docker : add server-first container images (#5157)

* feat: add Dockerfiles for each platform that user ./server instead of ./main

* feat: update .github/workflows/docker.yml to build server-first docker containers

* doc: add information about running the server with Docker to README.md

* doc: add information about running with docker to the server README

* doc: update n-gpu-layers to show correct GPU usage

* fix(doc): update container tag from `server` to `server-cuda` for README example on running server container with CUDA

.devops/server-cuda.Dockerfile

@@ -0,0 +1,32 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG CUDA_VERSION=11.7.1
+# Target the CUDA build image
+ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the CUDA runtime image
+ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
+
+RUN make
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
+
+COPY --from=build /app/server /server
+
+ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -0,0 +1,25 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+ARG UBUNTU_VERSION=22.04
+
+FROM intel/hpckit:$ONEAPI_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
+    cmake --build . --config Release --target main server
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -0,0 +1,45 @@
+ARG UBUNTU_VERSION=22.04
+
+# This needs to generally match the container host's environment.
+ARG ROCM_VERSION=5.6
+
+# Target the CUDA build image
+ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
+
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
+
+# Unless otherwise specified, we make a fat build.
+# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# This is mostly tied to rocBLAS supported archs.
+ARG ROCM_DOCKER_ARCH=\
+    gfx803 \
+    gfx900 \
+    gfx906 \
+    gfx908 \
+    gfx90a \
+    gfx1010 \
+    gfx1030 \
+    gfx1100 \
+    gfx1101 \
+    gfx1102
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+# Set nvcc architecture
+ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+# Enable ROCm
+ENV LLAMA_HIPBLAS=1
+ENV CC=/opt/rocm/llvm/bin/clang
+ENV CXX=/opt/rocm/llvm/bin/clang++
+
+RUN make
+
+ENTRYPOINT [ "/app/server" ]

.devops/server.Dockerfile

@@ -0,0 +1,20 @@
+ARG UBUNTU_VERSION=22.04
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git
+
+WORKDIR /app
+
+COPY . .
+
+RUN make
+
+FROM ubuntu:$UBUNTU_VERSION as runtime
+
+COPY --from=build /app/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.github/workflows/docker.yml

@@ -28,14 +28,18 @@ jobs:
         config:
           - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
           #                     have disabled them for now until the reason why
           #                     is understood.
           - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3

README.md

@@ -931,17 +931,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 * Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
-We have two Docker images available for this project:
+We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 
 - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
@@ -967,6 +970,12 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
 ### Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -976,6 +985,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
 docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -989,6 +999,7 @@ The resulting images, are essentially the same as the non-CUDA images:
 
 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
 
 #### Usage
 
@@ -997,6 +1008,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ### Contributing

common/sampling.cpp

@@ -13,6 +13,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
         // will be empty (default) if there are parse errors
         if (result->parsed_grammar.rules.empty()) {
             fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            delete result;
             return nullptr;
         }
 

examples/infill/infill.cpp

@@ -241,7 +241,7 @@ int main(int argc, char ** argv) {
     LOG("add_bos: %d\n", add_bos);
 
     bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+    if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
         params.input_suffix.erase(0, 1);
         suff_rm_leading_spc = false;
     }

examples/llava/clip.cpp

@@ -98,6 +98,7 @@ static std::string format(const char * fmt, ...) {
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
+    PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_UNKNOWN,
 };
@@ -304,10 +305,18 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
-    struct ggml_tensor * mm_0_w;
-    struct ggml_tensor * mm_0_b;
-    struct ggml_tensor * mm_2_w;
-    struct ggml_tensor * mm_2_b;
+    struct ggml_tensor * mm_0_w = NULL;
+    struct ggml_tensor * mm_0_b = NULL;
+    struct ggml_tensor * mm_2_w = NULL;
+    struct ggml_tensor * mm_2_b = NULL;
+
+    // Yi type models with mlp+normalization projection
+    struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4
+    struct ggml_tensor * mm_1_b = NULL;
+    struct ggml_tensor * mm_3_w = NULL;
+    struct ggml_tensor * mm_3_b = NULL;
+    struct ggml_tensor * mm_4_w = NULL;
+    struct ggml_tensor * mm_4_b = NULL;
 
     // MobileVLM projection
     struct ggml_tensor * mm_model_mlp_1_w;
@@ -460,6 +469,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     // pre-layernorm
     {
         embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "pre_ln");
 
         embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
@@ -575,6 +585,27 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+
+        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+            embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
+            // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
+            // First LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w),
+                                model.mm_1_b);
+
+            // GELU activation
+            embeddings = ggml_gelu(ctx0, embeddings);
+
+            // Second linear layer
+            embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings);
+            embeddings = ggml_add(ctx0, embeddings, model.mm_3_b);
+
+            // Second LayerNorm
+            embeddings = ggml_norm(ctx0, embeddings, eps);
+            embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w),
+                                model.mm_4_b);
         }
         else if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projector
@@ -808,6 +839,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         else {
             new_clip->proj_type = PROJECTOR_TYPE_MLP;
         }
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+            if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) {
+                new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM;
+            }
+        }
     }
 
 #ifdef GGML_USE_CUBLAS
@@ -956,11 +992,29 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
 
         // LLaVA projection
-        if (new_clip->proj_type == PROJECTOR_TYPE_MLP) {
+        if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
             vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
-            vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
-            vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            try {
+                // Yi-type llava
+                vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight"));
+                vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // missing in Yi-type llava
+                vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+                vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight"));
+                vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias"));
+            } catch (std::runtime_error & e) {  }
+            try {
+                // Yi-type llava
+                vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight"));
+                vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias"));
+            } catch (std::runtime_error & e) {  }
         }
         else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
@@ -1277,7 +1331,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
         ".*weight",
     };
 
-    std::vector<uint8_t> read_data(512);
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
     std::vector<int64_t> hist_all(1 << 4, 0);
@@ -1433,6 +1486,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     }
     else if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
+    } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        return ctx->vision_model.mm_3_b->ne[0];
     }
     else {
         std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];

examples/llava/llava-cli.cpp

@@ -148,10 +148,35 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
 
-    // llava chat format is "<system_prompt>\nUSER:<image_embeddings>\n<textual_prompt>\nASSISTANT:"
-    eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant.  The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos);
+    std::string system_prompt, user_prompt;
+    size_t image_pos = prompt.find("<image>");
+    if (image_pos != std::string::npos) {
+        // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
+
+        system_prompt = prompt.substr(0, image_pos);
+        user_prompt = prompt.substr(image_pos + std::string("<image>").length());
+        // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string
+        size_t pos = 0;
+        while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) {
+            user_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+        while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) {
+            system_prompt.replace(pos, 2, "\n");
+            pos += 1; // Advance past the replaced newline
+        }
+
+        printf("system_prompt: %s\n", system_prompt.c_str());
+        printf("user_prompt: %s\n", user_prompt.c_str());
+    } else {
+        // llava-1.5 native mode
+        system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
+        user_prompt = prompt + "\nASSISTANT:";
+    }
+
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
     llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
-    eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false);
+    eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
 
     // generate the response
 
@@ -162,6 +187,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
         if (strcmp(tmp, "</s>") == 0) break;
+        if (strstr(tmp, "###")) break; // Yi-VL behavior
 
         printf("%s", tmp);
         fflush(stdout);

examples/server/README.md

@@ -30,7 +30,8 @@ Command line options:
 -   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
 -   `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 -   `--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`
 ## Build
 
 server is build alongside everything else from the root of the project
@@ -65,6 +66,14 @@ server.exe -m models\7B\ggml-model.gguf -c 2048
 The above command will start a server that by default listens on `127.0.0.1:8080`.
 You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url.
 
+### Docker:
+```bash
+docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+
+# or, with CUDA:
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+```
+
 ## Testing with CURL
 
 Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.

examples/server/server.cpp

@@ -184,6 +184,12 @@ struct llama_client_slot
     struct llama_sampling_params sparams;
     llama_sampling_context *ctx_sampling = nullptr;
 
+    int32_t ga_i = 0;   // group-attention state
+    int32_t ga_n = 1;// group-attention factor
+    int32_t ga_w = 512; // group-attention width
+
+    int32_t n_past_se = 0; // self-extend
+
     // multimodal
     std::vector<slot_image> images;
 
@@ -212,7 +218,8 @@ struct llama_client_slot
         sent_count             = 0;
         sent_token_probs_index = 0;
         infill                 = false;
-
+        ga_i                   = 0;
+        n_past_se  = 0;
         generated_token_probs.clear();
 
         for (slot_image & img : images)
@@ -399,9 +406,26 @@ struct llama_server_context
 
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
-            slot.reset();
 
             LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
+
+            const int ga_n = params.grp_attn_n;
+            const int ga_w = params.grp_attn_w;
+
+            if (ga_n != 1) {
+                GGML_ASSERT(ga_n > 0                    && "ga_n must be positive");                     // NOLINT
+                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);
+            }
+
+            slot.ga_i = 0;
+            slot.ga_n = ga_n;
+            slot.ga_w = ga_w;
+
+            slot.reset();
+
             slots.push_back(slot);
         }
 
@@ -657,7 +681,7 @@ struct llama_server_context
                     while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
                         size_t end_prefix = pos;
                         pos += pattern.length();
-                        size_t end_pos = prompt.find("]", pos);
+                        size_t end_pos = prompt.find(']', pos);
                         if (end_pos != std::string::npos)
                         {
                             std::string image_id = prompt.substr(pos, end_pos - pos);
@@ -1349,32 +1373,35 @@ struct llama_server_context
 
         for (llama_client_slot &slot : slots)
         {
-            if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx)
+            if (slot.ga_n == 1)
             {
-                // Shift context
-                const int n_left    = slot.n_past - slot.params.n_keep - 1;
-                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, slot.params.n_keep, n_left, n_discard);
-                llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard);
-
-                for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
+                if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
-                    slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
+                    // Shift context
+                    const int n_left    = slot.n_past - slot.params.n_keep - 1;
+                    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, slot.params.n_keep, n_left, n_discard);
+                    llama_kv_cache_seq_rm   (ctx, slot.id, slot.params.n_keep + 1            , slot.params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard);
+
+                    for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++)
+                    {
+                        slot.cache_tokens[i - n_discard] = slot.cache_tokens[i];
+                    }
+
+                    slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard);
+
+                    slot.n_past -= n_discard;
+
+                    slot.truncated = true;
+
+                    LOG_VERBOSE("context shift", {
+                        { "n_ctx", n_ctx },
+                        { "n_keep", params.n_keep },
+                        { "n_left", n_left },
+                    });
                 }
-
-                slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard);
-
-                slot.n_past -= n_discard;
-
-                slot.truncated = true;
-
-                LOG_VERBOSE("context shift", {
-                                                {"n_ctx",  n_ctx},
-                                                {"n_keep", params.n_keep},
-                                                {"n_left", n_left},
-                                            });
             }
         }
 
@@ -1401,7 +1428,8 @@ struct llama_server_context
 
             slot.i_batch = batch.n_tokens;
 
-            llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true);
+            const int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
+            llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true);
 
             slot.n_past += 1;
         }
@@ -1499,6 +1527,8 @@ struct llama_server_context
                         llama_sampling_reset(slot.ctx_sampling);
 
                         slot.n_past = 0;
+                        slot.n_past_se = 0;
+                        slot.ga_i = 0;
                         slot.num_prompt_tokens_processed = slot.num_prompt_tokens;
                     }
                     else
@@ -1512,6 +1542,25 @@ struct llama_server_context
                         slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
                         slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
 
+                        if (slot.ga_n != 1)
+                        {
+                            int ga_i = 0;
+                            int32_t ga_n = slot.ga_n;
+                            int32_t ga_w = slot.ga_w;
+                            int32_t slot_npast = 0;
+                            for (int k = 0; k < slot.n_past; ++k)
+                            {
+                                while (slot_npast >= ga_i + ga_w) {
+                                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                                    slot_npast -= bd;
+                                    ga_i += ga_w/ga_n;
+                                }
+                                slot_npast++;
+                            }
+                            slot.n_past_se = slot_npast;
+                            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);
                     }
 
@@ -1526,6 +1575,10 @@ struct llama_server_context
                         // 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);
                         slot.n_past--;
+                        if (slot.ga_i > 0)
+                        {
+                            slot.n_past_se--;
+                        }
                     }
 
                     LOG_VERBOSE("prompt ingested", {
@@ -1538,9 +1591,22 @@ struct llama_server_context
 
                     // process the prefix of first image
                     std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
+                    int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
+                    int ga_i = slot.ga_i;
+                    int32_t ga_n = slot.ga_n;
+                    int32_t ga_w = slot.ga_w;
                     for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past)
                     {
-                       llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false);
+                        if (slot.ga_n != 1)
+                        {
+                            while (slot_npast >= ga_i + ga_w) {
+                                const int bd = (ga_w/ga_n)*(ga_n - 1);
+                                slot_npast -= bd;
+                                ga_i += ga_w/ga_n;
+                            }
+                        }
+                        llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot_npast, {slot.id }, false);
+                        slot_npast += 1;
                     }
 
                     if (has_images && !ingest_images(slot, n_batch))
@@ -1570,6 +1636,36 @@ struct llama_server_context
         for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
         {
             const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            for (auto & slot : slots)
+            {
+                if (slot.ga_n != 1)
+                {
+                    // context extension via Self-Extend
+                    while (slot.n_past_se >= slot.ga_i + slot.ga_w)
+                    {
+                        const int ib = (slot.ga_n * slot.ga_i) / slot.ga_w;
+                        const int bd = (slot.ga_w / slot.ga_n) * (slot.ga_n - 1);
+                        const int dd = (slot.ga_w / slot.ga_n) - ib * bd - slot.ga_w;
+
+                        LOG_TEE("\n");
+                        LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i, slot.n_past_se, ib * bd, slot.ga_i + ib * bd, slot.n_past_se + ib * bd);
+                        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_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);
+
+                        slot.n_past_se -= bd;
+
+                        slot.ga_i += slot.ga_w / slot.ga_n;
+
+                        LOG_TEE("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", slot.n_past_se + bd, slot.n_past_se, slot.ga_i);
+                    }
+                    slot.n_past_se += n_tokens;
+                }
+            }
             llama_batch batch_view =
             {
                 n_tokens,
@@ -1583,6 +1679,7 @@ struct llama_server_context
             };
 
             const int ret = llama_decode(ctx, batch_view);
+
             if (ret != 0)
             {
                 if (n_batch == 1 || ret < 0)
@@ -1728,6 +1825,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
+    printf("  -gan N, --grp-attn-n 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`");
+    printf("  -gaw N, --grp-attn-w N    Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
     printf("\n");
 }
 
@@ -1913,6 +2012,25 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_threads = std::stoi(argv[i]);
         }
+        else if (arg == "--grp-attn-n" || arg == "-gan")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_n = std::stoi(argv[i]);
+        }
+        else if (arg == "--grp-attn-w" || arg == "-gaw")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_w = std::stoi(argv[i]);
+        }
         else if (arg == "--threads-batch" || arg == "-tb")
         {
             if (++i >= argc)

examples/server/utils.hpp

@@ -249,6 +249,7 @@ struct llama_server_queue {
     }
 
     // Start the main loop. This call is blocking
+    [[noreturn]]
     void start_loop() {
         while (true) {
             // new task arrived

ggml.c

@@ -7498,7 +7498,12 @@ static void ggml_compute_forward_add(
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add_f32(params, src0, src1, dst);
+                if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add_f32(params, src0, src1, dst);
+                }
+                else {
+                    GGML_ASSERT(false);
+                }
             } break;
         case GGML_TYPE_F16:
             {

pocs/vdot/vdot.cpp

@@ -243,7 +243,6 @@ int main(int argc, char** argv) {
     if (useQ4_1) q41.resize(n4);
     else q40.resize(n4);
     std::vector<block_q8_0> q8(n8);
-    std::vector<int64_t> H(16, 0);
     double sumt = 0, sumt2 = 0, maxt = 0;
     double sumqt = 0, sumqt2 = 0, maxqt = 0;
     double sum = 0, sumq = 0, exactSum = 0;

scripts/sync-ggml.last

@@ -1 +1 @@
-6c1ce0bd591a430c1d3f6797d905194581c878c1
+c2448f88d17395452a587d0176d19ed87e0f7ce1

tests/test-backend-ops.cpp

@@ -102,7 +102,6 @@ static std::vector<float> tensor_to_float(const ggml_tensor * t) {
                     } else if (t->type == GGML_TYPE_I8) {
                         tv.push_back((float)*(int8_t *) &buf[i]);
                     } else if (quantized) {
-                        std::vector<float> vq(ggml_blck_size(t->type));
                         tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type));
                         tv.insert(tv.end(), vq.begin(), vq.end());
                     } else {

tests/test-llama-grammar.cpp

@@ -190,7 +190,6 @@ int main()
         index++;
     }
 
-    std::vector<std::vector<const llama_grammar_element *>> next_stacks;
     std::vector<llama_grammar_candidate> next_candidates;
     next_candidates.resize(24);