Update llama.cpp

Author: abetlen

llama_cpp/llama_cpp.py

@@ -79,6 +79,10 @@ def _load_shared_library(lib_base_name: str):
 
 # llama.h bindings
 
+GGML_USE_CUBLAS = hasattr(_lib, "ggml_init_cublas")
+GGML_CUDA_MAX_DEVICES = ctypes.c_int(16)
+LLAMA_MAX_DEVICES = GGML_CUDA_MAX_DEVICES if GGML_USE_CUBLAS else ctypes.c_int(1)
+
 # #define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
 LLAMA_FILE_MAGIC_GGJT = ctypes.c_uint(0x67676A74)
 # #define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
@@ -142,9 +146,12 @@ class llama_token_data_array(Structure):
 
 
 # struct llama_context_params {
-#     int n_ctx;        // text context
-#     int n_gpu_layers; // number of layers to store in VRAM
-#     int seed;         // RNG seed, -1 for random
+#     int n_ctx;                             // text context
+#     int n_batch;                           // prompt processing batch size
+#     int n_gpu_layers;                      // number of layers to store in VRAM
+#     int main_gpu;                          // the GPU that is used for scratch and small tensors
+#     float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+#     int seed;                              // RNG seed, -1 for random
 
 #     bool f16_kv;     // use fp16 for KV cache
 #     bool logits_all; // the llama_eval() call computes all logits, not just the last one
@@ -153,7 +160,6 @@ class llama_token_data_array(Structure):
 #     bool use_mlock;  // force system to keep model in RAM
 #     bool embedding;  // embedding mode only
 
-
 #     // called with a progress value between 0 and 1, pass NULL to disable
 #     llama_progress_callback progress_callback;
 #     // context pointer passed to the progress callback
@@ -162,7 +168,10 @@ class llama_token_data_array(Structure):
 class llama_context_params(Structure):
     _fields_ = [
         ("n_ctx", c_int),
+        ("n_batch", c_int),
         ("n_gpu_layers", c_int),
+        ("main_gpu", c_int),
+        ("tensor_split", c_float * LLAMA_MAX_DEVICES.value),
         ("seed", c_int),
         ("f16_kv", c_bool),
         (