update llama.cpp submodule to ceca1ae (#3064)

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;
-@@ -664,6 +668,10 @@ struct server_context {
+ struct server_params {
+@@ -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_) {
-@@ -3499,6 +3507,7 @@ int main(int argc, char ** argv) {
+     bool load_model(const gpt_params &params_)
+@@ -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;