feat: Update llama.cpp to ggerganov/llama.cpp@91e84fed6

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]

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

llama_cpp/mtmd_cpp.py

@@ -134,6 +134,23 @@ class mtmd_decoder_pos(Structure):
     ]
 
 
+# struct mtmd_caps {
+#     bool inp_vision;
+#     bool inp_audio;
+# };
+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 +532,13 @@ 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]: