update llama.cpp submodule to `66c1968f7` (#2618)

llm/dyn_ext_server.go

@@ -106,7 +106,12 @@ func newDynExtServer(library, model string, adapters, projectors []string, opts
 	sparams.memory_f16 = C.bool(opts.F16KV)
 	sparams.use_mlock = C.bool(opts.UseMLock)
 	sparams.use_mmap = C.bool(opts.UseMMap)
-	sparams.numa = C.bool(opts.UseNUMA)
+
+	if opts.UseNUMA {
+		sparams.numa = C.int(1)
+	} else {
+		sparams.numa = C.int(0)
+	}
 
 	sparams.lora_adapters = nil
 	for i := 0; i < len(adapters); i++ {

llm/ext_server/ext_server.cpp

@@ -80,7 +80,7 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
     params.main_gpu = sparams->main_gpu;
     params.use_mlock = sparams->use_mlock;
     params.use_mmap = sparams->use_mmap;
-    params.numa = sparams->numa;
+    params.numa = (ggml_numa_strategy)sparams->numa;
     params.embedding = sparams->embedding;
     if (sparams->model != NULL) {
       params.model = sparams->model;
@@ -111,7 +111,8 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
     }
 #endif
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // load the model
     if (!llama->load_model(params)) {

llm/ext_server/ext_server.h

@@ -41,7 +41,7 @@ typedef struct ext_server_params {
   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
+  int numa;              // attempt optimizations that help on some NUMA systems
   bool embedding;        // get only sentence embedding
   ext_server_lora_adapter_t *lora_adapters;
   char *mmproj;

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit 6c00a066928b0475b865a2e3e709e2166e02d548
+Subproject commit 66c1968f7a2e895675425e875b6589f1233a1b52

llm/patches/03-cudaleaks.diff → llm/patches/02-cudaleaks.diff

@@ -1,30 +1,29 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 3102762c..568ac1d0 100644
+index 7800c6e7..be30db23 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -307,6 +307,10 @@ struct llama_client_slot
-     }
- };
+@@ -30,6 +30,10 @@
+ #include <atomic>
+ #include <signal.h>
  
 +#ifdef GGML_USE_CUBLAS
 +extern "C" GGML_CALL void ggml_free_cublas(void);
 +#endif
 +
- struct llama_server_context
- {
-     llama_model *model = nullptr;
-@@ -353,6 +357,10 @@ struct llama_server_context
+ using json = nlohmann::json;
+ 
+ struct server_params
+@@ -353,6 +357,9 @@ struct llama_server_context
              llama_free_model(model);
              model = nullptr;
          }
 +#ifdef GGML_USE_CUBLAS
 +        ggml_free_cublas();
 +#endif
-+
      }
  
      bool load_model(const gpt_params &params_)
-@@ -3093,6 +3101,7 @@ int main(int argc, char **argv)
+@@ -3143,6 +3150,7 @@ int main(int argc, char **argv)
      sigemptyset (&sigint_action.sa_mask);
      sigint_action.sa_flags = 0;
      sigaction(SIGINT, &sigint_action, NULL);
@@ -32,13 +31,8 @@ index 3102762c..568ac1d0 100644
  #elif defined (_WIN32)
      auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
          return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-@@ -3106,3 +3115,4 @@ int main(int argc, char **argv)
-     llama_backend_free();
-     return 0;
- }
-+
 diff --git a/ggml-cuda.cu b/ggml-cuda.cu
-index 96976f24..3543920e 100644
+index 933ebbc4..88a4f664 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
 @@ -39,6 +39,7 @@
@@ -49,30 +43,30 @@ index 96976f24..3543920e 100644
  #define cublasGemmEx hipblasGemmEx
  #define cublasGemmBatchedEx hipblasGemmBatchedEx
  #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -7928,10 +7929,11 @@ GGML_CALL bool ggml_cublas_loaded(void) {
+@@ -7991,10 +7992,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
      return g_cublas_loaded;
  }
  
-+static bool g_cublas_initialized = false;
-+
- GGML_CALL void ggml_init_cublas() {
+-GGML_CALL void ggml_init_cublas() {
 -    static bool initialized = false;
++static bool g_cublas_initialized = false;
  
 -    if (!initialized) {
++GGML_CALL void ggml_init_cublas() {
 +    if (!g_cublas_initialized) {
  
  #ifdef __HIP_PLATFORM_AMD__
          // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -7941,7 +7943,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8004,7 +8005,7 @@ GGML_CALL void ggml_init_cublas() {
  #endif
  
          if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
 -            initialized = true;
 +            g_cublas_initialized = true;
              g_cublas_loaded = false;
+             fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
              return;
-         }
-@@ -8011,7 +8013,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8075,7 +8076,7 @@ GGML_CALL void ggml_init_cublas() {
          // configure logging to stdout
          // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
  
@@ -81,25 +75,30 @@ index 96976f24..3543920e 100644
          g_cublas_loaded = true;
      }
  }
-@@ -11528,3 +11530,17 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
+@@ -11604,3 +11605,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
      }
      return device_count;
  }
 +
++
 +extern "C" GGML_CALL void ggml_free_cublas(void);
 +GGML_CALL void ggml_free_cublas(void) {
 +    for (int id = 0; id < g_device_count; ++id) {
-+#if !defined(GGML_USE_HIPBLAS)
-+        CU_CHECK(cuMemUnmap(g_cuda_pool_addr[id], g_cuda_pool_size[id]));
-+        g_cuda_pool_size[id] = 0;
-+        g_cuda_pool_addr[id] = 0;
++#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
++        if (g_device_caps[id].vmm) {
++            CU_CHECK(cuMemUnmap(g_cuda_pool_addr[id], g_cuda_pool_size[id]));
++            g_cuda_pool_size[id] = 0;
++            g_cuda_pool_addr[id] = 0;
++        }
 +#endif
++        // TODO: free legacy non-vmm memory
++        // destroy cublas handle
 +        CUBLAS_CHECK(cublasDestroy(g_cublas_handles[id]));
 +        g_cublas_handles[id] = nullptr;
 +    }
++
 +    g_cublas_initialized = false;
 +}
-\ No newline at end of file
 diff --git a/ggml-cuda.h b/ggml-cuda.h
 index b1ebd61d..b4c80c2c 100644
 --- a/ggml-cuda.h

llm/patches/02-shutdown.diff

@@ -1,96 +0,0 @@
-diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index a0b46970..7800c6e7 100644
---- a/examples/server/server.cpp
-+++ b/examples/server/server.cpp
-@@ -28,6 +28,7 @@
- #include <chrono>
- #include <condition_variable>
- #include <atomic>
-+#include <signal.h>
- 
- using json = nlohmann::json;
- 
-@@ -2511,6 +2512,9 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
-     }
- }
- 
-+std::function<void(int)> shutdown_handler;
-+inline void signal_handler(int signal) { shutdown_handler(signal); }
-+
- int main(int argc, char **argv)
- {
- #if SERVER_VERBOSE != 1
-@@ -3128,8 +3132,25 @@ int main(int argc, char **argv)
-         std::placeholders::_2,
-         std::placeholders::_3
-     ));
--    llama.queue_tasks.start_loop();
- 
-+    shutdown_handler = [&](int) {
-+        llama.queue_tasks.terminate();
-+    };
-+
-+#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-+    struct sigaction sigint_action;
-+    sigint_action.sa_handler = signal_handler;
-+    sigemptyset (&sigint_action.sa_mask);
-+    sigint_action.sa_flags = 0;
-+    sigaction(SIGINT, &sigint_action, NULL);
-+#elif defined (_WIN32)
-+    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-+        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-+    };
-+    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-+#endif
-+    llama.queue_tasks.start_loop();
-+    svr.stop();
-     t.join();
- 
-     llama_backend_free();
-diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp
-index 54854896..0ee670db 100644
---- a/examples/server/utils.hpp
-+++ b/examples/server/utils.hpp
-@@ -220,6 +220,7 @@ inline std::string format_chatml(std::vector<json> messages)
- struct llama_server_queue {
-     int id = 0;
-     std::mutex mutex_tasks;
-+    bool running;
-     // queues
-     std::vector<task_server> queue_tasks;
-     std::vector<task_server> queue_tasks_deferred;
-@@ -278,9 +279,18 @@ struct llama_server_queue {
-         queue_tasks_deferred.clear();
-     }
- 
--    // Start the main loop. This call is blocking
--    [[noreturn]]
-+    // end the start_loop routine
-+    void terminate() {
-+        {
-+            std::unique_lock<std::mutex> lock(mutex_tasks);
-+            running = false;
-+        }
-+        condition_tasks.notify_all();
-+    }
-+
-+    // Start the main loop.
-     void start_loop() {
-+        running = true;
-         while (true) {
-             // new task arrived
-             LOG_VERBOSE("have new task", {});
-@@ -324,8 +334,12 @@ struct llama_server_queue {
-             {
-                 std::unique_lock<std::mutex> lock(mutex_tasks);
-                 if (queue_tasks.empty()) {
-+                    if (!running) {
-+                        LOG_VERBOSE("ending start_loop", {});
-+                        return;
-+                    }
-                     condition_tasks.wait(lock, [&]{
--                        return !queue_tasks.empty();
-+                        return (!queue_tasks.empty() || !running);
-                     });
-                 }
-             }