feat: Update llama.cpp to ggerganov/llama.cpp@5d6f18a63

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@5d6f18a63 and sync Python bindings
 - fix: Correct batched embedding outputs for multi-sequence `embed()` calls by @Anai-Guo in #2205
 - fix: Configure embedding contexts with enough sequence slots for batched `embed()` calls
 

llama_cpp/llama_cpp.py

@@ -199,6 +199,8 @@ def _warn_deprecated(symbol: str, hint: str) -> None:
 llama_token_p = ctypes.POINTER(llama_token)
 # typedef int32_t llama_seq_id;
 llama_seq_id = ctypes.c_int32
+# typedef uint32_t llama_state_seq_flags;
+llama_state_seq_flags = ctypes.c_uint32
 
 
 # enum llama_vocab_type {
@@ -2835,6 +2837,92 @@ def llama_state_seq_load_file(
 ) -> int: ...
 
 
+# for backwards-compat
+# define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+LLAMA_STATE_SEQ_FLAGS_SWA_ONLY = 1
+
+# work only with partial states, such as SWA KV cache or recurrent cache
+# (e.g. Mamba)
+# define LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY 1
+LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY = 1
+
+# keeps the tensor data on device buffers
+# (i.e. not accessible in host memory, but faster save/load)
+# define LLAMA_STATE_SEQ_FLAGS_ON_DEVICE 2
+LLAMA_STATE_SEQ_FLAGS_ON_DEVICE = 2
+
+
+# LLAMA_API size_t llama_state_seq_get_size_ext(
+#         struct llama_context * ctx,
+#                 llama_seq_id   seq_id,
+#        llama_state_seq_flags   flags);
+@ctypes_function(
+    "llama_state_seq_get_size_ext",
+    [llama_context_p_ctypes, llama_seq_id, llama_state_seq_flags],
+    ctypes.c_size_t,
+)
+def llama_state_seq_get_size_ext(
+    ctx: llama_context_p,
+    seq_id: llama_seq_id,
+    flags: llama_state_seq_flags,
+    /,
+) -> int: ...
+
+
+# LLAMA_API size_t llama_state_seq_get_data_ext(
+#         struct llama_context * ctx,
+#                      uint8_t * dst,
+#                       size_t   size,
+#                 llama_seq_id   seq_id,
+#        llama_state_seq_flags   flags);
+@ctypes_function(
+    "llama_state_seq_get_data_ext",
+    [
+        llama_context_p_ctypes,
+        ctypes.POINTER(ctypes.c_uint8),
+        ctypes.c_size_t,
+        llama_seq_id,
+        llama_state_seq_flags,
+    ],
+    ctypes.c_size_t,
+)
+def llama_state_seq_get_data_ext(
+    ctx: llama_context_p,
+    dst: CtypesArray[ctypes.c_uint8],
+    size: Union[ctypes.c_size_t, int],
+    seq_id: llama_seq_id,
+    flags: llama_state_seq_flags,
+    /,
+) -> int: ...
+
+
+# LLAMA_API size_t llama_state_seq_set_data_ext(
+#         struct llama_context * ctx,
+#                const uint8_t * src,
+#                       size_t   size,
+#                 llama_seq_id   dest_seq_id,
+#        llama_state_seq_flags   flags);
+@ctypes_function(
+    "llama_state_seq_set_data_ext",
+    [
+        llama_context_p_ctypes,
+        ctypes.POINTER(ctypes.c_uint8),
+        ctypes.c_size_t,
+        llama_seq_id,
+        llama_state_seq_flags,
+    ],
+    ctypes.c_size_t,
+)
+def llama_state_seq_set_data_ext(
+    ctx: llama_context_p,
+    src: CtypesArray[ctypes.c_uint8],
+    size: Union[ctypes.c_size_t, int],
+    dest_seq_id: llama_seq_id,
+    flags: llama_state_seq_flags,
+    /,
+) -> int: ...
+
+
 # //
 # // Decoding
 # //