From 71e3e4c435826677ff671f1d82748fe1dd4d64e1 Mon Sep 17 00:00:00 2001
From: Andrei Betlen <abetlen@gmail.com>
Date: Wed, 31 Jan 2024 10:41:42 -0500
Subject: [PATCH] Update llama.cpp

---
 llama_cpp/llama_cpp.py | 41 ++++++++++++++++++++++++++++++++++-------
 vendor/llama.cpp       |  2 +-
 2 files changed, 35 insertions(+), 8 deletions(-)

diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
index 2168579..431a99f 100644
--- a/llama_cpp/llama_cpp.py
+++ b/llama_cpp/llama_cpp.py
@@ -98,7 +98,7 @@ ggml_backend_sched_eval_callback = ctypes.CFUNCTYPE(c_bool, c_void_p, c_bool, c_
 # llama.h bindings
 
 _lib.llama_max_devices.argtypes = []
-_lib.llama_max_devices.restype = ctypes.c_int32
+_lib.llama_max_devices.restype = ctypes.c_size_t
 
 LLAMA_MAX_DEVICES = _lib.llama_max_devices()
 
@@ -390,7 +390,7 @@ class llama_model_kv_override(Structure):
 #     // LLAMA_SPLIT_LAYER: ignored
 #     int32_t main_gpu;
 
-#     // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
+#     // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
 #     const float * tensor_split;
 
 #     // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
@@ -417,7 +417,7 @@ class llama_model_params(Structure):
         n_gpu_layers (int): number of layers to store in VRAM
         split_mode (int): how to split the model across multiple GPUs
         main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored
-        tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
+        tensor_split (ctypes.Array[ctypes.c_float]): proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         progress_callback (llama_progress_callback): called with a progress value between 0.0 and 1.0. Pass NULL to disable. If the provided progress_callback returns true, model loading continues. If it returns false, model loading is immediately aborted.
         progress_callback_user_data (ctypes.c_void_p): context pointer passed to the progress callback
         kv_overrides (ctypes.Array[llama_model_kv_override]): override key-value pairs of the model meta data
@@ -760,16 +760,43 @@ _lib.llama_time_us.argtypes = []
 _lib.llama_time_us.restype = ctypes.c_int64
 
 
-# LLAMA_API int32_t  llama_max_devices(void);
+# LLAMA_API size_t llama_max_devices(void);
 def llama_max_devices() -> int:
     return _lib.llama_max_devices()
 
 
 _lib.llama_max_devices.argtypes = []
-_lib.llama_max_devices.restype = ctypes.c_int32
+_lib.llama_max_devices.restype = ctypes.c_size_t
 
 
-# LLAMA_API bool llama_mmap_supported (void);
+# LLAMA_API bool llama_supports_mmap       (void);
+def llama_supports_mmap() -> bool:
+    return _lib.llama_supports_mmap()
+
+
+_lib.llama_supports_mmap.argtypes = []
+_lib.llama_supports_mmap.restype = c_bool
+
+
+# LLAMA_API bool llama_supports_mlock      (void);
+def llama_supports_mlock() -> bool:
+    return _lib.llama_supports_mlock()
+
+
+_lib.llama_supports_mlock.argtypes = []
+_lib.llama_supports_mlock.restype = c_bool
+
+
+# LLAMA_API bool llama_supports_gpu_offload(void);
+def llama_supports_gpu_offload() -> bool:
+    return _lib.llama_supports_gpu_offload()
+
+
+_lib.llama_supports_gpu_offload.argtypes = []
+_lib.llama_supports_gpu_offload.restype = c_bool
+
+
+# LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead");
 def llama_mmap_supported() -> bool:
     return _lib.llama_mmap_supported()
 
@@ -778,7 +805,7 @@ _lib.llama_mmap_supported.argtypes = []
 _lib.llama_mmap_supported.restype = c_bool
 
 
-# LLAMA_API bool llama_mlock_supported(void);
+# LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead");
 def llama_mlock_supported() -> bool:
     return _lib.llama_mlock_supported()
 
diff --git a/vendor/llama.cpp b/vendor/llama.cpp
index 1560630..5cb04db 160000
--- a/vendor/llama.cpp
+++ b/vendor/llama.cpp
@@ -1 +1 @@
-Subproject commit 15606309a05ccf7fadbaad5538cb7c32acb1e06b
+Subproject commit 5cb04dbc16d1da38c8fdcc0111b40e67d00dd1c3