From a8b3a3739c668db8f857b8f90d64af4335da9113 Mon Sep 17 00:00:00 2001
From: Edwin <Edwin>
Date: Thu, 6 Aug 2026 20:49:07 +0100
Subject: [PATCH] Add memory breakdown bindings

---
 llama_cpp/_internals.py | 13 +++++++++++++
 llama_cpp/llama_cpp.py  | 23 +++++++++++++++++++++++
 2 files changed, 36 insertions(+)

diff --git a/llama_cpp/_internals.py b/llama_cpp/_internals.py
index b0fe94d01f..3621f85950 100644
--- a/llama_cpp/_internals.py
+++ b/llama_cpp/_internals.py
@@ -348,6 +348,19 @@ def get_logits(self):
     def get_logits_ith(self, i: int):
         return llama_cpp.llama_get_logits_ith(self.ctx, i)
 
+    def get_memory_breakdown(self):
+        count = ctypes.c_size_t()
+
+        entries = llama_cpp.llama_get_memory_breakdown_entries(
+            self.ctx,
+            ctypes.byref(count),
+        )
+
+        return [
+            entries[i]
+            for i in range(count.value)
+        ]
+
     def get_embeddings(self):
         return llama_cpp.llama_get_embeddings(self.ctx)
 
diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
index 64399fe317..221712e146 100644
--- a/llama_cpp/llama_cpp.py
+++ b/llama_cpp/llama_cpp.py
@@ -589,6 +589,14 @@ class llama_token_data_array(ctypes.Structure):
 
 llama_token_data_array_p = ctypes.POINTER(llama_token_data_array)
 
+class llama_memory_breakdown_entry(ctypes.Structure):
+    _fields_ = [
+        ("name", ctypes.c_char_p),
+        ("model", ctypes.c_size_t),
+        ("context", ctypes.c_size_t),
+        ("compute", ctypes.c_size_t),
+    ]
+
 # typedef bool (*llama_progress_callback)(float progress, void * user_data);
 llama_progress_callback = ctypes.CFUNCTYPE(
     ctypes.c_bool, ctypes.c_float, ctypes.c_void_p
@@ -3250,6 +3258,21 @@ def llama_get_logits_ith(
     llama_get_logits(ctx) + i*n_vocab"""
     ...
 
+# // Get memory breakdown by backend buffer type.
+# LLAMA_API const llama_memory_breakdown_entry * llama_get_memory_breakdown_entries(
+#         const struct llama_context * ctx,
+#         size_t * count);
+@ctypes_function(
+    "llama_get_memory_breakdown_entries",
+    [llama_context_p_ctypes, ctypes.POINTER(ctypes.c_size_t)],
+    ctypes.POINTER(llama_memory_breakdown_entry),
+)
+def llama_get_memory_breakdown_entries(
+    ctx: llama_context_p,
+    count: CtypesPointer[ctypes.c_size_t],
+    /,
+) -> CtypesPointer[llama_memory_breakdown_entry]:
+    ...
 
 # // Get all output token embeddings.
 # // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,