update llama.cpp submodule to c29af7e (#2868)

llm/ext_server/ext_server.cpp

@@ -146,9 +146,9 @@ void llama_server_start() {
       llama->queue_tasks.on_new_task(std::bind(
         &llama_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));
-      llama->queue_tasks.on_all_tasks_finished(std::bind(
-          &llama_server_context::run_on_all_tasks_finished, llama));
+        &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
+      llama->queue_tasks.on_run_slots(std::bind(
+        &llama_server_context::update_slots, llama));
       llama->queue_results.on_multitask_update(std::bind(
           &llama_server_queue::update_multitask,
           &llama->queue_tasks,

llm/patches/01-cache.diff

@@ -1,9 +1,9 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index d86d7e04..2694e92e 100644
+index 2b2f4a0f..afac49af 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -901,13 +901,15 @@ struct llama_server_context
-                 slot.sent_count += result.text_to_send.size();
+@@ -997,13 +997,15 @@ struct llama_server_context
+                 slot.n_sent_text += result.text_to_send.size();
                  // add the token to slot queue and cache
              }
 -            slot.add_token_string(result);

llm/patches/02-cudaleaks.diff

@@ -1,8 +1,8 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index 7800c6e7..be30db23 100644
+index 2b2f4a0f..25857bdd 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -30,6 +30,10 @@
+@@ -31,6 +31,10 @@
  #include <atomic>
  #include <signal.h>
 
@@ -12,8 +12,8 @@ index 7800c6e7..be30db23 100644
 +
  using json = nlohmann::json;
 
- struct server_params
-@@ -353,6 +357,9 @@ struct llama_server_context
+ struct server_params {
+@@ -363,6 +367,9 @@ struct llama_server_context
              llama_free_model(model);
              model = nullptr;
          }
@@ -23,7 +23,7 @@ index 7800c6e7..be30db23 100644
      }
 
      bool load_model(const gpt_params &params_)
-@@ -3143,6 +3150,7 @@ int main(int argc, char **argv)
+@@ -3494,6 +3501,7 @@ int main(int argc, char **argv)
      sigemptyset (&sigint_action.sa_mask);
      sigint_action.sa_flags = 0;
      sigaction(SIGINT, &sigint_action, NULL);
@@ -32,18 +32,18 @@ index 7800c6e7..be30db23 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 933ebbc4..88a4f664 100644
+index 0c6501e9..75c12723 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
-@@ -39,6 +39,7 @@
+@@ -43,6 +43,7 @@
  #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
  #define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
  #define cublasCreate hipblasCreate
 +#define cublasDestroy hipblasDestroy
  #define cublasGemmEx hipblasGemmEx
  #define cublasGemmBatchedEx hipblasGemmBatchedEx
  #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -7991,10 +7992,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
+@@ -8694,10 +8695,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
      return g_cublas_loaded;
  }
 
@@ -57,7 +57,7 @@ index 933ebbc4..88a4f664 100644
 
  #ifdef __HIP_PLATFORM_AMD__
          // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -8004,7 +8005,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8707,7 +8708,7 @@ GGML_CALL void ggml_init_cublas() {
  #endif
 
          if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
@@ -66,7 +66,7 @@ index 933ebbc4..88a4f664 100644
              g_cublas_loaded = false;
              fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
              return;
-@@ -8075,7 +8076,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8778,7 +8779,7 @@ GGML_CALL void ggml_init_cublas() {
          // configure logging to stdout
          // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
 
@@ -75,12 +75,11 @@ index 933ebbc4..88a4f664 100644
          g_cublas_loaded = true;
      }
  }
-@@ -11604,3 +11605,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
+@@ -12345,3 +12346,22 @@ 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) {
@@ -100,16 +99,16 @@ index 933ebbc4..88a4f664 100644
 +    g_cublas_initialized = false;
 +}
 diff --git a/ggml-cuda.h b/ggml-cuda.h
-index b1ebd61d..b4c80c2c 100644
+index b1ebd61d..6dd58ddf 100644
 --- a/ggml-cuda.h
 +++ b/ggml-cuda.h
-@@ -20,6 +20,9 @@ extern "C" {
- // Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
- GGML_API GGML_CALL void   ggml_init_cublas(void);
+@@ -23,6 +23,9 @@ GGML_API GGML_CALL void   ggml_init_cublas(void);
+ // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
+ GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
 
 +// Release CUDA resources
 +GGML_API GGML_CALL void   ggml_free_cublas(void);
 +
- // Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
- GGML_API GGML_CALL bool   ggml_cublas_loaded(void);
+ GGML_API GGML_CALL void * ggml_cuda_host_malloc(size_t size);
+ GGML_API GGML_CALL void   ggml_cuda_host_free(void * ptr);