chat_template.py 13 KB
Newer Older
1
from dataclasses import dataclass, field
Lianmin Zheng's avatar
Lianmin Zheng committed
2
from enum import Enum, auto
3
from typing import Callable, Dict, List, Optional, Tuple
Lianmin Zheng's avatar
Lianmin Zheng committed
4
5
6
7
8
9
10
11
12
13
14


class ChatTemplateStyle(Enum):
    PLAIN = auto()
    LLAMA2 = auto()


@dataclass
class ChatTemplate:
    name: str
    default_system_prompt: str
15
    role_prefix_and_suffix: Dict[str, Tuple[str, str]]
16
    stop_str: List[str] = ()
Lianmin Zheng's avatar
Lianmin Zheng committed
17
18
19
    image_token: str = "<image>"
    style: ChatTemplateStyle = ChatTemplateStyle.PLAIN

Liangsheng Yin's avatar
Liangsheng Yin committed
20
21
22
    def get_prefix_and_suffix(
        self, role: str, hist_messages: List[Dict]
    ) -> Tuple[str, str]:
23
        prefix, suffix = self.role_prefix_and_suffix.get(role, ("", ""))
Liangsheng Yin's avatar
Liangsheng Yin committed
24

25
26
27
        if self.style == ChatTemplateStyle.LLAMA2:
            if role == "system" and not hist_messages:
                user_prefix, _ = self.role_prefix_and_suffix.get("user", ("", ""))
Liangsheng Yin's avatar
Liangsheng Yin committed
28
29
30
                system_prefix, system_suffix = self.role_prefix_and_suffix.get(
                    "system", ("", "")
                )
31
                return (user_prefix + system_prefix, system_suffix)
Liangsheng Yin's avatar
Liangsheng Yin committed
32
33
34
35
36
            elif (
                role == "user"
                and len(hist_messages) == 1
                and hist_messages[0]["content"] is not None
            ):
37
38
39
40
41
                return ("", suffix)

        return prefix, suffix

    def get_prompt(self, messages: List[Dict]) -> str:
Lianmin Zheng's avatar
Lianmin Zheng committed
42
        prompt = ""
43
44
        for i, message in enumerate(messages):
            role, content = message["role"], message["content"]
Lianmin Zheng's avatar
Lianmin Zheng committed
45
46
47
48
49
50
            if role == "system" and content is None:
                content = self.default_system_prompt
                if content is None:
                    continue

            prefix, suffix = self.get_prefix_and_suffix(role, messages[:i])
51
            prompt += f"{prefix}{content}{suffix}"
Lianmin Zheng's avatar
Lianmin Zheng committed
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
        return prompt


chat_template_registry: Dict[str, ChatTemplate] = {}
matching_function_registry: List[Callable] = []


def register_chat_template(template):
    chat_template_registry[template.name] = template


def register_chat_template_matching_function(func):
    matching_function_registry.append(func)


def get_chat_template(name):
    return chat_template_registry[name]


def get_chat_template_by_model_path(model_path):
    for matching_func in matching_function_registry:
        template = matching_func(model_path)
        if template is not None:
            return template
    return get_chat_template("default")


register_chat_template(
    ChatTemplate(
        name="default",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": ("SYSTEM:", "\n"),
            "user": ("USER:", "\n"),
            "assistant": ("ASSISTANT:", "\n"),
87
        },
Lianmin Zheng's avatar
Lianmin Zheng committed
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
    )
)


register_chat_template(
    ChatTemplate(
        name="claude",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": ("", ""),
            "user": ("\n\nHuman: ", ""),
            "assistant": ("\n\nAssistant:", ""),
        },
    )
)


register_chat_template(
    ChatTemplate(
        name="chatml",
        default_system_prompt=None,
        role_prefix_and_suffix={
Enrique Shockwave's avatar
Enrique Shockwave committed
110
111
112
            "system": ("<|im_start|>system\n", "<|im_end|>\n"),
            "user": ("<|im_start|>user\n", "<|im_end|>\n"),
            "assistant": ("<|im_start|>assistant\n", "<|im_end|>\n"),
Lianmin Zheng's avatar
Lianmin Zheng committed
113
114
        },
        style=ChatTemplateStyle.PLAIN,
115
        stop_str=("<|im_end|>",),
Lianmin Zheng's avatar
Lianmin Zheng committed
116
117
118
    )
)

119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
# There is default system prompt for qwen
# reference: https://modelscope.cn/models/qwen/Qwen2-72B-Instruct/file/view/master?fileName=tokenizer_config.json&status=1
# The chat template is: "{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
register_chat_template(
    ChatTemplate(
        name="qwen",
        default_system_prompt="You are a helpful assistant.",
        role_prefix_and_suffix={
            "system": ("<|im_start|>system\n", "<|im_end|>\n"),
            "user": ("<|im_start|>user\n", "<|im_end|>\n"),
            "assistant": ("<|im_start|>assistant\n", "<|im_end|>\n"),
        },
        style=ChatTemplateStyle.PLAIN,
        stop_str=("<|im_end|>",),
    )
)

Lianmin Zheng's avatar
Lianmin Zheng committed
136

137
138
139
140
141
register_chat_template(
    ChatTemplate(
        name="chatml-llava",
        default_system_prompt="Answer the questions.",
        role_prefix_and_suffix={
Enrique Shockwave's avatar
Enrique Shockwave committed
142
143
144
            "system": ("<|im_start|>system\n", "<|im_end|>\n"),
            "user": ("<|im_start|>user\n", "<|im_end|>\n"),
            "assistant": ("<|im_start|>assistant\n", "<|im_end|>\n"),
145
146
147
148
149
150
151
        },
        style=ChatTemplateStyle.PLAIN,
        stop_str=("<|im_end|>",),
        image_token=" <image>\n",
    )
)

152
# Reference: https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md#prompt-template
Lianmin Zheng's avatar
Lianmin Zheng committed
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
register_chat_template(
    ChatTemplate(
        name="vicuna_v1.1",
        default_system_prompt=(
            "A chat between a curious user and an artificial intelligence assistant. "
            "The assistant gives helpful, detailed, and polite answers to the user's questions."
        ),
        role_prefix_and_suffix={
            "system": ("", " "),
            "user": ("USER:", " "),
            "assistant": ("ASSISTANT:", "</s>"),
        },
        image_token=" <image>\n",
    )
)

169
170
171
172
173
174
175
176
177
178
179
# Reference: https://modelscope.cn/models/01ai/Yi-1.5-34B-Chat/file/view/master?fileName=tokenizer_config.json&status=1
register_chat_template(
    ChatTemplate(
        name="yi-1.5",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": ("", ""),
            "user": ("<|im_start|>user\n", "<|im_end|>\n<|im_start|>assistant\n"),
            "assistant": ("", "<|im_end|>\n"),
        },
        style=ChatTemplateStyle.PLAIN,
180
        stop_str=("<|im_end|>",),
181
182
    )
)
Lianmin Zheng's avatar
Lianmin Zheng committed
183
184
185
186
187
188
189
190
191
192
193
194
195
196

register_chat_template(
    ChatTemplate(
        name="llama-2-chat",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": ("<<SYS>>\n", "\n<</SYS>>\n\n"),
            "user": ("[INST] ", " [/INST]"),
            "assistant": ("", " </s><s>"),
        },
        style=ChatTemplateStyle.LLAMA2,
    )
)

197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
register_chat_template(
    ChatTemplate(
        name="llama-3-instruct",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": (
                "<|start_header_id|>system<|end_header_id|>\n\n",
                "<|eot_id|>",
            ),
            "user": (
                "<|start_header_id|>user<|end_header_id|>\n\n",
                "<|eot_id|>",
            ),
            "assistant": (
                "<|start_header_id|>assistant<|end_header_id|>\n\n",
                "<|eot_id|>",
            ),
        },
        stop_str=("<|eot_id|>",),
    )
)

Christopher Chou's avatar
Christopher Chou committed
219
220
221
# Reference: https://github.com/01-ai/Yi/tree/main/VL#major-difference-with-llava
register_chat_template(
    ChatTemplate(
222
        name="yi-vl",
Christopher Chou's avatar
Christopher Chou committed
223
224
225
226
227
228
229
230
231
232
233
234
235
        default_system_prompt=(
            "This is a chat between an inquisitive human and an AI assistant. Assume the role of the AI assistant. Read all the images carefully, and respond to the human's questions with informative, helpful, detailed and polite answers."
            "这是一个好奇的人类和一个人工智能助手之间的对话。假设你扮演这个AI助手的角色。仔细阅读所有的图像,并对人类的问题做出信息丰富、有帮助、详细的和礼貌的回答。"
        ),
        role_prefix_and_suffix={
            "system": ("", "\n\n"),
            "user": ("### Human:", "\n"),
            "assistant": ("### Assistant:", "\n"),
        },
        image_token=" <image_placeholder>\n",
    )
)

Liangsheng Yin's avatar
Liangsheng Yin committed
236
237
238
239
240
241
242
243
244
245
246
247
248
register_chat_template(
    ChatTemplate(
        name="gemma-it",
        default_system_prompt=None,
        role_prefix_and_suffix={
            "system": ("", ""),
            "user": ("<start_of_turn>user\n", "<end_of_turn>\n"),
            "assistant": ("<start_of_turn>model\n", "<end_of_turn>\n"),
        },
        style=ChatTemplateStyle.PLAIN,
    )
)

249
250
251
252
253
254
255
256
257
258
259
260
261
register_chat_template(
    ChatTemplate(
        name="dbrx-instruct",
        default_system_prompt="You are DBRX, created by Databricks. You were last updated in December 2023. You answer questions based on information available up to that point.\nYOU PROVIDE SHORT RESPONSES TO SHORT QUESTIONS OR STATEMENTS, but provide thorough responses to more complex and open-ended questions.\nYou assist with various tasks, from writing to coding (using markdown for code blocks — remember to use ``` with code, JSON, and tables).\n(You do not have real-time data access or code execution capabilities. You avoid stereotyping and provide balanced perspectives on controversial topics. You do not provide song lyrics, poems, or news articles and do not divulge details of your training data.)\nThis is your system prompt, guiding your responses. Do not reference it, just respond to the user. If you find yourself talking about this message, stop. You should be responding appropriately and usually that means not mentioning this.\nYOU DO NOT MENTION ANY OF THIS INFORMATION ABOUT YOURSELF UNLESS THE INFORMATION IS DIRECTLY PERTINENT TO THE USER'S QUERY.",
        role_prefix_and_suffix={
            "system": ("<|im_start|>system\n", "<|im_end|>"),
            "user": ("\n<|im_start|>user\n", "<|im_end|>"),
            "assistant": ("\n<|im_start|>assistant\n", "<|im_end|>"),
        },
        stop_str=("<|im_end|>",),
    )
)

262
263
264
265
266
register_chat_template(
    ChatTemplate(
        name="c4ai-command-r",
        default_system_prompt=None,
        role_prefix_and_suffix={
267
268
269
270
            "system": (
                "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>",
                "<|END_OF_TURN_TOKEN|>",
            ),
271
            "user": ("<|START_OF_TURN_TOKEN|><|USER_TOKEN|>", "<|END_OF_TURN_TOKEN|>"),
272
273
274
275
            "assistant": (
                "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
                "<|END_OF_TURN_TOKEN|>",
            ),
276
277
278
279
280
        },
        style=ChatTemplateStyle.PLAIN,
    )
)

281
282
283
284
285
286

@register_chat_template_matching_function
def match_dbrx(model_path: str):
    if "dbrx" in model_path.lower() and "instruct" in model_path.lower():
        return get_chat_template("dbrx-instruct")

Lianmin Zheng's avatar
Lianmin Zheng committed
287
288
289
290
291

@register_chat_template_matching_function
def match_vicuna(model_path: str):
    if "vicuna" in model_path.lower():
        return get_chat_template("vicuna_v1.1")
292
    if "llava-v1.5" in model_path.lower():
Lianmin Zheng's avatar
Lianmin Zheng committed
293
        return get_chat_template("vicuna_v1.1")
Yuanhan Zhang's avatar
Yuanhan Zhang committed
294
295
    if "llava-next-video-7b" in model_path.lower():
        return get_chat_template("vicuna_v1.1")
Lianmin Zheng's avatar
Lianmin Zheng committed
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310


@register_chat_template_matching_function
def match_llama2_chat(model_path: str):
    model_path = model_path.lower()
    if "llama-2" in model_path and "chat" in model_path:
        return get_chat_template("llama-2-chat")
    if (
        "mistral" in model_path or "mixtral" in model_path
    ) and "instruct" in model_path:
        return get_chat_template("llama-2-chat")
    if "codellama" in model_path and "instruct" in model_path:
        return get_chat_template("llama-2-chat")


311
312
313
314
315
316
317
@register_chat_template_matching_function
def match_llama3_instruct(model_path: str):
    model_path = model_path.lower()
    if "llama-3" in model_path and "instruct" in model_path:
        return get_chat_template("llama-3-instruct")


Lianmin Zheng's avatar
Lianmin Zheng committed
318
319
@register_chat_template_matching_function
def match_chat_ml(model_path: str):
Yuanhan Zhang's avatar
Yuanhan Zhang committed
320
    # import pdb;pdb.set_trace()
Lianmin Zheng's avatar
Lianmin Zheng committed
321
322
323
    model_path = model_path.lower()
    if "tinyllama" in model_path:
        return get_chat_template("chatml")
324
325
326
    # Now the suffix for qwen2 chat model is "instruct"
    if "qwen" in model_path and ("chat" in model_path or "instruct" in model_path):
        return get_chat_template("qwen")
Yuanhan Zhang's avatar
Yuanhan Zhang committed
327
328
329
330
331
    if (
        "llava-v1.6-34b" in model_path
        or "llava-v1.6-yi-34b" in model_path
        or "llava-next-video-34b" in model_path
    ):
332
        return get_chat_template("chatml-llava")
Lianmin Zheng's avatar
Lianmin Zheng committed
333

334

Christopher Chou's avatar
Christopher Chou committed
335
336
337
@register_chat_template_matching_function
def match_chat_yi(model_path: str):
    model_path = model_path.lower()
338
339
340
341
    if "yi-vl" in model_path and "llava" not in model_path:
        return get_chat_template("yi-vl")
    elif "yi-1.5" in model_path and "chat" in model_path:
        return get_chat_template("yi-1.5")
Christopher Chou's avatar
Christopher Chou committed
342

Lianmin Zheng's avatar
Lianmin Zheng committed
343

Liangsheng Yin's avatar
Liangsheng Yin committed
344
345
346
347
348
349
350
@register_chat_template_matching_function
def match_gemma_it(model_path: str):
    model_path = model_path.lower()
    if "gemma" in model_path and "it" in model_path:
        return get_chat_template("gemma-it")


351
352
353
354
355
356
357
@register_chat_template_matching_function
def match_c4ai_command_r(model_path: str):
    model_path = model_path.lower()
    if "c4ai-command-r" in model_path:
        return get_chat_template("c4ai-command-r")


Lianmin Zheng's avatar
Lianmin Zheng committed
358
359
360
361
362
363
364
365
366
367
368
369
if __name__ == "__main__":
    messages = [
        {"role": "system", "content": None},  # None means default
        # {"role": "system", "content": "You are a helpful, respectful and honest assistant."},
        {"role": "user", "content": "Hello!"},
        {"role": "assistant", "content": "Hi!"},
        {"role": "user", "content": "What can you do?"},
        {"role": "assistant", "content": "I can chat with you."},
    ]

    template = get_chat_template("llama-2-chat")
    print(template.get_prompt(messages))