update llama.cpp submodule to ceca1ae (#3064)

llm/ext_server/ext_server.cpp

@@ -26,7 +26,7 @@
 #endif // GGML_USE_CUBLAS
 
 // Expose the llama server as a callable extern "C" API
-server_context *llama = NULL;
+llama_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 server_context;
+    llama = new llama_server_context;
     gpt_params params;
     params.n_ctx = sparams->n_ctx;
     params.n_batch = sparams->n_batch;
@@ -125,7 +125,7 @@ void llama_server_init(ext_server_params *sparams, ext_server_resp_t *err) {
       return;
     }
 
-    llama->init();
+    llama->initialize();
   } catch (std::exception &e) {
     err->id = -1;
     snprintf(err->msg, err->msg_len, "exception %s", e.what());
@@ -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(
-        &server_context::process_single_task, llama, std::placeholders::_1));
+        &llama_server_context::process_single_task, llama, std::placeholders::_1));
       llama->queue_tasks.on_finish_multitask(std::bind(
-        &server_context::on_finish_multitask, llama, std::placeholders::_1));
+        &llama_server_context::on_finish_multitask, llama, std::placeholders::_1));
       llama->queue_tasks.on_run_slots(std::bind(
-        &server_context::update_slots, llama));
+        &llama_server_context::update_slots, llama));
       llama->queue_results.on_multitask_update(std::bind(
-          &server_queue::update_multitask,
+          &llama_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, -1, data, false, false);
+    llama->request_completion(resp->id, data, false, false, -1);
   } 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);
-    server_task_result result = llama->queue_results.recv(task_id);
+    task_result result = llama->queue_results.recv(task_id);
     result_json =
-        result.data.dump(-1, ' ', false, json::error_handler_t::replace);
+        result.result_json.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, -1, {{"prompt", prompt}, {"n_predict", 0}}, false, true);
+    llama->request_completion(task_id, {{"prompt", prompt}, {"n_predict", 0}}, false, true, -1);
     atomicRecv ar(recv_counter);
-    server_task_result result = llama->queue_results.recv(task_id);
+    task_result result = llama->queue_results.recv(task_id);
-    std::string result_json = result.data.dump();
+    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());

llm/generate/gen_darwin.sh

@@ -18,19 +18,6 @@ sign() {
     fi
 }
 
-# bundle_metal bundles ggml-common.h and ggml-metal.metal into a single file
-bundle_metal() {
-    grep -v '#include "ggml-common.h"' "${LLAMACPP_DIR}/ggml-metal.metal" | grep -v '#pragma once' > "${LLAMACPP_DIR}/ggml-metal.metal.temp"
-    echo '#define GGML_COMMON_IMPL_METAL' > "${LLAMACPP_DIR}/ggml-metal.metal"
-    cat "${LLAMACPP_DIR}/ggml-common.h" | grep -v '#pragma once' >> "${LLAMACPP_DIR}/ggml-metal.metal"
-    cat  "${LLAMACPP_DIR}/ggml-metal.metal.temp" >> "${LLAMACPP_DIR}/ggml-metal.metal"
-    rm "${LLAMACPP_DIR}/ggml-metal.metal.temp"
-}
-
-cleanup_metal() {
-    (cd ${LLAMACPP_DIR} && git checkout ggml-metal.metal)
-}
-
 COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.0 -DCMAKE_SYSTEM_NAME=Darwin"
 
 case "${GOARCH}" in
@@ -76,11 +63,9 @@ case "${GOARCH}" in
     CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
     BUILD_DIR="${LLAMACPP_DIR}/build/darwin/${ARCH}/metal"
     EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
-    bundle_metal
     build
     sign ${LLAMACPP_DIR}/build/darwin/${ARCH}/metal/lib/libext_server.dylib
     compress_libs
-    cleanup_metal
     ;;
 *)
     echo "GOARCH must be set"

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit 77d1ac7e00bf049b9f2bba1b5a310a78318c49c4
+Subproject commit ceca1aef0738b57951cd12c603c3477e75312dec

llm/patches/01-cache.diff

@@ -1,19 +1,21 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index f255ad76..914ecfdd 100644
+index 8fe5e0b1..3e82acb9 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -1101,12 +1101,13 @@ struct server_context {
+@@ -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);
-             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

@@ -1,10 +1,10 @@
 diff --git a/examples/server/server.cpp b/examples/server/server.cpp
-index b14cca61..02bfd4b1 100644
+index 8fe5e0b1..53bf39c1 100644
 --- a/examples/server/server.cpp
 +++ b/examples/server/server.cpp
-@@ -29,6 +29,10 @@
+@@ -31,6 +31,10 @@
+ #include <atomic>
  #include <signal.h>
- #include <memory>
  
 +#ifdef GGML_USE_CUBLAS
 +extern "C" GGML_CALL void ggml_free_cublas(void);
@@ -12,8 +12,8 @@ index b14cca61..02bfd4b1 100644
 +
  using json = nlohmann::json;
  
- bool server_verbose = false;
+ struct server_params {
-@@ -664,6 +668,10 @@ struct server_context {
+@@ -363,6 +367,10 @@ struct llama_server_context
              llama_free_model(model);
              model = nullptr;
          }
@@ -23,8 +23,8 @@ index b14cca61..02bfd4b1 100644
 +#endif
      }
  
-     bool load_model(const gpt_params & params_) {
+     bool load_model(const gpt_params &params_)
-@@ -3499,6 +3507,7 @@ int main(int argc, char ** argv) {
+@@ -3543,6 +3551,7 @@ int main(int argc, char **argv)
      sigemptyset (&sigint_action.sa_mask);
      sigint_action.sa_flags = 0;
      sigaction(SIGINT, &sigint_action, NULL);
@@ -33,10 +33,10 @@ index b14cca61..02bfd4b1 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 c207ff87..945708a4 100644
+index 72bcec8c..6c934e8c 100644
 --- a/ggml-cuda.cu
 +++ b/ggml-cuda.cu
-@@ -46,6 +46,7 @@
+@@ -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
@@ -44,7 +44,7 @@ index c207ff87..945708a4 100644
  #define cublasGemmEx hipblasGemmEx
  #define cublasGemmBatchedEx hipblasGemmBatchedEx
  #define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-@@ -8014,10 +8015,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
+@@ -8751,10 +8752,10 @@ GGML_CALL bool ggml_cublas_loaded(void) {
      return g_cublas_loaded;
  }
  
@@ -58,7 +58,7 @@ index c207ff87..945708a4 100644
  
  #ifdef __HIP_PLATFORM_AMD__
          // Workaround for a rocBLAS bug when using multiple graphics cards:
-@@ -8027,7 +8028,7 @@ GGML_CALL void ggml_init_cublas() {
+@@ -8764,7 +8765,7 @@ GGML_CALL void ggml_init_cublas() {
  #endif
  
          if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
@@ -67,7 +67,7 @@ index c207ff87..945708a4 100644
              g_cublas_loaded = false;
              fprintf(stderr, "%s: no " GGML_CUDA_NAME " devices found, " GGML_CUDA_NAME " will be disabled\n", __func__);
              return;
-@@ -8098,7 +8099,7 @@ 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));
  
@@ -76,7 +76,7 @@ index c207ff87..945708a4 100644
          g_cublas_loaded = true;
      }
  }
-@@ -11753,3 +11754,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
+@@ -12490,3 +12491,23 @@ GGML_CALL int ggml_backend_cuda_reg_devices() {
      }
      return device_count;
  }
@@ -100,6 +100,7 @@ index c207ff87..945708a4 100644
 +
 +    g_cublas_initialized = false;
 +}
+\ No newline at end of file
 diff --git a/ggml-cuda.h b/ggml-cuda.h
 index b1ebd61d..6dd58ddf 100644
 --- a/ggml-cuda.h

llm/patches/03-locale.diff

@@ -1,13 +0,0 @@
-diff --git a/llama.cpp b/llama.cpp
-index b19616e8..519b9602 100644
---- a/llama.cpp
-+++ b/llama.cpp
-@@ -9938,7 +9938,7 @@ struct llm_tokenizer_wpm {
-     }
- 
-     uint32_t to_lower(uint32_t code) {
--        static const std::locale locale("en_US.UTF-8");
-+        static const std::locale locale("");
- #if defined(_WIN32)
-         if (code > 0xFFFF) {
-             return code;