'update'

__init__.py

@@ -60,9 +60,12 @@ class MZ_LLamaCPPModelConfig_ManualSelect:
     @classmethod
     def INPUT_TYPES(s):
         gguf_files = Utils.get_gguf_files()
+
+        chat_format = mz_llama_cpp.get_llama_cpp_chat_handlers()
         return {
             "required": {
                 "llama_cpp_model": (gguf_files,),
+                "chat_format": (["auto"] + chat_format, {"default": "auto"}),
             },
             "optional": {
             },
@@ -82,9 +85,13 @@ class MZ_LLamaCPPModelConfig_ManualSelect:
             llama_cpp_model = os.path.join(
                 Utils.get_gguf_models_path(), llama_cpp_model)
 
+        chat_format = kwargs.get("chat_format", "auto")
+        if chat_format == "auto":
+            chat_format = None
         return ({
             "type": "ManualSelect",
             "model_path": llama_cpp_model,
+            "chat_format": chat_format,
         },)
 
 
@@ -100,9 +107,11 @@ class MZ_LLamaCPPModelConfig_DownloaderSelect:
         model_names = [
             model["model"] for model in optional_models
         ]
+        chat_format = mz_llama_cpp.get_llama_cpp_chat_handlers()
         return {
             "required": {
                 "model_name": (model_names,),
+                "chat_format": (["auto"] + chat_format, {"default": "auto"}),
             },
             "optional": {
             },
@@ -118,9 +127,13 @@ class MZ_LLamaCPPModelConfig_DownloaderSelect:
         kwargs = kwargs.copy()
 
         model_name = kwargs.get("model_name", "")
+        chat_format = kwargs.get("chat_format", "auto")
+        if chat_format == "auto":
+            chat_format = None
         return ({
             "type": "DownloaderSelect",
             "model_name": model_name,
+            "chat_format": chat_format,
         },)
 
 
@@ -211,6 +224,9 @@ class MZ_LLamaCPPOptions:
         opt = {}
         for key in kwargs:
             opt[key] = kwargs[key]
+
+        # if opt.get("chat_format", None) == "auto":
+        #     opt["chat_format"] = None
         return (opt,)
 
 
@@ -370,6 +386,7 @@ class MZ_ImageInterrogatorCLIPTextEncode:
         return {
             "required": {
                 "resolution": ("INT", {"default": 512, "min": 128, "max": 0xffffffffffffffff}),
+                "post_processing": ([False, True], {"default": True}),
                 "keep_device": ([False, True], {"default": False}),
                 "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
             },
@@ -406,10 +423,12 @@ class MZ_ImageInterrogatorModelConfig_ManualSelect:
     @classmethod
     def INPUT_TYPES(s):
         gguf_files = Utils.get_gguf_files()
+        chat_format = mz_llama_cpp.get_llama_cpp_chat_handlers()
         return {
             "required": {
                 "llama_cpp_model": (gguf_files,),
                 "mmproj_model": (["auto"] + gguf_files,),
+                "chat_format": (["auto"] + chat_format, {"default": "auto"}),
             },
             "optional": {
             },
@@ -434,10 +453,14 @@ class MZ_ImageInterrogatorModelConfig_ManualSelect:
             mmproj_model = os.path.join(
                 Utils.get_gguf_models_path(), mmproj_model)
 
+        chat_format = kwargs.get("chat_format", "auto")
+        if chat_format == "auto":
+            chat_format = None
         return ({
             "type": "ManualSelect",
             "model_path": llama_cpp_model,
             "mmproj_model": mmproj_model,
+            "chat_format": chat_format,
         },)
 
 
@@ -459,10 +482,12 @@ class MZ_ImageInterrogatorModelConfig_DownloaderSelect:
             model["model"] for model in optional_models
         ]
 
+        chat_format = mz_llama_cpp.get_llama_cpp_chat_handlers()
         return {
             "required": {
                 "model_name": (model_names,),
-                "mmproj_model_name": (mmproj_model_names,),
+                "mmproj_model_name": (["auto"] + mmproj_model_names,),
+                "chat_format": (["auto"] + chat_format, {"default": "auto"}),
             },
             "optional": {
             },
@@ -478,10 +503,14 @@ class MZ_ImageInterrogatorModelConfig_DownloaderSelect:
         kwargs = kwargs.copy()
         model_name = kwargs.get("model_name")
         mmproj_model_name = kwargs.get("mmproj_model_name", "auto")
+        chat_format = kwargs.get("chat_format", "auto")
+        if chat_format == "auto":
+            chat_format = None
         return ({
             "type": "DownloaderSelect",
             "model_name": model_name,
             "mmproj_model_name": mmproj_model_name,
+            "chat_format": chat_format,
         },)
 
 

half_json.py

@@ -0,0 +1,349 @@
+# coding=utf8
+import json
+from typing import Any, List, NamedTuple, Optional, Tuple
+import json.decoder
+from json.decoder import JSONDecodeError as PyJSONDecodeError, JSONDecoder, py_scanstring
+from json.scanner import py_make_scanner
+from typing import Any, Dict, NamedTuple, Optional, Tuple, Union
+
+
+class FixResult(NamedTuple):
+    success: bool
+    line: str
+    origin: bool
+
+
+class JSONFixer:
+    def __init__(self, max_try: int = 20, max_stack: int = 3, *, js_style: bool = False) -> None:
+        self._max_try = max_try
+        self._max_stack = max_stack
+        self._js_style = js_style
+        self.last_fix: Optional[bool] = None
+        self.fix_stack: List[str] = []
+
+    def fix(self, line: str, *, strict: bool = True) -> FixResult:
+        try:
+            json.loads(line, strict=strict)
+            return FixResult(success=True, line=line, origin=True)
+        except Exception:
+            pass
+
+        ok, new_line = self.fixwithtry(line, strict=strict)
+        return FixResult(success=ok, line=new_line, origin=False)
+
+    def fixwithtry(self, line: str, *, strict: bool = True) -> Tuple[bool, str]:
+        if self._max_try <= 0:
+            return False, line
+
+        self.fix_stack = []
+        self.last_fix = None
+
+        ok = False
+        for _ in range(self._max_try):
+            ok, new_line = self.patch_line(line, strict=strict)
+            if ok:
+                return ok, new_line
+
+            self.last_fix = line != new_line
+            if self.last_fix:
+                self.fix_stack.insert(0, new_line)
+                self.fix_stack = self.fix_stack[: self._max_stack]
+
+            line = new_line
+        return ok, line
+
+    def patch_line(self, line: str, *, strict: bool = True) -> Tuple[bool, str]:
+        result = decode_line(line, strict=strict)
+        if result.success:
+            return True, line
+
+        if isinstance(result.exception, ValueError):
+            return self.patch_value_error(line, result.err_info)
+
+        if isinstance(result.exception, StopIteration):
+            return self.patch_stop_iteration(line)
+
+        if result.exception is None:
+            return self.patch_half_parse(line, result.err_info)
+
+        return False, line
+
+    def patch_value_error(self, line: str, err_info: Any) -> Tuple[bool, str]:
+        if err_info["error"] is None:
+            return False, line
+
+        error = err_info["error"]
+        pos = err_info["pos"]
+        nextchar = line[pos: pos + 1]
+        lastchar = line[pos - 1: pos]
+        nextline = line[pos:]
+        lastline = line[:pos]
+
+        if error == errors.StringUnterminatedString:
+            return False, insert_line(line, '"', len(line))
+        if error == errors.ObjectExceptKey:
+            if nextchar == "":
+                return False, insert_line(line, "}", pos)
+            if nextchar == ":":
+                return False, insert_line(line, '""', pos)
+            if lastchar in "{," and nextchar == ",":
+                return False, remove_line(line, pos, pos + 1)
+            if lastchar == "," and nextchar == "}":
+                return False, remove_line(line, pos - 1, pos)
+            if nextchar in "[{":
+                return False, insert_line(line, '"":', pos)
+            if self._js_style:
+                # find 'abc'
+                if nextchar == "'":
+                    nextline = remove_line(nextline, 0, 1)
+                    idx = nextline.find(":")
+                    if idx != -1 and idx != 0 and nextline[idx - 1] == "'":
+                        nextline = remove_line(nextline, idx - 1, idx)
+
+                    return False, lastline + nextline
+                # abc:1 --> "aabc":1
+                idx = nextline.find(":")
+                if idx != -1:
+                    line = lastline + insert_line(nextline, '"', idx)
+                    return False, insert_line(line, '"', pos)
+            # TODO process more case "
+            return False, insert_line(line, '"', pos)
+        if error == errors.ObjectExceptColon:
+            return False, insert_line(line, ":", pos)
+        if error == errors.ObjectExceptObject:
+            if nextchar == "":
+                if lastchar == "{":
+                    return False, insert_line(line, "}", pos)
+                return False, insert_line(line, "null}", pos)
+            if nextchar == "}":
+                return False, insert_line(line, "null", pos)
+            # TODO guess more
+            return False, insert_line(line, '"', pos)
+        if error == errors.ObjectExceptComma:
+            if nextchar == "":
+                return False, insert_line(line, "}", pos)
+            return False, insert_line(line, ",", pos)
+        if error == errors.ArrayExceptObject:
+            if nextchar == "," and lastchar == "[":
+                return False, remove_line(line, pos, pos + 1)
+            if nextchar == ",":
+                return False, insert_line(line, "null", pos)
+            if nextchar == "]":
+                return False, remove_line(line, pos - 1, pos)
+            if nextchar == "":
+                if lastchar == "[":
+                    return False, insert_line(line, "]", pos)
+                return False, insert_line(line, "null]", pos)
+            # TODO guess more?
+            return False, insert_line(line, "{", pos)
+        if error == errors.ArrayExceptComma:
+            if len(line) == pos:
+                return False, insert_line(line, "]", pos)
+            return False, insert_line(line, ",", pos)
+        # TODO unknonwn
+        return False, line
+
+    def patch_stop_iteration(self, line: str) -> Tuple[bool, str]:
+        # TODO clean
+        # TODO fix
+        # 1. }]
+        # 2. ]}
+        # 3. constans
+        # 4. -
+        # 先 patch 完 {[]}
+        # TODO: process number
+        if line.startswith("-."):
+            new_line = "-0." + line[2:]
+            return False, new_line
+        # patch
+        left = patch_lastest_left_object_and_array(line)
+        if left == "":
+            if not self.last_fix:
+                left = patch_guess_left(line)
+
+        new_line = left + line
+        return False, new_line
+
+    def patch_half_parse(self, line: str, err_info: Any) -> Tuple[bool, str]:
+        obj, end = err_info
+        nextline = line[end:].strip()
+        nextchar = nextline[:1]
+        left = patch_lastest_left_object_and_array(nextline)
+        # ??
+        if left == "":
+            if nextchar == ",":
+                left = "["
+            elif nextchar == ":" and isinstance(obj, str):
+                left = "{"
+            else:
+                if not self.last_fix:
+                    left = patch_guess_left(nextline)
+
+        new_line = left + line[:end] + nextline
+        return False, new_line
+
+
+# TODO better name
+def patch_lastest_left_object_and_array(line: str) -> str:
+    # '}]{[' --> '[{}]{['
+    pairs = {"}": "{", "]": "["}
+    breaks = "{["
+    left = ""
+    for char in line:
+        if char in breaks:
+            break
+        if char in pairs:
+            left = pairs[char] + left
+
+    return left
+
+
+# TODO better name
+# TODO 改成 lastest
+# TODO {}}]]]] --> { not [
+def patch_guess_left(line: str) -> str:
+    miss_object = line.count("}") - line.count("{")
+    miss_array = line.count("]") - line.count("[")
+    if miss_object == miss_array == 0:
+        if line[-1:] == '"' and line.count('"') == 1:
+            return '"'
+    elif miss_object >= miss_array:
+        return "{"
+    else:
+        return "["
+    return ""
+
+
+def insert_line(line: str, value: str, pos: int) -> str:
+    return line[:pos] + value + line[pos:]
+
+
+def remove_line(line: str, start: int, end: int) -> str:
+    return line[:start] + line[end:]
+
+
+class JSONDecodeError:
+    def __init__(self, parser, message):
+        self.message = message
+        self.parser = parser
+
+    def __eq__(self, err):
+        return err.parser == self.parser and self.message in err.message
+
+
+class errors:
+    StringInvalidUXXXXEscape = JSONDecodeError(
+        "py_scanstring", "Invalid \\uXXXX escape")
+    # 2 different case
+    StringUnterminatedString = JSONDecodeError(
+        "py_scanstring", "Unterminated string starting at")
+    StringInvalidControlCharacter = JSONDecodeError(
+        "py_scanstring", "Invalid control character")
+    StringInvalidEscape = JSONDecodeError("py_scanstring", "Invalid \\escape")
+    ObjectExceptColon = JSONDecodeError(
+        "JSONObject", "Expecting ':' delimiter")
+    ObjectExceptObject = JSONDecodeError("JSONObject", "Expecting value")
+    # 2 different case
+    ObjectExceptKey = JSONDecodeError(
+        "JSONObject", "Expecting property name enclosed in double quotes")
+    ObjectExceptComma = JSONDecodeError(
+        "JSONObject", "Expecting ',' delimiter")
+    ArrayExceptObject = JSONDecodeError("JSONArray", "Expecting value")
+    ArrayExceptComma = JSONDecodeError("JSONArray", "Expecting ',' delimiter")
+
+    @classmethod
+    def get_decode_error(cls, parser, message):
+        err = JSONDecodeError(parser, message)
+        for _, value in cls.__dict__.items():
+            if isinstance(value, JSONDecodeError):
+                if err == value:
+                    return value
+        return None
+
+    """
+    01 先不看,不研究
+    02 badcase: " --> "" success
+    03 控制符 pass
+    04 unicode \\u 的 pass
+    05 同上
+    06 object 后面没有跟随 " , badcase: {abc":1} --> {"abc":1}
+    07 object key 后面没有 : , badcase: {"abc"1} --> {"abc":1}
+    08 object 开始检测 Value 收到 StopIteration
+    08.1 要么后面没有了
+    08.2 要么后面不是 "/{/[/n[ull]/t[rue]/f[alse]/number/NaN/Infinity/-Infinity 开头的东西
+    -- 08.1 后面补上 null}
+    -- 08.2 无脑补一个 "
+    09 object 解析完一个 pair 后,下一个不是}, 期待一个 ','
+       badcase {"k":1"s":2}
+    10 在 09 的基础上解析完{"k":1, 发现下一个不是 ", 这个后面再优化(暂时和 06 一致)
+       badcase {"k":1,x":2}
+    11 array 开始检测 Value 收到 StopIteration
+    11.1 要么后面没有了,补上]
+    11.2 同 08.2,无脑补一个{ 看看
+    12 array 解析完前一个 object, 需要一个 ,
+        这里 nextchar 既不是 ] 也不是, 代表这个 nextchar 的 end 也已经+1 了，所以减 2
+    """
+
+
+def errmsg_inv(e: ValueError) -> Dict[str, Any]:
+    assert isinstance(e, PyJSONDecodeError)
+    parser = e.__dict__.get("parser", "")
+    errmsg = e.msg
+    localerr = errors.get_decode_error(parser, errmsg)
+    return {
+        "parsers": e.__dict__.get("parsers", []),
+        "error": localerr,
+        "lineno": e.lineno,
+        "colno": e.colno,
+        "pos": e.pos,
+    }
+
+
+def record_parser_name(parser: Any) -> Any:
+    def new_parser(*args: Any, **kwargs: Any) -> Any:
+        try:
+            return parser(*args, **kwargs)
+        except Exception as e:
+            if "parser" not in e.__dict__:
+                e.__dict__["parser"] = parser.__name__
+            if "parsers" not in e.__dict__:
+                e.__dict__["parsers"] = []
+            e.__dict__["parsers"].append(parser.__name__)
+            raise e
+
+    return new_parser
+
+
+def make_decoder(*, strict: bool = True) -> JSONDecoder:
+    json.decoder.scanstring = record_parser_name(py_scanstring)
+
+    decoder = JSONDecoder(strict=strict)
+    decoder.parse_object = record_parser_name(decoder.parse_object)
+    decoder.parse_array = record_parser_name(decoder.parse_array)
+    decoder.parse_string = record_parser_name(py_scanstring)
+    decoder.parse_object = record_parser_name(decoder.parse_object)
+
+    decoder.scan_once = py_make_scanner(decoder)
+    return decoder
+
+
+decoder = make_decoder()
+decoder_unstrict = make_decoder(strict=False)
+
+
+class DecodeResult(NamedTuple):
+    success: bool
+    exception: Optional[Exception]
+    err_info: Optional[Union[Dict[str, Any], Tuple[Any, Any]]]
+
+
+def decode_line(line: str, *, strict: bool = True) -> DecodeResult:
+    try:
+        obj, end = (decoder if strict else decoder_unstrict).scan_once(line, 0)
+        ok = end == len(line)
+        return DecodeResult(success=ok, exception=None, err_info=(obj, end))
+    except StopIteration as e:
+        return DecodeResult(success=False, exception=e, err_info=None)
+    except ValueError as e:
+        err_info = errmsg_inv(e)
+        return DecodeResult(success=False, exception=e, err_info=err_info)

mz_gen_translate.py

@@ -35,6 +35,7 @@ ZH_Replace_Map = {
     "ModelConfigManualSelect": "手动模型选择器",
     "ModelConfigDownloaderSelect": "预设模型下载器",
     "captioner_config": "打标器配置",
+    "post_processing": "后处理",
 }
 
 

mz_llama_core_nodes.py

@@ -56,16 +56,23 @@ def get_style_presets():
 def llama_cpp_node_encode(args_dict):
     importlib.reload(mz_prompts)
     importlib.reload(mz_llama_cpp)
+    # importlib.reload(mz_prompt_utils)
 
     model_config = args_dict.get("llama_cpp_model", {})
     mz_prompt_utils.Utils.print_log(f"model_config: {model_config}")
 
+    chat_format = model_config.get("chat_format", None)
+
     select_model_type = model_config.get("type", "ManualSelect")
     if select_model_type == "ManualSelect":
         model_file = model_config.get("model_path", "auto")
         if model_file == "auto":
             model_file = mz_prompt_utils.Utils.get_auto_model_fullpath(
                 "Meta-Llama-3-8B-Instruct.Q4_K_M")
+
+            if "llama-3" in mz_llama_cpp.get_llama_cpp_chat_handlers():
+                chat_format = "llama-3"
+
     elif select_model_type == "DownloaderSelect":
         model_name = model_config.get("model_name", "")
         model_file = mz_prompt_utils.Utils.get_auto_model_fullpath(
@@ -81,6 +88,7 @@ def llama_cpp_node_encode(args_dict):
     keep_device = args_dict.get("keep_device", False)
     seed = args_dict.get("seed", -1)
     options["seed"] = seed
+    options["chat_format"] = chat_format
 
     customize_instruct = args_dict.get("customize_instruct", None)
     mz_prompt_utils.Utils.print_log(
@@ -136,8 +144,19 @@ def llama_cpp_node_encode(args_dict):
                 schema=schema,
                 options=options,
             )
+            try:
+                response_json = json.loads(response_text)
+            except Exception as e:
+                from . import half_json
+                print("json.loads failed, try fix response_text: ", response_text)
+                json_fixer = half_json.JSONFixer()
+                fix_resp = json_fixer.fix(response_text)
+                if fix_resp.success:
+                    print("fix success, use fixed response_text: ", fix_resp.line)
+                    response_json = json.loads(fix_resp.line)
+                else:
+                    raise e
 
-            response_json = json.loads(response_text)
             mz_prompt_utils.Utils.print_log(
                 f"response_json: {json.dumps(response_json, indent=2)}")
 
@@ -322,6 +341,8 @@ def image_interrogator_node_encode(args_dict):
             "图片批量反推任务已完成 ; Image batch reverse push task completed")
 
     model_config = args_dict.get("image_interrogator_model", {})
+
+    chat_format = model_config.get("chat_format", None)
     llama_cpp_model = model_config.get("llama_cpp_model", "auto")
     mmproj_model = model_config.get("mmproj_model", "auto")
 
@@ -401,13 +422,20 @@ def image_interrogator_node_encode(args_dict):
     seed = args_dict.get("seed", -1)
     options = args_dict.get("llama_cpp_options", {})
     options["seed"] = seed
+    options["chat_format"] = chat_format
 
     image = mz_prompt_utils.Utils.resize_max(image, resolution, resolution)
 
     customize_instruct = args_dict.get("customize_instruct", None)
     if customize_instruct is None:
-        system_prompt = mz_prompts.M_ImageCaptioner_System
-        question = mz_prompts.M_ImageCaptioner_Prompt
+        # system_prompt = mz_prompts.GPT4VImageCaptioner_System
+        # question = mz_prompts.GPT4VImageCaptioner_Prompt
+
+        # system_prompt = mz_prompts.M_ImageCaptioner2_System
+        # question = mz_prompts.M_ImageCaptioner2_Prompt
+
+        system_prompt = "You are an assistant who perfectly describes images."
+        question = "Describe this image in detail please."
     else:
         system_prompt = customize_instruct.get("system", "")
         question = customize_instruct.get("instruct", "")
@@ -421,13 +449,116 @@ def image_interrogator_node_encode(args_dict):
         system=system_prompt,
         question=question,
     )
-    if response is not None:
-        response = response.strip()
-    if customize_instruct is None:
-        # 去除开头的In the image,
-        start_text = "In the image, "
-        if response.startswith(start_text):
-            response = response[len(start_text):]
+    response = response.strip()
+    if response is not None and response != "":
+
+        if args_dict.get("post_processing", False):
+
+            # 双引号换成空格
+            response = response.replace("\"", " ")
+            # 中括号换成空格
+            response = response.replace("[", " ")
+            response = response.replace("]", " ")
+
+            # 括号换成空格
+            response = response.replace("(", " ")
+            response = response.replace(")", " ")
+
+            # 去除多余空格
+            while response.find("  ") != -1:
+                response = response.replace("  ", " ")
+
+            # 从第一个为英文字母的地方开始截取
+            for i in range(len(response)):
+                if response[i].isalpha():
+                    response = response[i:]
+                    break
+
+            response = response.strip()
+            schema = get_schema_obj(
+                keys_type={
+                    "short_describes": get_schema_base_type("string"),
+                    "subject_tags": get_schema_array("string"),
+                    "action_tags": get_schema_array("string"),
+                    "light_tags": get_schema_array("string"),
+                    "scene_tags": get_schema_array("string"),
+                    "mood_tags": get_schema_array("string"),
+                    "style_tags": get_schema_array("string"),
+                    "object_tags": get_schema_array("string"),
+                },
+                required=[
+                    "short_describes",
+                    "subject_tags",
+                    "action_tags",
+                    "lights_tags",
+                    "scenes_tags",
+                    "moods_tags",
+                    "styles_tags",
+                    "objects_tags",
+                ]
+            )
+            response_json_str = mz_llama_cpp.llama_cpp_simple_interrogator_to_json(
+                model_file=llama_cpp_model,
+                system=mz_prompts.ImageCaptionerPostProcessing_System,
+                question=f"Content: {response}",
+                schema=schema,
+                options=options,
+            )
+
+            try:
+                response_json = json.loads(response_json_str)
+            except Exception as e:
+                from . import half_json
+                print("json.loads failed, try fix response_json_str: ",
+                      response_json_str)
+                json_fixer = half_json.JSONFixer()
+                fix_resp = json_fixer.fix(response_json_str)
+                if fix_resp.success:
+                    print("fix success, use fixed response_json_str: ",
+                          fix_resp.line)
+                    response_json = json.loads(fix_resp.line)
+                else:
+                    raise e
+
+            responses = []
+
+            def pure_words(text: str) -> bool:
+                number_of_spaces = text.count(" ")
+                if number_of_spaces > 2:
+                    return False
+                for c in text:
+                    if not c.isalpha() and c != "-" and c != "_" and c != " ":
+                        return False
+
+                return True
+
+            for key, value in response_json.items():
+                if type(value) == list:
+
+                    # 去除开头.和空格
+                    value = [v.strip().lstrip(".") for v in value]
+                    # 去除空字符串
+                    value = [v for v in value if v != ""]
+
+                    # 去除带有空格和标点符号的字符串
+                    value = [
+                        v for v in value if pure_words(v)]
+
+                    # 空格换成下划线
+                    value = [v.replace(" ", "_") for v in value]
+
+                    # 首字母小写
+                    value = [v.lower() for v in value]
+
+                    if len(value) > 0:
+                        responses.append(f"{', '.join(value)}")
+
+            description = response_json.get("short_describes", "")
+            if description != "":
+                responses.append(f"{description}")
+
+            # 对response进行去重
+            response = ", ".join(responses)
 
     if keep_device is False:
         mz_llama_cpp.freed_gpu_memory(model_file=llama_cpp_model)

mz_llama_cpp.py

@@ -1,3 +1,4 @@
+import copy
 import importlib
 import json
 import os
@@ -109,8 +110,19 @@ def check_llama_cpp_requirements():
             print("llama_cpp installed successfully. (llama_cpp安装成功)")
 
 
+def get_llama_cpp_chat_handlers():
+    check_llama_cpp_requirements()
+    from llama_cpp import llama_chat_format
+    chat_handlers = llama_chat_format.LlamaChatCompletionHandlerRegistry()._chat_handlers
+    chat_handlers = list(chat_handlers.keys())
+
+    return chat_handlers
+
+
 def LlamaCppOptions():
+    # chat_handlers = ["auto"] + get_llama_cpp_chat_handlers()
     return {
+        # "chat_format": chat_handlers,
         "n_ctx": 2048,
         "n_batch": 2048,
         "n_threads": 0,
@@ -170,6 +182,9 @@ def llama_cpp_messages(model_file, mmproj_file=None, messages=[], options={}):
 
     is_opts_changed = False
 
+    mz_prompt_utils.Utils.print_log(
+        f"llama_cpp_messages chat_format: {options.get('chat_format', None)}")
+
     if model_and_opt is not None:
         for opt in init_opts:
             if model_and_opt.get("options").get(opt) != options.get(opt):
@@ -259,6 +274,25 @@ def llama_cpp_messages(model_file, mmproj_file=None, messages=[], options={}):
         mirostat_mode = 1
     elif options.get("mirostat_mode", "none") == "mirostat_v2":
         mirostat_mode = 2
+
+    try:
+        debuf_messages = copy.deepcopy(messages)
+        for dindex in range(len(debuf_messages)):
+            if debuf_messages[dindex].get("role") == "user":
+                debuf_messages_content = debuf_messages[dindex].get(
+                    "content", [])
+                if type(debuf_messages_content) != list:
+                    continue
+                for ccindex in range(len(debuf_messages_content)):
+                    if debuf_messages_content[ccindex].get("type") == "image_url":
+                        debuf_messages[dindex]["content"][ccindex]["image_url"] = debuf_messages[
+                            dindex]["content"][ccindex]["image_url"] = None
+
+        mz_prompt_utils.Utils.print_log(
+            f"LLAMA_CPP messages: {json.dumps(debuf_messages, indent=4, ensure_ascii=False)}")
+    except Exception as e:
+        mz_prompt_utils.Utils.print_log(
+            f"LLAMA_CPP messages: {messages}")
     output = model.create_chat_completion(
         messages=messages,
         response_format=response_format,
@@ -276,6 +310,8 @@ def llama_cpp_messages(model_file, mmproj_file=None, messages=[], options={}):
         mirostat_mode=mirostat_mode,
         mirostat_tau=options.get("mirostat_tau", 5.0),
         mirostat_eta=options.get("mirostat_eta", 0.1),
+        tools=options.get("tools", None),
+        tool_choice=options.get("tool_choice", None),
     )
     mz_prompt_utils.Utils.print_log(f"LLAMA_CPP: \n{output}")
     choices = output.get("choices", [])
@@ -330,7 +366,9 @@ def llama_cpp_simple_interrogator_to_json(model_file, use_system=True, system=No
     }
 
     options["response_format"] = response_format
-    options["chat_format"] = "chatml"
+
+    # if options.get("chat_format", None) is None:
+    #     options["chat_format"] = "llama-2"
 
     result = llama_cpp_messages(model_file, None, messages, options=options)
     result = result.replace("\n", " ")
@@ -385,6 +423,9 @@ def llava_cpp_messages(model_file, mmproj_file, messages, options={}):
     options = options.copy()
     options["logits_all"] = True
     options["n_ctx"] = max(4096, options.get("n_ctx", 4096))
+
+    # if options.get("chat_format", None) is None:
+    #     options["chat_format"] = "llama-2"
     return llama_cpp_messages(model_file, mmproj_file, messages, options)
 
 

mz_prompt_utils.py

@@ -172,41 +172,81 @@ class Utils:
         os.makedirs(models_path, exist_ok=True)
         return models_path
 
-    def translate_text(text, from_code, to_code):
+    def get_translate_object(from_code, to_code):
         try:
-            import argostranslate
-            from argostranslate import translate
-        except ImportError:
-            subprocess.run([
-                sys.executable, "-m",
-                "pip", "install", "argostranslate"], check=True)
-            import argostranslate
-            from argostranslate import translate
+            is_disabel_argostranslate = Utils.cache_get(
+                "is_disabel_argostranslate")
 
-        try:
-            translation = translate.get_translation_from_codes(
-                from_code, to_code)
-            if translation is None:
-                raise Exception("Translation not found")
+            if is_disabel_argostranslate is not None:
+                return None
 
+            try:
+                import argostranslate
+                from argostranslate import translate, package
+            except ImportError:
+                subprocess.run([
+                    sys.executable, "-m",
+                    "pip", "install", "argostranslate"], check=True)
+
+                try:
+                    import argostranslate
+                    from argostranslate import translate, package
+                except ImportError:
+                    Utils.cache_set("is_disabel_argostranslate", True)
+                    print(
+                        "argostranslate not found and install failed , will disable it")
+                    return None
+
+            packages = package.get_installed_packages()
+            installed_packages = {}
+            for p in packages:
+                installed_packages[f"{p.from_code}_{p.to_code}"] = p
+
+            argosmodel_dir = os.path.join(
+                Utils.get_models_path(), "argosmodel")
+            if not os.path.exists(argosmodel_dir):
+                os.makedirs(argosmodel_dir)
+
+            model_name = None
+            if from_code == "zh" and to_code == "en":
+                model_name = "zh_en"
+            elif from_code == "en" and to_code == "zh":
+                model_name = "en_zh"
+            else:
+                return None
+
+            if Utils.cache_get(f"argostranslate_{model_name}") is not None:
+                return Utils.cache_get(f"argostranslate_{model_name}")
+
+            if installed_packages.get(model_name, None) is None:
+                if not os.path.exists(os.path.join(argosmodel_dir, f"translate-{model_name}-1_9.argosmodel")):
+                    argosmodel_file = Utils.download_file(
+                        url=f"https://www.modelscope.cn/api/v1/models/wailovet/MinusZoneAIModels/repo?Revision=master&FilePath=argosmodel%2Ftranslate-{model_name}-1_9.argosmodel",
+                        filepath=os.path.join(
+                            argosmodel_dir, f"translate-{model_name}-1_9.argosmodel"),
+                    )
+                else:
+                    argosmodel_file = os.path.join(
+                        argosmodel_dir, f"translate-{model_name}-1_9.argosmodel")
+                package.install_from_path(argosmodel_file)
+
+            translate_object = translate.get_translation_from_codes(
+                from_code=from_code, to_code=to_code)
+
+            Utils.cache_set(f"argostranslate_{model_name}", translate_object)
+
+            return translate_object
         except Exception as e:
-            print(e)
-            argostranslate.package.update_package_index()
-            available_packages = argostranslate.package.get_available_packages()
-            package_to_install = next(
-                filter(
-                    lambda x: (x.from_code == from_code and x.to_code ==
-                               to_code), available_packages,
-                )
-            )
-            download_path = package_to_install.download()
-            print("package_to_install.download():", download_path)
-            argostranslate.package.install_from_path(download_path)
+            Utils.cache_set("is_disabel_argostranslate", True)
+            print(
+                "argostranslate not found and install failed , will disable it")
+            print(f"get_translate_object error: {e}")
+            return None
 
-            translation = translate.get_translation_from_codes(
-                from_code, to_code)
-            if translation is None:
-                return text
+    def translate_text(text, from_code, to_code):
+        translation = Utils.get_translate_object(from_code, to_code)
+        if translation is None:
+            return text
 
         # Translate
         translatedText = translation.translate(
@@ -677,6 +717,9 @@ class Utils:
     def to_debug_prompt(p):
         if p is None:
             return ""
+        zh = Utils.en2zh(p)
+        if p == zh:
+            return p
         zh = Utils.split_en_to_zh(p)
         p = p.strip()
         return f"""
@@ -766,7 +809,7 @@ class Utils:
         cache_value = Utils.get_cache_by_local(cache_key)
         if cache_value is not None:
             return cache_value
-        
+
         sha256 = Utils.file_hash(file_path, hashlib.sha256)
         Utils.set_cache_by_local(cache_key, sha256)
         return sha256

mz_prompts.py

@@ -47,3 +47,23 @@ Describe this image in detail please.
 The language of reply is English only!!!
 Starts with "In the image," 
 """
+
+
+M_ImageCaptioner2_System = """
+You are an assistant who perfectly describes images. 
+"""
+
+M_ImageCaptioner2_Prompt = """
+Describe this image in detail please.
+The language of reply is English only!!!
+Starts with "In the image," 
+"""
+
+
+ImageCaptionerPostProcessing_System = """
+I want you to write me a detailed list of tips for Content.
+Write a very short description of the scene and put it in the 'short_describes' field
+Write complete [moods, styles, lights, elements, objects] of the word array and put it in the '$_tags' field
+Don't include anything that isn't in Content.
+The language of reply is English only!!!
+"""