From 2305061e64c4ebaccaade8d31e46da234f984e0c Mon Sep 17 00:00:00 2001
From: abetlen <abetlen@gmail.com>
Date: Fri, 15 May 2026 01:40:02 -0700
Subject: [PATCH 1/2] feat: update llama.cpp to 91e84fed6

---
 CHANGELOG.md           |  1 +
 llama_cpp/llama_cpp.py |  3 +++
 llama_cpp/mtmd_cpp.py  | 20 ++++++++++++++++++++
 vendor/llama.cpp       |  2 +-
 4 files changed, 25 insertions(+), 1 deletion(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 900176ea1..a0b63061c 100644
--- a/CHANGELOG.md
+++ b/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 ggerganov/llama.cpp@91e84fed6 and sync Python bindings
 - fix: Enable unified KV cache for embedding contexts to preserve full per-sequence context in batch embedding calls by @SanjanaB123 in #2217
 
 ## [0.3.23]
diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py
index a5ec5d190..a9c32a15b 100644
--- a/llama_cpp/llama_cpp.py
+++ b/llama_cpp/llama_cpp.py
@@ -2837,6 +2837,9 @@ def llama_state_seq_load_file(
 ) -> int: ...
 
 
+# define LLAMA_STATE_SEQ_FLAGS_NONE 0
+LLAMA_STATE_SEQ_FLAGS_NONE = 0
+
 # for backwards-compat
 # define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
 LLAMA_STATE_SEQ_FLAGS_SWA_ONLY = 1
diff --git a/llama_cpp/mtmd_cpp.py b/llama_cpp/mtmd_cpp.py
index 485dc5d8c..4c5584223 100644
--- a/llama_cpp/mtmd_cpp.py
+++ b/llama_cpp/mtmd_cpp.py
@@ -134,6 +134,19 @@ class mtmd_decoder_pos(Structure):
     ]
 
 
+class mtmd_caps(Structure):
+    """Capabilities exposed by an mmproj file."""
+
+    if TYPE_CHECKING:
+        inp_vision: bool
+        inp_audio: bool
+
+    _fields_ = [
+        ("inp_vision", c_bool),
+        ("inp_audio", c_bool),
+    ]
+
+
 ################################################
 # mtmd.h functions
 ################################################
@@ -515,6 +528,13 @@ def mtmd_get_output_embd(ctx: mtmd_context_p, /) -> Optional[CtypesArray[c_float
     ...
 
 
+# MTMD_API struct mtmd_caps mtmd_get_cap_from_file(const char * mmproj_fname);
+@ctypes_function("mtmd_get_cap_from_file", [c_char_p], mtmd_caps)
+def mtmd_get_cap_from_file(mmproj_fname: bytes, /) -> mtmd_caps:
+    """Get mmproj capabilities without initializing a full MTMD context."""
+    ...
+
+
 # MTMD_API mtmd_input_chunks * mtmd_test_create_input_chunks(void);
 @ctypes_function("mtmd_test_create_input_chunks", [], mtmd_input_chunks_p_ctypes)
 def mtmd_test_create_input_chunks() -> Optional[mtmd_input_chunks_p]:
diff --git a/vendor/llama.cpp b/vendor/llama.cpp
index 7d442abf5..91e84fed6 160000
--- a/vendor/llama.cpp
+++ b/vendor/llama.cpp
@@ -1 +1 @@
-Subproject commit 7d442abf5c6244117fd5a1dc51a5d19f00792491
+Subproject commit 91e84fed64329cd96202d68220724a1d92f5ec1f

From c58751b83e60fa05360d65188fde476eeb58ce5f Mon Sep 17 00:00:00 2001
From: abetlen <abetlen@gmail.com>
Date: Fri, 15 May 2026 01:44:18 -0700
Subject: [PATCH 2/2] chore: document mtmd_caps c declaration

---
 llama_cpp/mtmd_cpp.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/llama_cpp/mtmd_cpp.py b/llama_cpp/mtmd_cpp.py
index 4c5584223..f2b0ed2de 100644
--- a/llama_cpp/mtmd_cpp.py
+++ b/llama_cpp/mtmd_cpp.py
@@ -134,6 +134,10 @@ class mtmd_decoder_pos(Structure):
     ]
 
 
+# struct mtmd_caps {
+#     bool inp_vision;
+#     bool inp_audio;
+# };
 class mtmd_caps(Structure):
     """Capabilities exposed by an mmproj file."""