feat: update llama.cpp to 92e854ab8

CHANGELOG.md

@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- feat: update llama.cpp to ggml-org/llama.cpp@92e854ab8
 - fix: preserve recurrent/hybrid model state when the full prompt is already cached by @allthatido and @abetlen in #2306
 
 ## [0.3.31]

llama_cpp/llama_cpp.py

@@ -1744,6 +1744,11 @@ def llama_model_n_embd_out(model: llama_model_p, /) -> int:
 def llama_model_n_layer(model: llama_model_p, /) -> int: ...
 
 
+# LLAMA_API int32_t llama_model_n_layer_nextn(const struct llama_model * model);
+@ctypes_function("llama_model_n_layer_nextn", [llama_model_p_ctypes], ctypes.c_int32)
+def llama_model_n_layer_nextn(model: llama_model_p, /) -> int: ...
+
+
 # LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
 @ctypes_function("llama_model_n_head", [llama_model_p_ctypes], ctypes.c_int32)
 def llama_model_n_head(model: llama_model_p, /) -> int: ...

llama_cpp/llama_cpp_ext.py

@@ -62,6 +62,25 @@ def llama_set_embeddings_nextn(
     ...
 
 
+# LLAMA_API void llama_set_nextn_layer_offset(struct llama_context * ctx, int32_t offset);
+@_ctypes_function_from_names(
+    (
+        "llama_set_nextn_layer_offset",
+        "_Z28llama_set_nextn_layer_offsetP13llama_contexti",
+        "?llama_set_nextn_layer_offset@@YAXPEAUllama_context@@H@Z",
+    ),
+    [llama_cpp.llama_context_p_ctypes, ctypes.c_int32],
+    None,
+)
+def llama_set_nextn_layer_offset(
+    ctx: llama_cpp.llama_context_p,
+    offset: Union[ctypes.c_int32, int],
+    /,
+):
+    """Select which appended NextN block the decoder MTP graph runs."""
+    ...
+
+
 # LLAMA_API float * llama_get_embeddings_nextn(struct llama_context * ctx);
 @_ctypes_function_from_names(
     (

llama_cpp/mtmd_cpp.py

@@ -20,6 +20,7 @@
 )
 import pathlib
 from typing import (
+    Callable,
     Union,
     NewType,
     Optional,
@@ -84,6 +85,8 @@
 MTMD_INPUT_CHUNK_TYPE_IMAGE = 1
 MTMD_INPUT_CHUNK_TYPE_AUDIO = 2
 
+mtmd_progress_callback = CFUNCTYPE(c_bool, c_float, c_void_p)
+
 
 # Structures
 class mtmd_context_params(Structure):
@@ -106,6 +109,8 @@ class mtmd_context_params(Structure):
         cb_eval: llama_cpp.ggml_backend_sched_eval_callback
         cb_eval_user_data: c_void_p
         batch_max_tokens: int
+        progress_callback: Callable[[float, c_void_p], bool]
+        progress_callback_user_data: c_void_p
 
     _fields_ = [
         ("use_gpu", c_bool),
@@ -120,6 +125,8 @@ class mtmd_context_params(Structure):
         ("cb_eval", llama_cpp.ggml_backend_sched_eval_callback),
         ("cb_eval_user_data", c_void_p),
         ("batch_max_tokens", c_int),
+        ("progress_callback", mtmd_progress_callback),
+        ("progress_callback_user_data", c_void_p),
     ]