Introduce llama-run (#10291)

It's like simple-chat but it uses smart pointers to avoid manual
memory cleanups. Less memory leaks in the code now. Avoid printing
multiple dots. Split code into smaller functions. Uses no exception
handling.

Signed-off-by: Eric Curtin <ecurtin@redhat.com>

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -24,7 +24,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -41,14 +42,17 @@ body:
         description: Which GGML backends do you know to be affected?
         options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
+    validations:
+      required: true
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific compile flags, that information would be very much appreciated by us.
       placeholder: >
+        I'm trying to compile llama.cpp with CUDA support on a fresh install of Ubuntu and get error XY.
         Here are the exact commands that I used: ...
     validations:
       required: true

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

@@ -26,7 +26,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -43,6 +44,8 @@ body:
         description: Which GGML backends do you know to be affected?
         options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
         multiple: true
+    validations:
+      required: true
   - type: textarea
     id: hardware
     attributes:
@@ -55,20 +58,20 @@ body:
   - type: textarea
     id: model
     attributes:
-      label: Model
+      label: Models
       description: >
-        Which model at which quantization were you using when encountering the bug?
+        Which model(s) at which quantization were you using when encountering the bug?
         If you downloaded a GGUF file off of Huggingface, please provide a link.
       placeholder: >
         e.g. Meta LLaMA 3.1 Instruct 8b q4_K_M
     validations:
       required: false
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it.
         If you can narrow down the bug to specific hardware, compile flags, or command line arguments,
         that information would be very much appreciated by us.
       placeholder: >

.github/ISSUE_TEMPLATE/019-bug-misc.yml

@@ -14,7 +14,7 @@ body:
     id: version
     attributes:
       label: Name and Version
-      description: Which version of our software are you running? (use `--version` to get a version string)
+      description: Which version of our software is affected? (You can use `--version` to get a version string.)
       placeholder: |
         $./llama-cli --version
         version: 2999 (42b4109e)
@@ -24,7 +24,8 @@ body:
   - type: dropdown
     id: operating-system
     attributes:
-      label: Which operating systems do you know to be affected?
+      label: Operating systems
+      description: Which operating systems do you know to be affected?
       multiple: true
       options:
         - Linux
@@ -33,28 +34,30 @@ body:
         - BSD
         - Other? (Please let us know in description)
     validations:
-      required: true
+      required: false
   - type: dropdown
     id: module
     attributes:
       label: Which llama.cpp modules do you know to be affected?
       multiple: true
       options:
+        - Documentation/Github
         - libllama (core library)
         - llama-cli
         - llama-server
         - llama-bench
         - llama-quantize
         - Python/Bash scripts
+        - Test code
         - Other (Please specify in the next section)
     validations:
-      required: true
+      required: false
   - type: textarea
-    id: steps_to_reproduce
+    id: info
     attributes:
-      label: Steps to Reproduce
+      label: Problem description & steps to reproduce
       description: >
-        Please tell us how to reproduce the bug and any additional information that you think could be useful for fixing it.
+        Please give us a summary of the problem and tell us how to reproduce it (if applicable).
     validations:
       required: true
   - type: textarea
@@ -62,7 +65,7 @@ body:
     attributes:
       label: First Bad Commit
       description: >
-        If the bug was not present on an earlier version: when did it start appearing?
+        If the bug was not present on an earlier version and it's not trivial to track down: when did it start appearing?
         If possible, please do a git bisect and identify the exact commit that introduced the bug.
     validations:
       required: false
@@ -71,8 +74,8 @@ body:
     attributes:
       label: Relevant log output
       description: >
-          Please copy and paste any relevant log output, including the command that you entered and any generated text.
+          If applicable, please copy and paste any relevant log output, including the command that you entered and any generated text.
           This will be automatically formatted into code, so no need for backticks.
       render: shell
     validations:
-      required: true
+      required: false

.github/workflows/docker.yml

@@ -10,12 +10,10 @@
 name: Publish Docker image
 
 on:
-  #pull_request:
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
-  workflow_dispatch: # allows manual triggering, useful for debugging
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    # Rebuild daily rather than on every push because it is expensive
+    - cron: '12 4 * * *'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -29,7 +27,6 @@ permissions:
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
-    #if: github.event.pull_request.draft == false
 
     runs-on: ubuntu-latest
     env:

CMakeLists.txt

@@ -163,8 +163,11 @@ if (GGML_TARGET_DEFINES)
     list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
 endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
-
-set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
+# all public headers
+set(LLAMA_PUBLIC_HEADERS
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
+set_target_properties(llama PROPERTIES PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
 configure_package_config_file(

Makefile

@@ -34,6 +34,7 @@ BUILD_TARGETS = \
 	llama-server \
 	llama-simple \
 	llama-simple-chat \
+	llama-run \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -1167,6 +1168,11 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-run: examples/run/run.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-simple: examples/simple/simple.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

common/arg.cpp

@@ -298,6 +298,27 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
     print_options(specific_options);
 }
 
+static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & value) {
+    std::vector<ggml_backend_dev_t> devices;
+    auto dev_names = string_split<std::string>(value, ',');
+    if (dev_names.empty()) {
+        throw std::invalid_argument("no devices specified");
+    }
+    if (dev_names.size() == 1 && dev_names[0] == "none") {
+        devices.push_back(nullptr);
+    } else {
+        for (const auto & device : dev_names) {
+            auto * dev = ggml_backend_dev_by_name(device.c_str());
+            if (!dev || ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU) {
+                throw std::invalid_argument(string_format("invalid device: %s", device.c_str()));
+            }
+            devices.push_back(dev);
+        }
+        devices.push_back(nullptr);
+    }
+    return devices;
+}
+
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
     auto ctx_arg = common_params_parser_init(params, ex, print_usage);
     const common_params params_org = ctx_arg.params; // the example can modify the default params
@@ -324,6 +345,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 }
 
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
+    // load dynamic backends
+    ggml_backend_load_all();
+
     common_params_context ctx_arg(params);
     ctx_arg.print_usage = print_usage;
     ctx_arg.ex          = ex;
@@ -1312,6 +1336,30 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_NUMA"));
+    add_opt(common_arg(
+        {"-dev", "--device"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.devices = parse_device_list(value);
+        }
+    ).set_env("LLAMA_ARG_DEVICE"));
+    add_opt(common_arg(
+        {"--list-devices"},
+        "print list of available devices and exit",
+        [](common_params &) {
+            printf("Available devices:\n");
+            for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                auto * dev = ggml_backend_dev_get(i);
+                if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+                    size_t free, total;
+                    ggml_backend_dev_memory(dev, &free, &total);
+                    printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
+                }
+            }
+            exit(0);
+        }
+    ));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -1336,10 +1384,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             } else if (arg_next == "layer") {
                 params.split_mode = LLAMA_SPLIT_MODE_LAYER;
             } else if (arg_next == "row") {
-#ifdef GGML_USE_SYCL
-                fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
-                exit(1);
-#endif // GGML_USE_SYCL
                 params.split_mode = LLAMA_SPLIT_MODE_ROW;
             } else {
                 throw std::invalid_argument("invalid value");
@@ -2042,6 +2086,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.n_ctx = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"-devd", "--device-draft"}, "<dev1,dev2,..>",
+        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            params.speculative.devices = parse_device_list(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-ngld", "--gpu-layers-draft", "--n-gpu-layers-draft"}, "N",
         "number of layers to store in VRAM for the draft model",

common/common.cpp

@@ -377,9 +377,6 @@ void common_init() {
 #endif
 
     LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
-
-    // load dynamic backends
-    ggml_backend_load_all();
 }
 
 std::string common_params_get_system_info(const common_params & params) {
@@ -982,9 +979,12 @@ void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_l
     }
 }
 
-struct llama_model_params common_model_params_to_llama(const common_params & params) {
+struct llama_model_params common_model_params_to_llama(common_params & params) {
     auto mparams = llama_model_default_params();
 
+    if (!params.devices.empty()) {
+        mparams.devices = params.devices.data();
+    }
     if (params.n_gpu_layers != -1) {
         mparams.n_gpu_layers = params.n_gpu_layers;
     }

common/common.h

@@ -156,6 +156,7 @@ struct common_params_sampling {
 };
 
 struct common_params_speculative {
+    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
     int32_t n_ctx        =     0; // draft context size
     int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
     int32_t n_min        =     5; // minimum number of draft tokens to use for speculative decoding
@@ -178,9 +179,6 @@ struct common_params {
     int32_t n_chunks              =    -1; // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            =     1; // number of parallel sequences to decode
     int32_t n_sequences           =     1; // number of sequences to decode
-    int32_t n_gpu_layers          =    -1; // number of layers to store in VRAM (-1 - use default)
-    int32_t main_gpu              =     0; // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]     =   {0}; // how split tensors should be distributed across GPUs
     int32_t grp_attn_n            =     1; // group-attention factor
     int32_t grp_attn_w            =   512; // group-attention width
     int32_t n_print               =    -1; // print token count every n tokens (-1 = disabled)
@@ -193,6 +191,13 @@ struct common_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          =  0.1f; // KV cache defragmentation threshold
 
+    // offload params
+    std::vector<ggml_backend_dev_t> devices;         // devices to use for offloading
+    int32_t n_gpu_layers                    =    -1; // number of layers to store in VRAM (-1 - use default)
+    int32_t main_gpu                        =     0; // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]               =   {0}; // how split tensors should be distributed across GPUs
+    enum llama_split_mode        split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -201,7 +206,6 @@ struct common_params {
 
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
-    enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
@@ -462,7 +466,7 @@ struct common_init_result {
 
 struct common_init_result     common_init_from_params(common_params & params);
 
-struct llama_model_params     common_model_params_to_llama  (const common_params & params);
+struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
 struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 

common/speculative.cpp

@@ -90,9 +90,10 @@ bool common_speculative_are_compatible(
     if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
         llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
         llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
-        llama_token_eos(model_tgt) != llama_token_eos(model_dft)
-    ) {
+        llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
         LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
+        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
         return false;
     }
 

convert_hf_to_gguf.py

@@ -3040,9 +3040,9 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@Model.register("Olmo1124ForCausalLM")
-class Olmo1124Model(Model):
-    model_arch = gguf.MODEL_ARCH.OLMO_1124
+@Model.register("Olmo2ForCausalLM")
+class Olmo2Model(Model):
+    model_arch = gguf.MODEL_ARCH.OLMO2
 
 
 @Model.register("OlmoeForCausalLM")

examples/CMakeLists.txt

@@ -36,6 +36,7 @@ else()
         add_subdirectory(server)
     endif()
     add_subdirectory(save-load-state)
+    add_subdirectory(run)
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)

examples/run/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-run)
+add_executable(${TARGET} run.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/run/README.md

@@ -0,0 +1,7 @@
+# llama.cpp/example/run
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
+
+```bash
+./llama-run Meta-Llama-3.1-8B-Instruct.gguf
+...

examples/run/run.cpp

@@ -0,0 +1,409 @@
+#if defined(_WIN32)
+#include <windows.h>
+#else
+#include <unistd.h>
+#endif
+
+#include <climits>
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "llama-cpp.h"
+
+typedef std::unique_ptr<char[]> char_array_ptr;
+
+struct Argument {
+    std::string flag;
+    std::string help_text;
+};
+
+struct Options {
+    std::string model_path, prompt_non_interactive;
+    int ngl = 99;
+    int n_ctx = 2048;
+};
+
+class ArgumentParser {
+   public:
+    ArgumentParser(const char * program_name) : program_name(program_name) {}
+
+    void add_argument(const std::string & flag, std::string & var, const std::string & help_text = "") {
+        string_args[flag] = &var;
+        arguments.push_back({flag, help_text});
+    }
+
+    void add_argument(const std::string & flag, int & var, const std::string & help_text = "") {
+        int_args[flag] = &var;
+        arguments.push_back({flag, help_text});
+    }
+
+    int parse(int argc, const char ** argv) {
+        for (int i = 1; i < argc; ++i) {
+            std::string arg = argv[i];
+            if (string_args.count(arg)) {
+                if (i + 1 < argc) {
+                    *string_args[arg] = argv[++i];
+                } else {
+                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
+                    print_usage();
+                    return 1;
+                }
+            } else if (int_args.count(arg)) {
+                if (i + 1 < argc) {
+                    if (parse_int_arg(argv[++i], *int_args[arg]) != 0) {
+                        fprintf(stderr, "error: invalid value for %s: %s\n", arg.c_str(), argv[i]);
+                        print_usage();
+                        return 1;
+                    }
+                } else {
+                    fprintf(stderr, "error: missing value for %s\n", arg.c_str());
+                    print_usage();
+                    return 1;
+                }
+            } else {
+                fprintf(stderr, "error: unrecognized argument %s\n", arg.c_str());
+                print_usage();
+                return 1;
+            }
+        }
+
+        if (string_args["-m"]->empty()) {
+            fprintf(stderr, "error: -m is required\n");
+            print_usage();
+            return 1;
+        }
+
+        return 0;
+    }
+
+   private:
+    const char * program_name;
+    std::unordered_map<std::string, std::string *> string_args;
+    std::unordered_map<std::string, int *> int_args;
+    std::vector<Argument> arguments;
+
+    int parse_int_arg(const char * arg, int & value) {
+        char * end;
+        const long val = std::strtol(arg, &end, 10);
+        if (*end == '\0' && val >= INT_MIN && val <= INT_MAX) {
+            value = static_cast<int>(val);
+            return 0;
+        }
+        return 1;
+    }
+
+    void print_usage() const {
+        printf("\nUsage:\n");
+        printf("  %s [OPTIONS]\n\n", program_name);
+        printf("Options:\n");
+        for (const auto & arg : arguments) {
+            printf("  %-10s %s\n", arg.flag.c_str(), arg.help_text.c_str());
+        }
+
+        printf("\n");
+    }
+};
+
+class LlamaData {
+   public:
+    llama_model_ptr model;
+    llama_sampler_ptr sampler;
+    llama_context_ptr context;
+    std::vector<llama_chat_message> messages;
+
+    int init(const Options & opt) {
+        model = initialize_model(opt.model_path, opt.ngl);
+        if (!model) {
+            return 1;
+        }
+
+        context = initialize_context(model, opt.n_ctx);
+        if (!context) {
+            return 1;
+        }
+
+        sampler = initialize_sampler();
+        return 0;
+    }
+
+   private:
+    // Initializes the model and returns a unique pointer to it
+    llama_model_ptr initialize_model(const std::string & model_path, const int ngl) {
+        llama_model_params model_params = llama_model_default_params();
+        model_params.n_gpu_layers = ngl;
+
+        llama_model_ptr model(llama_load_model_from_file(model_path.c_str(), model_params));
+        if (!model) {
+            fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        }
+
+        return model;
+    }
+
+    // Initializes the context with the specified parameters
+    llama_context_ptr initialize_context(const llama_model_ptr & model, const int n_ctx) {
+        llama_context_params ctx_params = llama_context_default_params();
+        ctx_params.n_ctx = n_ctx;
+        ctx_params.n_batch = n_ctx;
+
+        llama_context_ptr context(llama_new_context_with_model(model.get(), ctx_params));
+        if (!context) {
+            fprintf(stderr, "%s: error: failed to create the llama_context\n", __func__);
+        }
+
+        return context;
+    }
+
+    // Initializes and configures the sampler
+    llama_sampler_ptr initialize_sampler() {
+        llama_sampler_ptr sampler(llama_sampler_chain_init(llama_sampler_chain_default_params()));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_min_p(0.05f, 1));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_temp(0.8f));
+        llama_sampler_chain_add(sampler.get(), llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
+
+        return sampler;
+    }
+};
+
+// Add a message to `messages` and store its content in `owned_content`
+static void add_message(const char * role, const std::string & text, LlamaData & llama_data,
+                        std::vector<char_array_ptr> & owned_content) {
+    char_array_ptr content(new char[text.size() + 1]);
+    std::strcpy(content.get(), text.c_str());
+    llama_data.messages.push_back({role, content.get()});
+    owned_content.push_back(std::move(content));
+}
+
+// Function to apply the chat template and resize `formatted` if needed
+static int apply_chat_template(const LlamaData & llama_data, std::vector<char> & formatted, const bool append) {
+    int result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
+                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+    if (result > static_cast<int>(formatted.size())) {
+        formatted.resize(result);
+        result = llama_chat_apply_template(llama_data.model.get(), nullptr, llama_data.messages.data(),
+                                           llama_data.messages.size(), append, formatted.data(), formatted.size());
+    }
+
+    return result;
+}
+
+// Function to tokenize the prompt
+static int tokenize_prompt(const llama_model_ptr & model, const std::string & prompt,
+                           std::vector<llama_token> & prompt_tokens) {
+    const int n_prompt_tokens = -llama_tokenize(model.get(), prompt.c_str(), prompt.size(), NULL, 0, true, true);
+    prompt_tokens.resize(n_prompt_tokens);
+    if (llama_tokenize(model.get(), prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
+                       true) < 0) {
+        GGML_ABORT("failed to tokenize the prompt\n");
+    }
+
+    return n_prompt_tokens;
+}
+
+// Check if we have enough space in the context to evaluate this batch
+static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
+    const int n_ctx = llama_n_ctx(ctx.get());
+    const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
+    if (n_ctx_used + batch.n_tokens > n_ctx) {
+        printf("\033[0m\n");
+        fprintf(stderr, "context size exceeded\n");
+        return 1;
+    }
+
+    return 0;
+}
+
+// convert the token to a string
+static int convert_token_to_string(const llama_model_ptr & model, const llama_token token_id, std::string & piece) {
+    char buf[256];
+    int n = llama_token_to_piece(model.get(), token_id, buf, sizeof(buf), 0, true);
+    if (n < 0) {
+        GGML_ABORT("failed to convert token to piece\n");
+    }
+
+    piece = std::string(buf, n);
+    return 0;
+}
+
+static void print_word_and_concatenate_to_response(const std::string & piece, std::string & response) {
+    printf("%s", piece.c_str());
+    fflush(stdout);
+    response += piece;
+}
+
+// helper function to evaluate a prompt and generate a response
+static int generate(LlamaData & llama_data, const std::string & prompt, std::string & response) {
+    std::vector<llama_token> prompt_tokens;
+    const int n_prompt_tokens = tokenize_prompt(llama_data.model, prompt, prompt_tokens);
+    if (n_prompt_tokens < 0) {
+        return 1;
+    }
+
+    // prepare a batch for the prompt
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+    llama_token new_token_id;
+    while (true) {
+        check_context_size(llama_data.context, batch);
+        if (llama_decode(llama_data.context.get(), batch)) {
+            GGML_ABORT("failed to decode\n");
+        }
+
+        // sample the next token, check is it an end of generation?
+        new_token_id = llama_sampler_sample(llama_data.sampler.get(), llama_data.context.get(), -1);
+        if (llama_token_is_eog(llama_data.model.get(), new_token_id)) {
+            break;
+        }
+
+        std::string piece;
+        if (convert_token_to_string(llama_data.model, new_token_id, piece)) {
+            return 1;
+        }
+
+        print_word_and_concatenate_to_response(piece, response);
+
+        // prepare the next batch with the sampled token
+        batch = llama_batch_get_one(&new_token_id, 1);
+    }
+
+    return 0;
+}
+
+static int parse_arguments(const int argc, const char ** argv, Options & opt) {
+    ArgumentParser parser(argv[0]);
+    parser.add_argument("-m", opt.model_path, "model");
+    parser.add_argument("-p", opt.prompt_non_interactive, "prompt");
+    parser.add_argument("-c", opt.n_ctx, "context_size");
+    parser.add_argument("-ngl", opt.ngl, "n_gpu_layers");
+    if (parser.parse(argc, argv)) {
+        return 1;
+    }
+
+    return 0;
+}
+
+static int read_user_input(std::string & user) {
+    std::getline(std::cin, user);
+    return user.empty();  // Indicate an error or empty input
+}
+
+// Function to generate a response based on the prompt
+static int generate_response(LlamaData & llama_data, const std::string & prompt, std::string & response) {
+    // Set response color
+    printf("\033[33m");
+    if (generate(llama_data, prompt, response)) {
+        fprintf(stderr, "failed to generate response\n");
+        return 1;
+    }
+
+    // End response with color reset and newline
+    printf("\n\033[0m");
+    return 0;
+}
+
+// Helper function to apply the chat template and handle errors
+static int apply_chat_template_with_error_handling(const LlamaData & llama_data, std::vector<char> & formatted,
+                                                   const bool is_user_input, int & output_length) {
+    const int new_len = apply_chat_template(llama_data, formatted, is_user_input);
+    if (new_len < 0) {
+        fprintf(stderr, "failed to apply the chat template\n");
+        return -1;
+    }
+
+    output_length = new_len;
+    return 0;
+}
+
+// Helper function to handle user input
+static bool handle_user_input(std::string & user_input, const std::string & prompt_non_interactive) {
+    if (!prompt_non_interactive.empty()) {
+        user_input = prompt_non_interactive;
+        return true;  // No need for interactive input
+    }
+
+    printf("\033[32m> \033[0m");
+    return !read_user_input(user_input);  // Returns false if input ends the loop
+}
+
+// Function to tokenize the prompt
+static int chat_loop(LlamaData & llama_data, std::string & prompt_non_interactive) {
+    std::vector<char_array_ptr> owned_content;
+    std::vector<char> fmtted(llama_n_ctx(llama_data.context.get()));
+    int prev_len = 0;
+
+    while (true) {
+        // Get user input
+        std::string user_input;
+        if (!handle_user_input(user_input, prompt_non_interactive)) {
+            break;
+        }
+
+        add_message("user", prompt_non_interactive.empty() ? user_input : prompt_non_interactive, llama_data,
+                    owned_content);
+
+        int new_len;
+        if (apply_chat_template_with_error_handling(llama_data, fmtted, true, new_len) < 0) {
+            return 1;
+        }
+
+        std::string prompt(fmtted.begin() + prev_len, fmtted.begin() + new_len);
+        std::string response;
+        if (generate_response(llama_data, prompt, response)) {
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static void log_callback(const enum ggml_log_level level, const char * text, void *) {
+    if (level == GGML_LOG_LEVEL_ERROR) {
+        fprintf(stderr, "%s", text);
+    }
+}
+
+static bool is_stdin_a_terminal() {
+#if defined(_WIN32)
+    HANDLE hStdin = GetStdHandle(STD_INPUT_HANDLE);
+    DWORD mode;
+    return GetConsoleMode(hStdin, &mode);
+#else
+    return isatty(STDIN_FILENO);
+#endif
+}
+
+static std::string read_pipe_data() {
+    std::ostringstream result;
+    result << std::cin.rdbuf();  // Read all data from std::cin
+    return result.str();
+}
+
+int main(int argc, const char ** argv) {
+    Options opt;
+    if (parse_arguments(argc, argv, opt)) {
+        return 1;
+    }
+
+    if (!is_stdin_a_terminal()) {
+        if (!opt.prompt_non_interactive.empty()) {
+            opt.prompt_non_interactive += "\n\n";
+        }
+
+        opt.prompt_non_interactive += read_pipe_data();
+    }
+
+    llama_log_set(log_callback, nullptr);
+    LlamaData llama_data;
+    if (llama_data.init(opt)) {
+        return 1;
+    }
+
+    if (chat_loop(llama_data, opt.prompt_non_interactive)) {
+        return 1;
+    }
+
+    return 0;
+}

examples/server/README.md

@@ -412,7 +412,7 @@ node index.js
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
-    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
+    `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `true`
 
     `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["dry", "top_k", "typ_p", "top_p", "min_p", "xtc", "temperature"]` - these are all the available values.
 

examples/server/public/index.html

@@ -81,7 +81,13 @@
               <path d="M14.5 3a1 1 0 0 1-1 1H13v9a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V4h-.5a1 1 0 0 1-1-1V2a1 1 0 0 1 1-1H6a1 1 0 0 1 1-1h2a1 1 0 0 1 1 1h3.5a1 1 0 0 1 1 1zM4.118 4 4 4.059V13a1 1 0 0 0 1 1h6a1 1 0 0 0 1-1V4.059L11.882 4zM2.5 3h11V2h-11z"/>
             </svg>
           </button>
-
+          <button v-if="messages.length > 0" class="btn mr-1" @click="downloadConv(viewingConvId)" :disabled="isGenerating">
+              <!-- download conversation button -->
+              <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-download" viewBox="0 0 16 16">
+                  <path d="M.5 9.9a.5.5 0 0 1 .5.5v2.5a1 1 0 0 0 1 1h12a1 1 0 0 0 1-1v-2.5a.5.5 0 0 1 1 0v2.5a2 2 0 0 1-2 2H2a2 2 0 0 1-2-2v-2.5a.5.5 0 0 1 .5-.5"/>
+                  <path d="M7.646 11.854a.5.5 0 0 0 .708 0l3-3a.5.5 0 0 0-.708-.708L8.5 10.293V1.5a.5.5 0 0 0-1 0v8.793L5.354 8.146a.5.5 0 1 0-.708.708z"/>
+            </svg>
+          </button>
           <button class="btn" @click="showConfigDialog = true" :disabled="isGenerating">
             <!-- edit config button -->
             <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-gear" viewBox="0 0 16 16">
@@ -526,6 +532,23 @@
             this.fetchMessages();
           }
         },
+        downloadConv(convId) {
+          const conversation = StorageUtils.getOneConversation(convId);
+          if (!conversation) {
+            alert('Conversation not found.');
+            return;
+          }
+          const conversationJson = JSON.stringify(conversation, null, 2);
+          const blob = new Blob([conversationJson], { type: 'application/json' });
+          const url = URL.createObjectURL(blob);
+          const a = document.createElement('a');
+          a.href = url;
+          a.download = `conversation_${convId}.json`;
+          document.body.appendChild(a);
+          a.click();
+          document.body.removeChild(a);
+          URL.revokeObjectURL(url);
+        },
         async sendMessage() {
           if (!this.inputMsg) return;
           const currConvId = this.viewingConvId;

examples/server/server.cpp

@@ -111,7 +111,7 @@ struct server_static_file {
 
 struct slot_params {
     bool stream       = true;
-    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+    bool cache_prompt = true; // remember the prompt to avoid reprocessing all prompt
 
     int32_t n_keep    =  0; // number of tokens to keep from initial prompt
     int32_t n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
@@ -692,6 +692,7 @@ struct server_context {
 
             auto params_dft = params_base;
 
+            params_dft.devices      = params_base.speculative.devices;
             params_dft.model        = params_base.speculative.model;
             params_dft.n_ctx        = params_base.speculative.n_ctx;
             params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
@@ -882,7 +883,7 @@ struct server_context {
         }
 
         slot.params.stream           = json_value(data, "stream",             false);
-        slot.params.cache_prompt     = json_value(data, "cache_prompt",       false);
+        slot.params.cache_prompt     = json_value(data, "cache_prompt",       true);
         slot.params.n_predict        = json_value(data, "n_predict",          json_value(data, "max_tokens", defaults.n_predict));
         slot.params.n_indent         = json_value(data, "n_indent",           defaults.n_indent);
         slot.params.n_keep           = json_value(data, "n_keep",             defaults.n_keep);

examples/speculative-simple/speculative-simple.cpp

@@ -46,6 +46,7 @@ int main(int argc, char ** argv) {
     ctx_tgt   = llama_init_tgt.context;
 
     // load the draft model
+    params.devices      = params.speculative.devices;
     params.model        = params.speculative.model;
     params.n_ctx        = params.speculative.n_ctx;
     params.n_batch      = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;

examples/speculative/speculative.cpp

@@ -76,6 +76,7 @@ int main(int argc, char ** argv) {
     ctx_tgt = llama_init_tgt.context;
 
     // load the draft model
+    params.devices = params.speculative.devices;
     params.model = params.speculative.model;
     params.n_gpu_layers = params.speculative.n_gpu_layers;
     if (params.speculative.cpuparams.n_threads > 0) {

ggml/src/ggml-backend-reg.cpp

@@ -253,6 +253,15 @@ void ggml_backend_device_register(ggml_backend_dev_t device) {
 }
 
 // Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
 size_t ggml_backend_reg_count() {
     return get_reg().backends.size();
 }
@@ -265,7 +274,7 @@ ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
 ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
         ggml_backend_reg_t reg = ggml_backend_reg_get(i);
-        if (std::strcmp(ggml_backend_reg_name(reg), name) == 0) {
+        if (striequals(ggml_backend_reg_name(reg), name)) {
             return reg;
         }
     }
@@ -285,7 +294,7 @@ ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
 ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+        if (striequals(ggml_backend_dev_name(dev), name)) {
             return dev;
         }
     }

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

@@ -1927,7 +1927,7 @@ static void ggml_metal_encode_node(
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
-                int ne11_mm_min = 1;
+                int ne11_mm_min = 4;
 
 #if 0
                 // the numbers below are measured on M2 Ultra for 7B and 13B models

gguf-py/gguf/constants.py

@@ -243,7 +243,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R    = auto()
     DBRX         = auto()
     OLMO         = auto()
-    OLMO_1124    = auto()
+    OLMO2        = auto()
     OLMOE        = auto()
     OPENELM      = auto()
     ARCTIC       = auto()
@@ -405,7 +405,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
-    MODEL_ARCH.OLMO_1124:      "olmo_1124",
+    MODEL_ARCH.OLMO2:          "olmo2",
     MODEL_ARCH.OLMOE:          "olmoe",
     MODEL_ARCH.OPENELM:        "openelm",
     MODEL_ARCH.ARCTIC:         "arctic",
@@ -1071,7 +1071,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
-    MODEL_ARCH.OLMO_1124: [
+    MODEL_ARCH.OLMO2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.OUTPUT,

gguf-py/gguf/tensor_mapping.py

@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo_1124
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -54,7 +54,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo_1124
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -66,7 +66,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo_1124
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo2
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -145,7 +145,7 @@ class TensorNameMap:
 
         # Attention query
         MODEL_TENSOR.ATTN_Q: (
-            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.q_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wq",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.query",                  # bert
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
@@ -157,7 +157,7 @@ class TensorNameMap:
 
         # Attention key
         MODEL_TENSOR.ATTN_K: (
-            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.k_proj",                     # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
@@ -170,7 +170,7 @@ class TensorNameMap:
 
         # Attention value
         MODEL_TENSOR.ATTN_V: (
-            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
@@ -188,7 +188,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.out_proj",                       # mpt
             "transformer.h.{bid}.self_attention.dense",                     # falcon
             "h.{bid}.self_attention.dense",                                 # bloom
-            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo_1124
+            "model.layers.{bid}.self_attn.o_proj",                          # llama-hf nemotron olmoe olmo2
             "layers.{bid}.attention.wo",                                    # llama-pth
             "encoder.layer.{bid}.attention.output.dense",                   # bert
             "transformer.h.{bid}.attn.out_proj",                            # gpt-j
@@ -215,7 +215,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo_1124
+            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2
         ),
 
         # Rotary embeddings
@@ -250,7 +250,7 @@ class TensorNameMap:
 
         # Post feed-forward norm
         MODEL_TENSOR.FFN_POST_NORM: (
-            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo_1124
+            "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -273,7 +273,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.up_proj",                   # mpt
             "transformer.h.{bid}.mlp.dense_h_to_4h",                  # falcon
             "h.{bid}.mlp.dense_h_to_4h",                              # bloom
-            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo_1124
+            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo2
             "layers.{bid}.feed_forward.w3",                           # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
@@ -314,7 +314,7 @@ class TensorNameMap:
 
         # Feed-forward gate
         MODEL_TENSOR.FFN_GATE: (
-            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo_1124
+            "model.layers.{bid}.mlp.gate_proj",           # llama-hf refact olmo2
             "layers.{bid}.feed_forward.w1",               # llama-pth
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "transformer.h.{bid}.mlp.c_fc2",              # jais
@@ -346,7 +346,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.down_proj",                 # mpt
             "transformer.h.{bid}.mlp.dense_4h_to_h",                  # falcon
             "h.{bid}.mlp.dense_4h_to_h",                              # bloom
-            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo_1124
+            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo2
             "layers.{bid}.feed_forward.w2",                           # llama-pth
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
@@ -383,7 +383,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.q_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
@@ -392,7 +392,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
-            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo_1124
+            "model.layers.{bid}.self_attn.k_norm",                            # cohere olmoe chameleon olmo2
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm

include/llama-cpp.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include <memory>
+
+#include "llama.h"
+
+struct llama_model_deleter {
+    void operator()(llama_model * model) { llama_free_model(model); }
+};
+
+struct llama_context_deleter {
+    void operator()(llama_context * context) { llama_free(context); }
+};
+
+struct llama_sampler_deleter {
+    void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
+};
+
+typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
+typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
+typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;

include/llama.h

@@ -272,6 +272,9 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
+        ggml_backend_dev_t * devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 

src/llama.cpp

@@ -179,7 +179,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
-    LLM_ARCH_OLMO_1124,
+    LLM_ARCH_OLMO2,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
@@ -233,7 +233,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
-    { LLM_ARCH_OLMO_1124,       "olmo_1124"    },
+    { LLM_ARCH_OLMO2,           "olmo2"        },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
@@ -1210,7 +1210,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
         },
     },
     {
-        LLM_ARCH_OLMO_1124,
+        LLM_ARCH_OLMO2,
         {
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
@@ -5900,7 +5900,7 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
@@ -8593,7 +8593,7 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
-            case LLM_ARCH_OLMO_1124:
+            case LLM_ARCH_OLMO2:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -14483,7 +14483,7 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_olmo_1124() {
+    struct ggml_cgraph * build_olmo2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         // mutable variable, needed during the last layer of the computation to skip unused tokens
@@ -16799,9 +16799,9 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
             {
-                result = llm.build_olmo_1124();
+                result = llm.build_olmo2();
             } break;
         case LLM_ARCH_OLMOE:
             {
@@ -19364,6 +19364,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
 //
 struct llama_model_params llama_model_default_params() {
     struct llama_model_params result = {
+        /*.devices                     =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
@@ -19576,19 +19577,24 @@ struct llama_model * llama_load_model_from_file(
     }
 
     // create list of devices to use with this model
-    // currently, we use all available devices
-    // TODO: rework API to give user more control over device selection
-    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        switch (ggml_backend_dev_type(dev)) {
-            case GGML_BACKEND_DEVICE_TYPE_CPU:
-            case GGML_BACKEND_DEVICE_TYPE_ACCEL:
-                // skip CPU backends since they are handled separately
-                break;
+    if (params.devices) {
+        for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
+            model->devices.push_back(*dev);
+        }
+    } else {
+        // use all available devices
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            switch (ggml_backend_dev_type(dev)) {
+                case GGML_BACKEND_DEVICE_TYPE_CPU:
+                case GGML_BACKEND_DEVICE_TYPE_ACCEL:
+                    // skip CPU backends since they are handled separately
+                    break;
 
-            case GGML_BACKEND_DEVICE_TYPE_GPU:
-                model->devices.push_back(dev);
-                break;
+                case GGML_BACKEND_DEVICE_TYPE_GPU:
+                    model->devices.push_back(dev);
+                    break;
+            }
         }
     }
 
@@ -20078,7 +20084,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
-        case LLM_ARCH_OLMO_1124:
+        case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3: