Merge https://github.com/abetlen/llama-cpp-python

llama_cpp/llama.py

@@ -194,32 +194,32 @@ class Llama:
         self.model_params.use_mmap = use_mmap if lora_path is None else False
         self.model_params.use_mlock = use_mlock
 
+        # kv_overrides is the original python dict
         self.kv_overrides = kv_overrides
         if kv_overrides is not None:
-            n_overrides = len(kv_overrides)
-            self._kv_overrides_array = llama_cpp.llama_model_kv_override * (n_overrides + 1)
-            self._kv_overrides_array_keys = []
+            # _kv_overrides_array is a ctypes.Array of llama_model_kv_override Structs
+            kvo_array_len = len(kv_overrides) + 1  # for sentinel element
+            self._kv_overrides_array = (
+                llama_cpp.llama_model_kv_override * kvo_array_len
+            )()
 
-            for k, v in kv_overrides.items():
-                key_buf = ctypes.create_string_buffer(k.encode("utf-8"))
-                self._kv_overrides_array_keys.append(key_buf)
-                self._kv_overrides_array[i].key = key_buf
-                if isinstance(v, int):
+            for i, (k, v) in enumerate(kv_overrides.items()):
+                self._kv_overrides_array[i].key = k.encode("utf-8")
+                if isinstance(v, bool):
+                    self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL
+                    self._kv_overrides_array[i].value.bool_value = v
+                elif isinstance(v, int):
                     self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_INT
                     self._kv_overrides_array[i].value.int_value = v
                 elif isinstance(v, float):
                     self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_FLOAT
                     self._kv_overrides_array[i].value.float_value = v
-                elif isinstance(v, bool):
-                    self._kv_overrides_array[i].tag = llama_cpp.LLAMA_KV_OVERRIDE_BOOL
-                    self._kv_overrides_array[i].value.bool_value = v
                 else:
                     raise ValueError(f"Unknown value type for {k}: {v}")
 
-            self._kv_overrides_array_sentinel_key = b'\0'
-
-            # null array sentinel
-            self._kv_overrides_array[n_overrides].key = self._kv_overrides_array_sentinel_key
+            self._kv_overrides_array[
+                -1
+            ].key = b"\0"  # ensure sentinel element is zeroed
             self.model_params.kv_overrides = self._kv_overrides_array
 
         self.n_batch = min(n_ctx, n_batch)  # ???
@@ -329,7 +329,9 @@ class Llama:
             (n_ctx, self._n_vocab), dtype=np.single
         )
 
-        self._mirostat_mu = ctypes.c_float(2.0 * 5.0) # TODO: Move this to sampling context
+        self._mirostat_mu = ctypes.c_float(
+            2.0 * 5.0
+        )  # TODO: Move this to sampling context
 
         try:
             self.metadata = self._model.metadata()
@@ -537,7 +539,7 @@ class Llama:
                 candidates=self._candidates,
                 tau=mirostat_tau,
                 eta=mirostat_eta,
-                mu=ctypes.pointer(self._mirostat_mu)
+                mu=ctypes.pointer(self._mirostat_mu),
             )
         else:
             self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1)

llama_cpp/llama_chat_format.py

@@ -379,7 +379,8 @@ def hf_autotokenizer_to_chat_completion_handler(
 
 
 def hf_tokenizer_config_to_chat_formatter(
-    tokenizer_config: Dict[str, Any]
+    tokenizer_config: Dict[str, Any],
+    add_generation_prompt: bool = True,
 ) -> ChatFormatter:
     assert isinstance(tokenizer_config, dict)
 
@@ -401,31 +402,34 @@ def hf_tokenizer_config_to_chat_formatter(
         lstrip_blocks=True,
     ).from_string(chat_template)
 
-    def format_autotokenizer(
+    def format_tokenizer_config(
         messages: List[llama_types.ChatCompletionRequestMessage],
         **kwargs: Any,
     ) -> ChatFormatterResponse:
         # TODO: veryify this is correct
         # Add a blank assistant message to the end of the messages to prompt the model to generate a response
-        prompt = env.render(
-            messages=[
+        if add_generation_prompt:
+            messages = [
                 *messages,
                 llama_types.ChatCompletionRequestAssistantMessage(
                     role="assistant", content=""
                 ),
-            ],
+            ]
+        prompt = env.render(
+            messages=messages,
             bos_token=bos_token,
             eos_token=eos_token,
         )
-        return ChatFormatterResponse(prompt=prompt, stop=eos_token)
+        return ChatFormatterResponse(prompt=prompt, stop=[eos_token, bos_token])
 
-    return format_autotokenizer
+    return format_tokenizer_config
 
 
 def hf_tokenizer_config_to_chat_completion_handler(
     tokenizer_config: Dict[str, Any],
+    add_generation_prompt: bool = True,
 ) -> LlamaChatCompletionHandler:
-    chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config)
+    chat_formatter = hf_tokenizer_config_to_chat_formatter(tokenizer_config, add_generation_prompt=add_generation_prompt)
     return chat_formatter_to_chat_completion_handler(chat_formatter)
 
 

llama_cpp/llama_cpp.py

@@ -188,6 +188,7 @@ LLAMA_TOKEN_TYPE_BYTE = 6
 #     LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
 #     LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
 #     LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
+#     LLAMA_FTYPE_MOSTLY_Q3_K_XS       = 22, // except 1d tensors
 
 #     LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
 # };
@@ -211,6 +212,7 @@ LLAMA_FTYPE_MOSTLY_Q6_K = 18
 LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19
 LLAMA_FTYPE_MOSTLY_IQ2_XS = 20
 LLAMA_FTYPE_MOSTLY_Q2_K_S = 21
+LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22
 LLAMA_FTYPE_GUESSED = 1024
 
 # enum llama_rope_scaling_type {

vendor/llama.cpp

@@ -1 +1 @@
-Subproject commit 504dc37be8446fb09b1ede70300250ad41be32a2
+Subproject commit 26d607608d794efa56df3bdb6043a2f94c1d632c