update llama.cpp submodule to 6cdabe6 (#2999)

llm/ext_server/ext_server.cpp

@@ -26,7 +26,7 @@
 #endif // GGML_USE_CUBLAS
 
 // Expose the llama server as a callable extern "C" API
-llama_server_context *llama = NULL;
+server_context *llama = NULL;
 std::thread ext_server_thread;
 bool shutting_down = false;
 std::atomic_int recv_counter;
@@ -57,7 +57,7 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
   err->id = 0;
   err->msg[0] = '\0';
   try {
-    llama = new llama_server_context;
+    llama = new server_context;
     gpt_params params;
     params.n_ctx = sparams->n_ctx;
     params.n_batch = sparams->n_batch;
@@ -144,13 +144,13 @@ void llama_server_start() {
       LOG_TEE("llama server main loop starting\n");
       ggml_time_init();
       llama->queue_tasks.on_new_task(std::bind(
-        &llama_server_context::process_single_task, llama, std::placeholders::_1));
+        &server_context::process_single_task, llama, std::placeholders::_1));
       llama->queue_tasks.on_finish_multitask(std::bind(
-        &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
+        &server_context::on_finish_multitask, llama, std::placeholders::_1));
       llama->queue_tasks.on_run_slots(std::bind(
-        &llama_server_context::update_slots, llama));
+        &server_context::update_slots, llama));
       llama->queue_results.on_multitask_update(std::bind(
-          &llama_server_queue::update_multitask,
+          &server_queue::update_multitask,
           &llama->queue_tasks,
           std::placeholders::_1,
           std::placeholders::_2,
@@ -198,7 +198,7 @@ void llama_server_completion(const char *json_req, ext_server_resp_t *resp) {
     json data = json::parse(json_req);
     resp->id = llama->queue_tasks.get_new_id();
     llama->queue_results.add_waiting_task_id(resp->id);
-    llama->request_completion(resp->id, data, false, false, -1);
+    llama->request_completion(resp->id, -1, data, false, false);
   } catch (std::exception &e) {
     snprintf(resp->msg, resp->msg_len, "exception %s", e.what());
   } catch (...) {
@@ -216,9 +216,9 @@ void llama_server_completion_next_result(const int task_id,
   std::string result_json;
   try {
     atomicRecv ar(recv_counter);
-    task_result result = llama->queue_results.recv(task_id);
+    server_task_result result = llama->queue_results.recv(task_id);
     result_json =
-        result.result_json.dump(-1, ' ', false, json::error_handler_t::replace);
+        result.data.dump(-1, ' ', false, json::error_handler_t::replace);
     resp->id = result.id;
     resp->stop = result.stop;
     resp->error = result.error;
@@ -363,10 +363,10 @@ void llama_server_embedding(const char *json_req, char **json_resp,
     }
     const int task_id = llama->queue_tasks.get_new_id();
     llama->queue_results.add_waiting_task_id(task_id);
-    llama->request_completion(task_id, {{"prompt", prompt}, {"n_predict", 0}}, false, true, -1);
+    llama->request_completion(task_id, -1, {{"prompt", prompt}, {"n_predict", 0}}, false, true);
     atomicRecv ar(recv_counter);
-    task_result result = llama->queue_results.recv(task_id);
-    std::string result_json = result.result_json.dump();
+    server_task_result result = llama->queue_results.recv(task_id);
+    std::string result_json = result.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());

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit c29af7e2252d288f2ea58a7d437c1cb7c0abf160
+Subproject commit 6cdabe652695167263c8b447520987b11856f7ca

llm/patches/01-cache.diff

@@ -1,21 +1,19 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 2b2f4a0f..afac49af 100644
+index f255ad76..914ecfdd 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -997,13 +997,15 @@ struct llama_server_context
-                 slot.n_sent_text += result.text_to_send.size();
+@@ -1101,12 +1101,13 @@ struct server_context {
                  // add the token to slot queue and cache
              }
+ 
 -            slot.add_token_string(result);
-+
-             if (slot.params.stream)
-             {
+             if (slot.params.stream) {
                  send_partial_response(slot, result);
              }
          }
  
 +        slot.add_token_string(result);
 +
-         if (incomplete)
-         {
+         if (incomplete) {
              slot.has_next_token = true;
+         }

llm/patches/02-cudaleaks.diff

@@ -1,9 +1,9 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 2b2f4a0f..25857bdd 100644
+index f255ad76..5b83acb1 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -31,6 +31,10 @@
- #include <atomic>
+@@ -28,6 +28,10 @@
+ #include <thread>
  #include <signal.h>
  
 +#ifdef GGML_USE_CUBLAS
@@ -12,18 +12,19 @@ index 2b2f4a0f..25857bdd 100644
 +
  using json = nlohmann::json;
  
- struct server_params {
-@@ -363,6 +367,9 @@ struct llama_server_context
+ bool server_verbose = false;
+@@ -648,6 +652,10 @@ struct server_context {
              llama_free_model(model);
              model = nullptr;
          }
++
 +#ifdef GGML_USE_CUBLAS
 +        ggml_free_cublas();
 +#endif
      }
  
-     bool load_model(const gpt_params &params_)
-@@ -3494,6 +3501,7 @@ int main(int argc, char **argv)
+     bool load_model(const gpt_params & params_) {
+@@ -3339,6 +3347,7 @@ int main(int argc, char ** argv) {
      sigemptyset (&sigint_action.sa_mask);
      sigint_action.sa_flags = 0;
      sigaction(SIGINT, &sigint_action, NULL);
@@ -32,7 +33,7 @@ index 2b2f4a0f..25857bdd 100644
      auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
          return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
 diff --git a/ggml-cuda.cu b/ggml-cuda.cu
-index 0c6501e9..75c12723 100644
+index 72bcec8c..50a45e3d 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
 @@ -43,6 +43,7 @@
@@ -43,7 +44,7 @@ index 0c6501e9..75c12723 100644
  #define cublasGemmEx hipblasGemmEx
  #define cublasGemmBatchedEx hipblasGemmBatchedEx
  #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -8694,10 +8695,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
+@@ -8751,10 +8752,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
      return g_cublas_loaded;
  }
  
@@ -57,7 +58,7 @@ index 0c6501e9..75c12723 100644
  
  #ifdef __HIP_PLATFORM_AMD__
          // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -8707,7 +8708,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8764,7 +8765,7 @@ GGML_CALL void ggml_init_cublas() {
  #endif
  
          if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
@@ -66,7 +67,7 @@ index 0c6501e9..75c12723 100644
              g_cublas_loaded = false;
              fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
              return;
-@@ -8778,7 +8779,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8835,7 +8836,7 @@ GGML_CALL void ggml_init_cublas() {
          // configure logging to stdout
          // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
  
@@ -75,7 +76,7 @@ index 0c6501e9..75c12723 100644
          g_cublas_loaded = true;
      }
  }
-@@ -12345,3 +12346,22 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
+@@ -12490,3 +12491,22 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
      }
      return device_count;
  }