0001-Expose-callable-API-for-server.patch

From 087cf3300e973d7790db8f7cad01d2a790de38be Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <daniel@ollama.com>
Date: Mon, 13 Nov 2023 12:25:58 -0800
Subject: [PATCH] Expose callable API for server

This adds an extern "C" interface within the example server
---
 examples/server/CMakeLists.txt |  24 +++
 examples/server/server.cpp     | 274 +++++++++++++++++++++++++++++++++
 examples/server/server.h       |  89 +++++++++++
 ggml-cuda.cu                   |   1 +
 4 files changed, 388 insertions(+)
 create mode 100644 examples/server/server.h

diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt
index 859cd12..4ea47a7 100644
--- a/examples/server/CMakeLists.txt
+++ b/examples/server/CMakeLists.txt
@@ -11,3 +11,27 @@ if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET ext_server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+add_library(${TARGET} STATIC server.cpp)
+target_include_directories(${TARGET} PRIVATE ../../common)
+target_include_directories(${TARGET} PRIVATE ../..)
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_definitions(${TARGET} PUBLIC LLAMA_SERVER_LIBRARY=1)
+target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(ext_server PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(ext_server PRIVATE LLAMA_SHARED LLAMA_BUILD)
+    add_library(ext_server_shared SHARED $<TARGET_OBJECTS:ext_server>)
+    target_link_libraries(ext_server_shared PRIVATE ggml llama llava common ${CMAKE_THREAD_LIBS_INIT})
+    install(TARGETS ext_server_shared LIBRARY)
+endif()
+
+if (CUDAToolkit_FOUND)
+    target_include_directories(${TARGET} PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+    if (WIN32)
+        target_link_libraries(ext_server_shared PRIVATE nvml)
+    endif()
+endif()
\ No newline at end of file
diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index d0cd8e1..5f5d4c5 100644
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -5,6 +5,9 @@
 #include "../llava/clip.h"
 
 #include "stb_image.h"
+#if defined(LLAMA_SERVER_LIBRARY)
+#include "server.h"
+#endif
 
 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@@ -2632,6 +2635,7 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
     }
 }
 
+#ifndef LLAMA_SERVER_LIBRARY
 int main(int argc, char **argv)
 {
     // own arguments required by this example
@@ -3066,3 +3070,273 @@ int main(int argc, char **argv)
     llama_backend_free();
     return 0;
 }
+
+#else // LLAMA_SERVER_LIBRARY
+// Expose the llama server as a callable extern "C" API
+llama_server_context *llama = NULL;
+std::atomic<bool> ext_server_running(false);
+std::thread ext_server_thread;
+
+void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err)
+{
+    assert(err != NULL && sparams != NULL);
+    err->id = 0;
+    err->msg[0] = '\0';
+    try {
+        llama = new llama_server_context;
+        log_set_target(stdout);
+        gpt_params params;
+        params.n_ctx = sparams->n_ctx;
+        params.n_batch = sparams->n_batch;
+        params.n_threads = sparams->n_threads;
+        params.n_parallel = sparams->n_parallel;
+        params.rope_freq_base = sparams->rope_freq_base;
+        params.rope_freq_scale = sparams->rope_freq_scale;
+
+        if (sparams->memory_f16)  {
+            params.cache_type_k = "f16";
+            params.cache_type_v = "f16";
+        } else {
+            params.cache_type_k = "f32";
+            params.cache_type_v = "f32";
+        }
+
+        params.n_gpu_layers = sparams->n_gpu_layers;
+        params.main_gpu = sparams->main_gpu;
+        params.use_mlock = sparams->use_mlock;
+        params.use_mmap = sparams->use_mmap;
+        params.numa = sparams->numa;
+        params.embedding = sparams->embedding;
+        if (sparams->model != NULL) {
+            params.model = sparams->model;
+        }
+
+        for (ext_server_lora_adapter *la = sparams->lora_adapters; la != NULL; la = la->next) {
+            params.lora_adapter.push_back(std::make_tuple(la->adapter, la->scale));
+        }
+
+        if (sparams->mmproj != NULL) {
+            params.mmproj = std::string(sparams->mmproj);
+        }
+           
+        llama_backend_init(params.numa);
+
+        // load the model
+        if (!llama->load_model(params))
+        {
+            // TODO - consider modifying the logging logic or patching load_model so we can capture more detailed error messages
+            // and pass them back to the caller for better UX
+            err->id = -1;
+            snprintf(err->msg, err->msg_len, "error loading model %s", params.model.c_str());
+            return;
+        }
+
+        llama->initialize();
+    } catch (std::exception &e) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "exception %s", e.what());
+    } catch (...) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "Unknown exception initializing llama server");
+    }
+}
+
+void llama_server_start()
+{
+    assert(llama != NULL);
+     // TODO mutex to protect thread creation
+    ext_server_thread = std::thread([&]()
+    {
+        ext_server_running = true;
+        try {
+            LOG_TEE("llama server main loop starting\n");
+            ggml_time_init();
+            while (ext_server_running.load())
+            {
+                if (!llama->update_slots()) {
+                    LOG_TEE("unexpected error in llama server update_slots - exiting main loop\n");
+                    break;
+                }
+            }
+        } catch (std::exception &e) {
+            LOG_TEE("caught exception in llama server main loop: %s\n", e.what());
+        } catch (...) {
+            LOG_TEE("caught unknown exception in llama server main loop\n");
+        }
+        LOG_TEE("\nllama server shutting down\n");
+        llama_backend_free();
+    });
+}
+
+void llama_server_stop() {
+    assert(llama != NULL);
+    // TODO - too verbose, remove once things are solid
+    LOG_TEE("requesting llama server shutdown\n");
+    ext_server_running = false;
+    ext_server_thread.join();
+    delete llama;
+    llama = NULL;
+    LOG_TEE("llama server shutdown complete\n");
+}
+
+void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
+    assert(llama != NULL && json_req != NULL && resp != NULL);
+    resp->id = -1;
+    resp->msg[0] = '\0';
+    try {
+        json data = json::parse(json_req);
+        resp->id = llama->request_completion(data, false, false, -1);
+    } catch (std::exception &e) {
+        snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
+    } catch (...) {
+        snprintf(resp->msg, resp->msg_len, "Unknown exception during completion");
+    }
+}
+
+void llama_server_completion_next_result(const int task_id, ext_server_task_result_t *resp) {
+    assert(llama != NULL && resp != NULL);
+    std::string msg;
+    resp->id = -1;
+    resp->stop = false;
+    resp->error = false;
+    resp->json_resp = NULL;
+    std::string result_json;
+    try {
+        task_result result = llama->next_result(task_id);
+        result_json = result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
+        resp->id = result.id;
+        resp->stop = result.stop;
+        resp->error = result.error;
+        if (result.error) {
+            llama->request_cancel(task_id);
+        } else if (result.stop) {
+            llama->request_cancel(task_id);
+        }
+    } catch (std::exception &e) {
+        resp->error = true;
+        resp->id = -1;
+        result_json = "{\"error\":\"exception " + std::string(e.what()) + "\"}";
+    } catch (...) {
+        resp->error = true;
+        resp->id = -1;
+        result_json = "{\"error\":\"Unknown exception during completion\"}";
+    }
+    const std::string::size_type size = result_json.size() + 1;
+    resp->json_resp = new char[size];
+    snprintf(resp->json_resp, size, "%s", result_json.c_str());
+}
+
+void llama_server_release_task_result(ext_server_task_result_t *result) {
+    if (result == NULL || result->json_resp == NULL) {
+        return;
+    }
+    delete[] result->json_resp;
+}
+
+void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err) {
+    assert(llama != NULL && err != NULL);
+    err->id = 0;
+    err->msg[0] = '\0';
+    try {
+        llama->request_cancel(task_id);
+    } catch (std::exception &e) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "exception %s", e.what());
+    } catch (...) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "Unknown exception completion cancel in llama server");
+    }
+}
+
+void llama_server_tokenize(const char *json_req, char **json_resp, ext_server_resp_t *err) {
+    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
+    *json_resp = NULL;
+    err->id = 0;
+    err->msg[0] = '\0';
+    try {
+        const json body = json::parse(json_req);
+        std::vector<llama_token> tokens;
+        if (body.count("content") != 0)
+        {
+            tokens = llama->tokenize(body["content"], false);
+        }
+        const json data = format_tokenizer_response(tokens);
+        std::string result_json = data.dump();
+        const std::string::size_type size = result_json.size() + 1;
+        *json_resp = new char[size];
+        snprintf(*json_resp, size, "%s", result_json.c_str());
+    } catch (std::exception &e) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "exception %s", e.what());
+    } catch (...) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "Unknown exception during tokenize");
+    }
+}
+
+void llama_server_release_json_resp(char **json_resp) {
+    if (json_resp == NULL || *json_resp == NULL) {
+        return;
+    }
+    delete[] *json_resp;
+}
+
+void llama_server_detokenize(const char *json_req, char **json_resp, ext_server_resp_t *err) {
+    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
+    *json_resp = NULL;
+    err->id = 0;
+    err->msg[0] = '\0';
+    try {
+        const json body = json::parse(json_req);
+        std::string content;
+        if (body.count("tokens") != 0)
+        {
+            const std::vector<llama_token> tokens = body["tokens"];
+            content = tokens_to_str(llama->ctx, tokens.cbegin(), tokens.cend());
+        }
+        const json data = format_detokenized_response(content);
+        std::string result_json = data.dump();
+        const std::string::size_type size = result_json.size() + 1;
+        *json_resp = new char[size];
+        snprintf(*json_resp, size, "%s", result_json.c_str());
+    } catch (std::exception &e) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "exception %s", e.what());
+    } catch (...) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "Unknown exception during detokenize");
+    }
+}
+
+void llama_server_embedding(const char *json_req, char** json_resp, ext_server_resp_t *err) {
+    assert(llama != NULL && json_req != NULL && json_resp != NULL && err != NULL);
+    *json_resp = NULL;
+    err->id = 0;
+    err->msg[0] = '\0';
+    try {
+        const json body = json::parse(json_req);
+        json prompt;
+        if (body.count("content") != 0)
+        {
+            prompt = body["content"];
+        }
+        else
+        {
+            prompt = "";
+        }
+        const int task_id = llama->request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true, -1);
+        task_result result = llama->next_result(task_id);
+        std::string result_json = result.result_json.dump();
+        const std::string::size_type size = result_json.size() + 1;
+        *json_resp = new char[size];
+        snprintf(*json_resp, size, "%s", result_json.c_str());
+    } catch (std::exception &e) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "exception %s", e.what());
+    } catch (...) {
+        err->id = -1;
+        snprintf(err->msg, err->msg_len, "Unknown exception during embedding");
+    }
+}
+
+#endif // LLAMA_SERVER_LIBRARY
\ No newline at end of file
diff --git a/examples/server/server.h b/examples/server/server.h
new file mode 100644
index 0000000..d22f1b6
--- /dev/null
+++ b/examples/server/server.h
@@ -0,0 +1,89 @@
+#if defined(LLAMA_SERVER_LIBRARY)
+#ifndef LLAMA_SERVER_H
+#define LLAMA_SERVER_H
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdbool.h>
+
+// This exposes extern C entrypoints into the llama_server 
+// To enable the server compile with LLAMA_SERVER_LIBRARY
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+    typedef struct ext_server_resp {
+        int id; // < 0 on error
+        size_t msg_len; // caller must allocate msg and set msg_len
+        char *msg;
+    } ext_server_resp_t;
+
+    // Allocated and freed by caller
+    typedef struct ext_server_lora_adapter {
+        char *adapter;
+        float scale;
+        struct ext_server_lora_adapter *next;
+    } ext_server_lora_adapter_t;
+
+    // Allocated and freed by caller
+    typedef struct ext_server_params
+    {
+        char *model;            
+        uint32_t n_ctx;         // text context, 0 = from model
+        uint32_t n_batch;       // prompt processing maximum batch size
+        uint32_t n_threads;     // number of threads to use for generation
+        int32_t n_parallel;     // number of parallel sequences to decodewra
+        float rope_freq_base;   // RoPE base frequency, 0 = from model
+        float rope_freq_scale;  // RoPE frequency scaling factor, 0 = from model
+        bool memory_f16;        // use f16 instead of f32 for memory kv
+        int32_t n_gpu_layers;   // number of layers to store in VRAM (-1 - use default)
+        int32_t main_gpu;       // the GPU that is used for scratch and small tensors
+        bool use_mlock;         // force system to keep model in RAM
+        bool use_mmap;          // use mmap if possible
+        bool numa;              // attempt optimizations that help on some NUMA systems
+        bool embedding;         // get only sentence embedding
+        ext_server_lora_adapter_t* lora_adapters;
+        char *mmproj;
+    } ext_server_params_t;
+
+    typedef struct ext_server_task_result
+    {
+        int id;
+        bool stop;
+        bool error;
+        char* json_resp; // null terminated, memory managed by ext_server
+    } ext_server_task_result_t;
+
+    // Initialize the server once per process
+    // err->id = 0 for success and err->msg[0] = NULL
+    // err->id != 0 for failure, and err->msg contains error message
+    void llama_server_init(ext_server_params_t *sparams, ext_server_resp_t *err);
+
+    // Run the main loop, called once per init
+    void llama_server_start();
+    // Stop the main loop and free up resources allocated in init and start.  Init must be called again to reuse
+    void llama_server_stop();
+
+    // json_req null terminated string, memory managed by caller
+    // resp->id >= 0 on success (task ID)
+    // resp->id < 0 on error, and resp->msg contains error message
+    void llama_server_completion(const char *json_req, ext_server_resp_t *resp);
+
+    // Caller must call llama_server_release_task_result to free resp->json_resp
+    void llama_server_completion_next_result(const int task_id, ext_server_task_result_t *result);
+    void llama_server_completion_cancel(const int task_id, ext_server_resp_t *err);
+    void llama_server_release_task_result(ext_server_task_result_t *result);
+
+    // Caller must call llama_server_releaes_json_resp to free json_resp if err.id < 0
+    void llama_server_tokenize(const char *json_req, char **json_resp, ext_server_resp_t *err);
+    void llama_server_detokenize(const char *json_req, char **json_resp, ext_server_resp_t *err);
+    void llama_server_embedding(const char *json_req, char** json_resp, ext_server_resp_t *err);
+    void llama_server_release_json_resp(char **json_resp);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
+#endif // LLAMA_SERVER_LIBRARY
\ No newline at end of file
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 9e1acd3..ea64b55 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -6505,6 +6505,7 @@ static cudaError_t ggml_cuda_cpy_tensor_2d(
         CUDA_CHECK(cudaGetDevice(&id));
         src_ptr = (char *) extra->data_device[id];
     } else {
+        fprintf(stderr, "ggml_cuda_cpy_tensor_2d assert: backend: %d\n", src->backend);
         GGML_ASSERT(false);
     }
     char * dst_ptr = (char *) dst;
-- 
2.39.3 (Apple Git-145)