feat: Update llama.cpp

Author: abetlen

llama_cpp/_internals.py

@@ -805,15 +805,10 @@ def add_penalties(
         ignore_eos: bool,
     ):
         sampler = llama_cpp.llama_sampler_init_penalties(
-            n_vocab,
-            special_eos_id,
-            linefeed_id,
             penalty_last_n,
             penalty_repeat,
             penalty_freq,
             penalty_present,
-            penalize_nl,
-            ignore_eos,
         )
         self._add_sampler(sampler)
 

llama_cpp/llama_cpp.py

@@ -256,13 +256,17 @@
 # // note: these values should be synchronized with ggml_rope
 # // TODO: maybe move this enum to ggml.h (ggml_rope_type)
 # enum llama_rope_type {
-#     LLAMA_ROPE_TYPE_NONE = -1,
-#     LLAMA_ROPE_TYPE_NORM =  0,
-#     LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
+#     LLAMA_ROPE_TYPE_NONE   = -1,
+#     LLAMA_ROPE_TYPE_NORM   = 0,
+#     LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
+#     LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+#     LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
 # };
 LLAMA_ROPE_TYPE_NONE = -1
 LLAMA_ROPE_TYPE_NORM = 0
 LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX = 2
+LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE = 8
+LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION = 24
 
 
 # enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -1265,6 +1269,7 @@ def llama_rope_freq_scale_train(model: llama_model_p, /) -> float:
 # // Functions to access the model's GGUF metadata scalar values
 # // - The functions return the length of the string on success, or -1 on failure
 # // - The output string is always null-terminated and cleared on failure
+# // - When retrieving a string, an extra byte must be allocated to account for the null terminator
 # // - GGUF array values are not supported by these functions
 
 
@@ -1378,18 +1383,6 @@ def llama_model_n_params(model: llama_model_p, /) -> int:
     ...
 
 
-# // Get a llama model tensor
-# LLAMA_API struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name);
-@ctypes_function(
-    "llama_get_model_tensor", [llama_model_p_ctypes, ctypes.c_char_p], ctypes.c_void_p
-)
-def llama_get_model_tensor(
-    model: llama_model_p, name: Union[ctypes.c_char_p, bytes], /
-) -> ctypes.c_void_p:
-    """Get a llama model tensor"""
-    ...
-
-
 # // Returns true if the model contains an encoder that requires llama_encode() call
 # LLAMA_API bool llama_model_has_encoder(const struct llama_model * model);
 @ctypes_function("llama_model_has_encoder", [llama_model_p_ctypes], ctypes.c_bool)
@@ -3336,41 +3329,22 @@ def llama_sampler_init_grammar(
     ...
 
 
+# /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
 # LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
-#                          int32_t   n_vocab,         // llama_n_vocab()
-#                      llama_token   special_eos_id,  // llama_token_eos()
-#                      llama_token   linefeed_id,     // llama_token_nl()
-#                          int32_t   penalty_last_n,  // last n tokens to penalize (0 = disable penalty, -1 = context size)
-#                            float   penalty_repeat,  // 1.0 = disabled
-#                            float   penalty_freq,    // 0.0 = disabled
-#                            float   penalty_present, // 0.0 = disabled
-#                             bool   penalize_nl,     // consider newlines as a repeatable token
-#                             bool   ignore_eos);     // ignore the end-of-sequence token
+#                          int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
+#                            float   penalty_repeat,   // 1.0 = disabled
+#                            float   penalty_freq,     // 0.0 = disabled
+#                            float   penalty_present); // 0.0 = disabled
 @ctypes_function(
     "llama_sampler_init_penalties",
-    [
-        ctypes.c_int32,
-        llama_token,
-        llama_token,
-        ctypes.c_int32,
-        ctypes.c_float,
-        ctypes.c_float,
-        ctypes.c_float,
-        ctypes.c_bool,
-        ctypes.c_bool,
-    ],
+    [ctypes.c_int32, ctypes.c_float, ctypes.c_float, ctypes.c_float],
     llama_sampler_p_ctypes,
 )
 def llama_sampler_init_penalties(
-    n_vocab: int,
-    special_eos_id: int,
-    linefeed_id: int,
     penalty_last_n: int,
     penalty_repeat: float,
     penalty_freq: float,
     penalty_present: float,
-    penalize_nl: bool,
-    ignore_eos: bool,
     /,
 ) -> llama_sampler_p:
     ...