     if context.server_embeddings:
         server_args.append('--embedding')
+    if context.server_metrics:
+        server_args.append('--metrics')
     if context.model_alias is not None:
         server_args.extend(['--alias', context.model_alias])
     if context.n_ctx is not None:
@@ -765,6 +794,8 @@ def start_server_background(context):
         server_args.extend(['--api-key', context.server_api_key])
     if context.debug:
         server_args.append('--verbose')
+    if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
+        server_args.extend(['--log-format', "text"])
     print(f"starting server with: {context.server_path}", *server_args)
     context.server_process = subprocess.Popen(
         [str(arg) for arg in [context.server_path, *server_args]],

examples/server/tests/requirements.txt

@@ -1,3 +1,4 @@
 aiohttp~=3.9.3
 behave~=1.2.6
 openai~=0.25.0
+prometheus-client~=0.20.0

examples/server/utils.hpp

@@ -14,6 +14,7 @@
 using json = nlohmann::json;
 
 extern bool server_verbose;
+extern bool server_log_json;
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -27,14 +28,14 @@ extern bool server_verbose;
     {                                                                    \
         if (server_verbose)                                              \
         {                                                                \
-            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
         }                                                                \
     } while (0)
 #endif
 
-#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_ERROR(  MSG, ...) server_log("ERR",  __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
 
 //
 // parallel
@@ -50,7 +51,7 @@ enum task_type {
     TASK_TYPE_COMPLETION,
     TASK_TYPE_CANCEL,
     TASK_TYPE_NEXT_RESPONSE,
-    TASK_TYPE_SLOTS_DATA
+    TASK_TYPE_METRICS
 };
 
 struct task_server {
@@ -133,26 +134,48 @@ struct completion_token_output
     std::string text_to_send;
 };
 
-static inline void server_log(const char *level, const char *function, int line,
-                       const char *message, const nlohmann::ordered_json &extra)
+static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra)
 {
-    nlohmann::ordered_json log
-    {
+    std::stringstream ss_tid;
+    ss_tid << std::this_thread::get_id();
+    json log = nlohmann::ordered_json{
+        {"tid", ss_tid.str()},
         {"timestamp", time(nullptr)},
-        {"level",     level},
-        {"function",  function},
-        {"line",      line},
-        {"message",   message},
     };
 
-    if (!extra.empty())
-    {
-        log.merge_patch(extra);
-    }
+    if (server_log_json) {
+        log.merge_patch(
+                {
+                        {"level",     level},
+                        {"function",  function},
+                        {"line",      line},
+                        {"msg",       message},
+                });
+        if (!extra.empty()) {
+            log.merge_patch(extra);
+        }
 
-    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
-    printf("%.*s\n", (int)str.size(), str.data());
-    fflush(stdout);
+        std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
+    } else {
+        char buf[1024];
+        snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
+
+        if (!extra.empty()) {
+            log.merge_patch(extra);
+        }
+        std::stringstream ss;
+        ss << buf << " |";
+        for (const auto& el : log.items())
+        {
+            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
+            snprintf(buf, 1024, " %s=%s", el.key().c_str(), value.c_str());
+            ss << buf;
+        }
+
+        const std::string str = ss.str();
+        printf("%.*s\n", (int)str.size(), str.data());
+        fflush(stdout);
+    }
 }
 
 //
@@ -234,6 +257,7 @@ struct llama_server_queue {
         std::unique_lock<std::mutex> lock(mutex_tasks);
         if (task.id == -1) {
             task.id = id++;
+            LOG_VERBOSE("new task id", {{"new_id", task.id}});
         }
         queue_tasks.push_back(std::move(task));
         condition_tasks.notify_one();
@@ -249,7 +273,9 @@ struct llama_server_queue {
     // Get the next id for creating anew task
     int get_new_id() {
         std::unique_lock<std::mutex> lock(mutex_tasks);
-        return id++;
+        int new_id = id++;
+        LOG_VERBOSE("new task id", {{"new_id", new_id}});
+        return new_id;
     }
 
     // Register function to process a new task
@@ -290,8 +316,7 @@ struct llama_server_queue {
     void start_loop() {
         running = true;
         while (true) {
-            // new task arrived
-            LOG_VERBOSE("have new task", {});
+            LOG_VERBOSE("new task may arrive", {});
             {
                 while (true)
                 {
@@ -303,7 +328,7 @@ struct llama_server_queue {
                     task_server task = queue_tasks.front();
                     queue_tasks.erase(queue_tasks.begin());
                     lock.unlock();
-                    LOG_VERBOSE("callback_new_task", {});
+                    LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
                     callback_new_task(task);
                 }
                 LOG_VERBOSE("callback_all_task_finished", {});
@@ -384,11 +409,13 @@ struct llama_server_response {
     std::condition_variable condition_results;
 
     void add_waiting_task_id(int task_id) {
+        LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.insert(task_id);
     }
 
     void remove_waiting_task_id(int task_id) {
+        LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
         std::unique_lock<std::mutex> lock(mutex_results);
         waiting_task_ids.erase(task_id);
     }
@@ -401,7 +428,6 @@ struct llama_server_response {
             condition_results.wait(lock, [&]{
                 return !queue_results.empty();
             });
-            LOG_VERBOSE("condition_results unblock", {});
 
             for (int i = 0; i < (int) queue_results.size(); i++)
             {
@@ -426,22 +452,22 @@ struct llama_server_response {
     // Send a new result to a waiting task_id
     void send(task_result result) {
         std::unique_lock<std::mutex> lock(mutex_results);
-        LOG_VERBOSE("send new result", {});
+        LOG_VERBOSE("send new result", {{"task_id", result.id}});
         for (auto& task_id : waiting_task_ids) {
             // LOG_TEE("waiting task id %i \n", task_id);
             // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
             if (result.multitask_id == task_id)
             {
-                LOG_VERBOSE("callback_update_multitask", {});
+                LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
                 callback_update_multitask(task_id, result.id, result);
                 continue;
             }
 
             if (result.id == task_id)
             {
-                LOG_VERBOSE("queue_results.push_back", {});
+                LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
                 queue_results.push_back(result);
-                condition_results.notify_one();
+                condition_results.notify_all();
                 return;
             }
         }

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1708118438,
-        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
+        "lastModified": 1708655239,
+        "narHash": "sha256-ZrP/yACUvDB+zbqYJsln4iwotbH6CTZiTkANJ0AgDv4=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
+        "rev": "cbc4211f0afffe6dfd2478a62615dd5175a13f9a",
         "type": "github"
       },
       "original": {

ggml-quants.c

@@ -462,6 +462,30 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
     return res;
 }
 
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
 #else
 
 #define ggml_int16x8x2_t  int16x8x2_t
@@ -476,6 +500,7 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
 #define ggml_vld1q_s8_x2  vld1q_s8_x2
 #define ggml_vld1q_s8_x4  vld1q_s8_x4
 #define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
 
 #endif
 
@@ -9488,8 +9513,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
             qs += 16;
 
             vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | (signs[1] << 16)));
-            vs.val[1] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vceqq_u8(vs.val[0], mask2);
             vs.val[1] = vceqq_u8(vs.val[1], mask2);
 
@@ -9497,8 +9522,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
             q3s.val[1] = vsubq_s8(vreinterpretq_s8_u8(veorq_u8(vs.val[1], vreinterpretq_u8_u32(aux32x4_1))), vreinterpretq_s8_u8(vs.val[1]));
 
             vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | (signs[3] << 16)));
-            vs.val[1] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
             vs.val[0] = vceqq_u8(vs.val[0], mask2);
             vs.val[1] = vceqq_u8(vs.val[1], mask2);
 

ggml-sycl.cpp

@@ -8126,23 +8126,51 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-static void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale,
-                         const sycl::nd_item<3> &item_ct1, float *buf) {
+
+template <bool vals_smem, int ncols_template, int block_size_template>
+static void soft_max_f32(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
+                         const int nrows_y, const float scale, const float max_bias, const float m0,
+                         const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
     const int tid = item_ct1.get_local_id(2);
     const int rowx = item_ct1.get_group(2);
     const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
 
-    const int block_size = item_ct1.get_local_range(2);
+    const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template;
 
     const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
     const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
 
+    float slope = 0.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const uint32_t h = rowx/nrows_y; // head index
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = sycl::pow(base, float(exp));
+    }
+
+    float * vals = vals_smem ? buf + WARP_SIZE : dst + rowx*ncols;
     float max_val = -INFINITY;
 
-    for (int col = tid; col < ncols; col += block_size) {
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
-        max_val = sycl::max(max_val, x[ix] * scale + (y ? y[iy] : 0.0f));
+
+        const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
+
+        vals[col] = val;
+        max_val = sycl::max(max_val, val);
     }
 
     // find the max value in the block
@@ -8151,30 +8179,12 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
         if (warp_id == 0) {
             buf[lane_id] = -INFINITY;
         }
-        /*
-        DPCT1118:12: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
-        /*
-        DPCT1065:60: Consider replacing sycl::nd_item::barrier() with
-        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
-        better performance if there is no access to global memory.
-        */
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
         if (lane_id == 0) {
             buf[warp_id] = max_val;
         }
-        /*
-        DPCT1118:13: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
-        /*
-        DPCT1065:61: Consider replacing sycl::nd_item::barrier() with
-        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
-        better performance if there is no access to global memory.
-        */
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
         max_val = buf[lane_id];
         max_val = warp_reduce_max(max_val, item_ct1);
@@ -8182,13 +8192,16 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
 
     float tmp = 0.f;
 
-    for (int col = tid; col < ncols; col += block_size) {
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-        const float val =
-            sycl::native::exp((x[ix] * scale + (y ? y[iy] : 0.0f)) - max_val);
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+                if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const float val = sycl::native::exp(vals[col] - max_val);
         tmp += val;
-        dst[ix] = val;
+        vals[col] = val;
     }
 
     // find the sum of exps in the block
@@ -8197,40 +8210,29 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
         if (warp_id == 0) {
             buf[lane_id] = 0.f;
         }
-        /*
-        DPCT1118:14: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
-        /*
-        DPCT1065:62: Consider replacing sycl::nd_item::barrier() with
-        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
-        better performance if there is no access to global memory.
-        */
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
         if (lane_id == 0) {
             buf[warp_id] = tmp;
         }
-        /*
-        DPCT1118:15: SYCL group functions and algorithms must be encountered in
-        converged control flow. You may need to adjust the code.
-        */
-        /*
-        DPCT1065:63: Consider replacing sycl::nd_item::barrier() with
-        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
-        better performance if there is no access to global memory.
-        */
-        item_ct1.barrier();
+        item_ct1.barrier(sycl::access::fence_space::local_space);
 
         tmp = buf[lane_id];
         tmp = warp_reduce_sum(tmp, item_ct1);
     }
 
-    const float inv_tmp = 1.f / tmp;
+    const float inv_sum = 1.f / tmp;
 
-    for (int col = tid; col < ncols; col += block_size) {
-        const int i = rowx*ncols + col;
-        dst[i] *= inv_tmp;
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            return;
+        }
+
+        const int idst = rowx*ncols + col;
+        dst[idst] = vals[col] * inv_sum;
     }
 }
 
@@ -10867,35 +10869,96 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst,
                          });
 }
 
-static void soft_max_f32_sycl(const float *x, const float *y, float *dst,
-                              const int ncols_x, const int nrows_x,
-                              const int nrows_y, const float scale,
+template <bool vals_smem, int ncols_template, int block_size_template>
+static void soft_max_f32_submitter(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
+                                   const int nrows_y, const float scale, const float max_bias, const float m0,
+                                   const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
+                                   const size_t n_local_scratch, dpct::queue_ptr stream) {
+    stream->submit([&](sycl::handler &cgh) {
+        sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
+                soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, pos, dst, ncols_par,
+                                                                             nrows_y, scale, max_bias, m0,
+                                                                             m1, n_head_log2, item_ct1,
+                                                                             local_buf_acc.get_pointer());
+            });
+    });
+}
+
+static void soft_max_f32_sycl(const float * x, const float * mask, const float * pos,
+                              float * dst, const int ncols_x, const int nrows_x,
+                              const int nrows_y, const float scale, const float max_bias,
                               dpct::queue_ptr stream) {
     int nth = WARP_SIZE;
     while (nth < ncols_x && nth < SYCL_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, 1, nrows_x);
-    /*
-    DPCT1049:46: The work-group size passed to the SYCL kernel may exceed the
-    limit. To get the device limit, query info::device::max_work_group_size.
-    Adjust the work-group size if needed.
-    */
-    stream->submit([&](sycl::handler &cgh) {
-        /*
-        DPCT1101:96: 'SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE' expression was
-        replaced with a value. Modify the code to use the original expression,
-        provided in comments, if it is correct.
-        */
-        sycl::local_accessor<float, 1> buf_acc_ct1(
-            sycl::range<1>(32 /*SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE*/), cgh);
+    const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE);
+    static_assert(SYCL_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-        cgh.parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-                soft_max_f32(x, y, dst, ncols_x, nrows_y, scale, item_ct1,
-                             buf_acc_ct1.get_pointer());
-            });
-    });
+    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
+    if (n_local_scratch*sizeof(float) < local_mem_size) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f32_submitter<true, 32, 32>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                     max_bias, m0, m1, n_head_log2, block_nums,
+                                                     block_dims, n_local_scratch, stream);
+                break;
+            case 64:
+                soft_max_f32_submitter<true, 64, 64>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                     max_bias, m0, m1, n_head_log2, block_nums,
+                                                     block_dims, n_local_scratch, stream);
+                break;
+            case 128:
+                soft_max_f32_submitter<true, 128, 128>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 256:
+                soft_max_f32_submitter<true, 256, 256>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 512:
+                soft_max_f32_submitter<true, 512, 512>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 1024:
+                soft_max_f32_submitter<true, 1024, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            case 2048:
+                soft_max_f32_submitter<true, 2048, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            case 4096:
+                soft_max_f32_submitter<true, 4096, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            default:
+                soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                                   max_bias, m0, m1, n_head_log2, block_nums,
+                                                   block_dims, n_local_scratch, stream);
+                break;
+        }
+    } else {
+        soft_max_f32_submitter<false, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                                            max_bias, m0, m1, n_head_log2, block_nums,
+                                            block_dims, WARP_SIZE, stream);
+    }
 }
 
 template <typename T>
@@ -12435,14 +12498,35 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
 
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
-    const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
+    const int64_t nrows_y = src0->ne[1];
 
     float scale = 1.0f;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float max_bias = 0.0f;
 
-    soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+    memcpy(&scale, dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, dst->op_params + 1, sizeof(float));
 
-    (void) dst;
+    // positions tensor
+    float * src2_dd = nullptr;
+    sycl_pool_alloc<float> src2_f;
+
+    ggml_tensor * src2 = dst->src[2];
+    const bool use_src2 = src2 != nullptr;
+
+    if (use_src2) {
+        const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
+
+        if (src2_on_device) {
+            ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
+            src2_dd = (float *) src2_extra->data_device[g_main_device];
+        } else {
+            src2_dd = src2_f.alloc(ggml_nelements(src2));
+            SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
+        }
+    }
+
+    soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00,
+                      nrows_x, nrows_y, scale, max_bias, main_stream);
 }
 
 inline void ggml_sycl_op_scale(const ggml_tensor *src0, const ggml_tensor *src1,

llama.h

@@ -64,6 +64,15 @@ extern "C" {
         LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
     };
 
+    // note: these values should be synchronized with ggml_rope
+    // TODO: maybe move this enum to ggml.h (ggml_rope_type)
+    enum llama_rope_type {
+        LLAMA_ROPE_TYPE_NONE = -1,
+        LLAMA_ROPE_TYPE_NORM =  0,
+        LLAMA_ROPE_TYPE_NEOX =  2,
+        LLAMA_ROPE_TYPE_GLM  =  4,
+    };
+
     enum llama_token_type {
         LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
         LLAMA_TOKEN_TYPE_NORMAL       = 1,
@@ -360,6 +369,7 @@ extern "C" {
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
 
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);
 
     LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
     LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@@ -514,10 +524,12 @@ extern "C" {
                     llama_seq_id   seq_id);
 
     // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_cache_seq_shift(
+    LLAMA_API void llama_kv_cache_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -525,7 +537,9 @@ extern "C" {
                        llama_pos   delta);
 
     // Integer division of the positions by factor of `d > 1`
-    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // If the KV cache is RoPEd, the KV data is updated accordingly:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
     LLAMA_API void llama_kv_cache_seq_div(
@@ -535,6 +549,20 @@ extern "C" {
                        llama_pos   p1,
                              int   d);
 
+    // Returns the largest position present in the KV cache for the specified sequence
+    LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id);
+
+    // Defragment the KV cache
+    // This will be applied:
+    //   - lazily on next llama_decode()
+    //   - explicitly with llama_kv_cache_update()
+    LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
+
+    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
+    LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
+
     //
     // State / sessions
     //

unicode.h

@@ -404,7 +404,8 @@ static std::unordered_map<uint32_t, int> codepoint_type_map() {
 
 static int codepoint_type(uint32_t cp) {
     static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
-    return codepoint_types.find(cp) == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : codepoint_types.at(cp);
+    const auto it = codepoint_types.find(cp);
+    return it == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
 }
 
 static int codepoint_type(const std::string & utf8) {