From 906afdd98c2b7f79cb0fd2cc9d67caebb873a66c Mon Sep 17 00:00:00 2001
From: Laurent-Philippe Albou <contact@abstractcore.ai>
Date: Sun, 1 Nov 2026 14:37:36 +0100
Subject: [PATCH] feat: update llama.cpp to b12032 (dde879bc4)

- vendor/llama.cpp: 0c1e57098 -> dde879bc4 (tag b12032)
- llama_context_params: add the new moe_cache_size (size_t) field after type_v;
  without it the struct is one word short and every field after type_v
  (abort_callback, embeddings, kv_unified, samplers, ...) is read at the wrong
  offset by llama_context_default_params / llama_init_from_model
- llama_perf_context_data: add n_reallocs_same / n_reallocs_diff

No other llama.h API change between the two commits (the declared LLAMA_API
functions are identical). Brings the decision-model support merged in
ggml-org/llama.cpp#29818 (<arch>.decision.* metadata, projected embeddings
output for kev/laya) to the Python bindings.
---
 llama_cpp/llama_cpp.py | 7 +++++++
 vendor/llama.cpp       | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
index 94e5aede3..ac00bfb00 100644
--- a/llama_cpp/llama_cpp.py
+++ b/llama_cpp/llama_cpp.py
@@ -971,6 +971,8 @@ class llama_sampler_seq_config(ctypes.Structure):
 
 #     enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
 #     enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
+#
+#     size_t moe_cache_size; // device cache in bytes for the experts kept in host memory, split among the devices like the layers, 0 = disabled [EXPERIMENTAL]
 
 #     // Abort callback
 #     // if it returns true, execution of llama_decode() will be aborted
@@ -1027,6 +1029,7 @@ class llama_context_params(ctypes.Structure):
         cb_eval_user_data (ctypes.ctypes.c_void_p): user data for cb_eval
         type_k (int): data type for K cache
         type_v (int): data type for V cache
+        moe_cache_size (int): device cache in bytes for the experts kept in host memory, 0 = disabled
         abort_callback (ggml_abort_callback): abort callback if it returns true, execution of llama_decode() will be aborted
         abort_callback_data (ctypes.ctypes.c_void_p): data for abort_callback
         embeddings (bool): if true, extract embeddings (together with logits)
@@ -1067,6 +1070,7 @@ class llama_context_params(ctypes.Structure):
         cb_eval_user_data: ctypes.c_void_p
         type_k: int
         type_v: int
+        moe_cache_size: int
         abort_callback: Callable[[ctypes.c_void_p], bool]
         abort_callback_data: ctypes.c_void_p
         embeddings: bool
@@ -1106,6 +1110,7 @@ class llama_context_params(ctypes.Structure):
         ("cb_eval_user_data", ctypes.c_void_p),
         ("type_k", ctypes.c_int),
         ("type_v", ctypes.c_int),
+        ("moe_cache_size", ctypes.c_size_t),
         ("abort_callback", ggml_abort_callback),
         ("abort_callback_data", ctypes.c_void_p),
         ("embeddings", ctypes.c_bool),
@@ -5248,6 +5253,8 @@ class llama_perf_context_data(ctypes.Structure):
         ("n_p_eval", ctypes.c_int32),
         ("n_eval", ctypes.c_int32),
         ("n_reused", ctypes.c_int32),
+        ("n_reallocs_same", ctypes.c_int32),
+        ("n_reallocs_diff", ctypes.c_int32),
     ]
 
 
diff --git a/vendor/llama.cpp b/vendor/llama.cpp
index 0c1e57098..dde879bc4 160000
--- a/vendor/llama.cpp
+++ b/vendor/llama.cpp
@@ -1 +1 @@
-Subproject commit 0c1e57098bba43ac29e6e3b677cdceebdd22334f
+Subproject commit dde879bc4f954b809892fd0d660515f42536394a