llama.cpp/examples/openai/prompting.py

412 lines
20 KiB
Python
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

from enum import Enum
import jinja2
import json
from pathlib import Path
import random
import re
import sys
from typing import Optional, Tuple, Callable
from typeguard import typechecked
from examples.json_schema_to_grammar import SchemaConverter
from examples.openai.api import Tool, Message, FunctionCall, ToolCall
from examples.openai.gguf_kvs import GGUFKeyValues, Keys
from examples.openai.ts_converter import SchemaToTypeScriptConverter
@typechecked
def raise_exception(msg: str):
raise Exception(msg)
@typechecked
class ChatFormat:
def __init__(self, template: str, eos_token: str, bos_token: str):
env = jinja2.Environment(loader=jinja2.BaseLoader(), trim_blocks=True, lstrip_blocks=True)
self.template = env.from_string(template)
self.eos_token = eos_token
self.bos_token = bos_token
self.strict_user_assistant_alternation = "{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception" in template
if "<|recipient|>' + tool_call['function']['name']" in template:
self.tool_style = ToolsPromptStyle.TYPESCRIPT_FUNCTIONARY_V2
else:
self.tool_style = ToolsPromptStyle.TOOLS_LONG
def __str__(self):
return f"ChatFormat(template={self.template}, eos_token={self.eos_token}, bos_token={self.bos_token})"
def add_system_prompt(self, messages: list[Message], system_prompt: Message) -> list[Message]:
assert system_prompt.role == "system"
# TODO: add to last system message, or create a new one just before the last user message
system_message = next(((i, m) for i, m in enumerate(messages) if m.role == "system"), None)
if system_message is not None:
(i, m) = system_message
return messages[:i] + [Message(role="system", content=m.content + '\n' + system_prompt.content)] + messages[i+1:]
else:
return [system_prompt] + messages
@staticmethod
def from_gguf(metadata: GGUFKeyValues):
tokens = metadata[Keys.Tokenizer.LIST]
return ChatFormat(
template = metadata[Keys.Tokenizer.CHAT_TEMPLATE],
bos_token = tokens[metadata[Keys.Tokenizer.BOS_ID]],
eos_token = tokens[metadata[Keys.Tokenizer.EOS_ID]])
def render(self, messages: list[Message], add_generation_prompt: bool, omit_bos: bool = False):
if self.strict_user_assistant_alternation and any(m.role not in ('user', 'assistant') for m in messages):
new_messages=[]
i = 0
n = len(messages)
while i < n:
if messages[i].role == 'system':
assert messages[i+1].role == 'user'
new_messages.append(Message(
role="user",
content=f'[SYS]{messages[i].content}[/SYS]\n{messages[i+1].content}'))
i += 2
else:
new_messages.append(messages[i])
i += 1
# print(f'new_messages={json.dumps(new_messages, indent=2)}')
messages = new_messages
# print(f'messages={messages}')
return self.template.render(
messages=messages,
eos_token=self.eos_token,
bos_token='' if omit_bos else self.bos_token,
raise_exception=raise_exception,
add_generation_prompt=add_generation_prompt,
)
# While the API will be usable with a generic tools usage like OpenAI,
# (see https://cookbook.openai.com/examples/how_to_call_functions_with_chat_models),
# each model may need specific prompting (and/or constrained output,
# especially for models not fine-tuned for tool usage / function calling).
class ToolsPromptStyle(Enum):
# Short prompt w/ <tools>schemas</tools>
TOOLS_SHORT = 1
# Longer prompt w/ <tools>schemas</tools>
TOOLS_LONG = 2
# Large prompt for https://huggingface.co/NousResearch/Hermes-2-Pro-Mistral-7B
# Requires:
# - git clone https://github.com/NousResearch/Hermes-Function-Calling examples/openai/hermes_function_calling
# - Set large context length as their prompts are super long
TOOLS_HERMES_2_PRO = 3
# Short prompt w/ TypeScript definitions for https://github.com/MeetKai/functionary
# https://github.com/MeetKai/functionary/blob/main/functionary/prompt_template/prompt_template_v2.py
# Note: see this prior attempt to support Functionary: https://github.com/ggerganov/llama.cpp/pull/5695
TYPESCRIPT_FUNCTIONARY_V2 = 4
@typechecked
def make_tools_prompt(chat_format: ChatFormat, tools: list[Tool], indent=2) -> Message:
if chat_format.tool_style == ToolsPromptStyle.TOOLS_SHORT:
return Message(
role="system",
content='\n'.join([
'Here are the tools available:',
'<tools>',
*(json.dumps(tool.model_dump(), indent=indent) for tool in tools),
'</tools>',
])
)
elif chat_format.tool_style == ToolsPromptStyle.TOOLS_LONG:
return Message(
role="system",
content='\n'.join([
'''You are a function calling AI model. You are provided with function signatures within <tools></tools> XML tags.''',
'''You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. Here are the available tools:''',
'''<tools>''',
*(json.dumps(tool.model_dump(), indent=indent) for tool in tools),
'''</tools>''',
'',
'''Use the following json schema for each tool call you will make: {"properties": {"arguments": {"title": "Arguments", "type": "object"}, "name": {"title": "Name", "type": "string"}}, "required": ["arguments", "name"], "title": "FunctionCall", "type": "object"}''',
'',
# '''For each function call return a json object with function name and arguments within <tool_call></tool_call> XML tags as follows:''',
'''To call each function, give its name and arguments within <tool_call></tool_call> XML tags as follows:''',
'''<tool_call>''',
'''{"arguments": <args-dict>, "name": <function-name>}''',
'''</tool_call>''',
'''This is not hypothetical, you're not asked what you would do. If you need a tool called, just call it.''',
])
)
elif chat_format.tool_style == ToolsPromptStyle.TYPESCRIPT_FUNCTIONARY_V2:
ts_converter = SchemaToTypeScriptConverter()
return Message(
role="system",
content='\n'.join([
'// Supported function definitions that should be called when necessary.'
'namespace functions {',
*[
'// ' + tool.function.description.replace('\n', '\n// ') + '\n' + ''
'type ' + tool.function.name + ' = (_: ' + ts_converter.visit(tool.function.parameters) + ") => any;\n"
for tool in tools
],
'} // namespace functions',
])
)
elif chat_format.tool_style == ToolsPromptStyle.TOOLS_HERMES_2_PRO:
# Hackily import https://github.com/NousResearch/Hermes-Function-Calling
path = str(Path(__file__).parent / "hermes_function_calling")
if path not in sys.path: sys.path.insert(0, path)
try:
from examples.openai.hermes_function_calling.prompter import PromptManager
except ImportError:
raise ImportError(f"Please `git clone https://github.com/NousResearch/Hermes-Function-Calling {path}`")
prompt = PromptManager().generate_prompt(user_prompt=[], tools=[json.dumps(tool) for tool in tools])
assert len(prompt) == 1 and prompt[0]["role"] == "system"
return Message(**prompt[0])
else:
raise ValueError(f"Unsupported tool call style: {chat_format.tool_style}")
@typechecked
def _outputs_tool_call_tags(style: ToolsPromptStyle) -> bool:
return style in (
ToolsPromptStyle.TOOLS_SHORT,
ToolsPromptStyle.TOOLS_LONG,
ToolsPromptStyle.TOOLS_HERMES_2_PRO,
)
_tool_call_re = re.compile(
'<tool_call>(.*?)</tool_call>', re.DOTALL)
_recipient_content_re = re.compile(r'(?:(?:<\|(?:stop|from)\|>)+ *assistant\n<\|recipient\|>|^) *([^ <|>\n]+) *\n<\|content\|>(.*?)(?:$|<\|stop\|>\s*$|(?=(?:<\|(?:stop|from)\|>)+ *assistant\n))', re.DOTALL)
def gen_callid():
return f'call_{random.randint(0, 1000000)}'
@typechecked
def make_grammar(chat_format: ChatFormat, tools: list[Tool], response_schema: Optional[dict], indent=2) -> Tuple[Optional[str], Callable[[str], Optional[list[Message]]]]:
converter = SchemaConverter(prop_order={}, allow_fetch=False, dotall=False, raw_pattern=False)
response_rule = converter.visit(response_schema, "response") if response_schema else None
delimiter = '<%$[SAMPLE]$%>'
user_msg = Message(role="user", content="Hey")
empty_prompt = chat_format.render([user_msg], add_generation_prompt=True).strip()
planted_prompt = chat_format.render([user_msg, Message(role="assistant", content=delimiter)], add_generation_prompt=False).strip()
assert planted_prompt.startswith(empty_prompt), f"Planted prompt does not start with empty prompt: {planted_prompt} vs {empty_prompt}"
[prefix, suffix] = planted_prompt[len(empty_prompt):].split(delimiter)
def strip_suffix(s: str) -> str:
if s.endswith(suffix):
return s[:-len(suffix)]
else:
sys.stderr.write(f"Expected suffix ({suffix}) not found: {s}\n")
return s
if tools:
if _outputs_tool_call_tags(chat_format.tool_style):
escapes_underscores = chat_format.tool_style != ToolsPromptStyle.TOOLS_HERMES_2_PRO
tool_rules = [
converter.visit(
dict(
type="object",
properties=dict(
name=dict(type="string", pattern='^' + tool.function.name.replace('_', f'\\?_') + '$') if escapes_underscores \
else dict(const=tool.function.name),
arguments=tool.function.parameters,
),
required=['name', 'arguments']
),
f'{tool.function.name}-tool-call'
)
for tool in tools
]
def format_literal(s: str) -> str:
if escapes_underscores:
return ' "\\\\"? "_" '.join((converter._format_literal(part) for part in s.split('_')))
else:
return converter._format_literal(s)
tool_call_rule = converter._add_rule(
'tool_call',
format_literal("<tool_call>") + " (" +
' | '.join(tool_rules) +
") " + format_literal("</tool_call>"))
# Ideally we'd want a negative lookahead of /<tool\\?_call>/, but it's just too hard to express in GBNF for now.
# So we just over-constrain the content rule to not contain literals dangerously getting close to <tool_call>
content_rule = converter._add_rule('content', '[^<] | "<" [^t<]? | "<t" [^o<]?')
# content_rule = converter._add_rule('content', converter.not_literal('<tool_call>'))
converter._add_rule(
'root',
f'{content_rule}* ({tool_call_rule}+ {content_rule}*)?')
# # Constrain the output to be a non-tool-call message (constrained to a JSON schema or not)
# # OR a tool-call message respecting the schema of any of the tools
# converter._add_rule(
# "root",
# converter._format_literal(prefix) + " (" +
# (response_rule or converter.not_literal("<tool_call>")) + " | " +
# converter._format_literal("<tool_call>") + " (" +
# ' | '.join(tool_rules) +
# ") " + converter._format_literal("</tool_call>") +
# ")") # + converter._format_literal(suffix))
@typechecked
def parse(s: str) -> Optional[Message]:
s = strip_suffix(s)
if r'<tool\_call>' in s:
# Some weird escaping of underscores is happening w/ Mixtral 8x7B Instruct
s = s.replace(r'\_', '_')
parts = _tool_call_re.split(s)
if len(parts) == 1:
return Message(role="assistant", content=s)
else:
content = []
tool_calls = []
for i, part in enumerate(parts):
if i % 2 == 0:
content.append(part)
else:
try:
fc = json.loads(part)
except json.JSONDecodeError:
raise ValueError(f'Failed to parse tool call as JSON: {part}\nFull string: {s}')
tool_calls.append(
ToolCall(
id=gen_callid(),
function=FunctionCall(**fc)))
content = '(...)'.join(content).strip()
return Message(role="assistant", content=content if content else None, tool_calls=tool_calls)
# if '<tool_call>'.startswith(ls) or ls.startswith('<tool_call>'):
# if ls.startswith('<tool_call>') and ls.endswith('</tool_call>' + suffix):
# tool_call = ls[len('<tool_call>'):-len('</tool_call>' + suffix)]
# return Message(role="assistant", content=None, tool_calls=[json.loads(tool_call)])
# return None
# else:
# return Message(role="assistant", content=s)
return (converter.format_grammar(), parse)
elif chat_format.tool_style == ToolsPromptStyle.TYPESCRIPT_FUNCTIONARY_V2:
# Only allowing a single tool call at a time for now.
# Note that if there were more, they'd be separated by a '<|from|>assistant' literal
tool_rules = [
converter._add_rule(
tool.function.name + '-call',
converter._format_literal(tool.function.name) + ' ' + converter._format_literal('\n<|content|>\n') + ' ' +
converter.visit(tool.function.parameters, tool.function.name + '-args') + ' ' +
converter._format_literal('\n'))
# converter.visit(
# dict(
# type="object",
# properties=dict(
# name=dict(const=tool.function.name),
# arguments=tool.function.parameters,
# ),
# required=['name', 'arguments']
# ),
# f'{tool.function.name}-tool-call'
# )
for i, tool in enumerate(tools)
]
not_from_rule = converter._add_rule('not_from', converter.not_literal("<|from|>"))
content_without_start_rule = converter._add_rule(
'content_without_start',
converter._format_literal("all\n<|content|>") + ' ' + not_from_rule + '*')
start_rule = converter._add_rule('start', converter._format_literal('<|from|>assistant\n<|recipient|>'))
content_rule = converter._add_rule('content', start_rule + ' ' + content_without_start_rule)
tool_call_without_start_rule = converter._add_rule(
'tool_call_without_start',
' | '.join(tool_rules))
# + ' ' +
# converter.not_literal("all", dotall=False) + ' ' + converter._format_literal('\n<|content|>\n') + ' ' + not_from_rule + '*')
tool_call_rule = converter._add_rule('tool_call', f'{start_rule} {tool_call_without_start_rule}')
# converter._add_rule('root', f'({content_without_start_rule} ({content_rule})* ({tool_call_rule}+ {content_rule}*)? | {tool_call_without_start_rule} (* {tool_call_rule}{content_rule}*')
converter._add_rule(
'root',
f'{content_without_start_rule} {content_rule}* ({tool_call_rule}+ {content_rule}*)? | '
f'{tool_call_without_start_rule} {tool_call_rule}* {content_rule}*')
# converter._add_rule(
# "root",
# converter._format_literal(prefix) + " (" +
# (response_rule or converter.not_literal("<|recipient|>")) + " | " +
# (' | '.join(
# converter._format_literal(f"<|recipient|>{tool.function.name}\n<|content|>") + " " +
# converter.visit(tool.function.parameters, tool.function.name + '-args')
# for tool in tools
# )) +
# ") " +
# ")") # + converter._format_literal(suffix))
@typechecked
def parse(s: str) -> Optional[Message]:
s = strip_suffix(s)
parts = _recipient_content_re.split(s)
if len(parts) == 1:
return Message(role="assistant", content=s)
else:
text_content = []
tool_calls: list[ToolCall] = []
for i in range((len(parts) - 1) // 3):
assert parts[i * 3].strip() == '', f'Unexpected content before tool call: {parts[i * 3]}'
recipient = parts[i * 3 + 1].strip()
content = parts[i * 3 + 2]
if recipient == 'all':
text_content.append(content)
else:
try:
arguments = json.loads(content)
except json.JSONDecodeError:
raise ValueError(f'Failed to parse tool call content as JSON: {content}')
tool_calls.append(
ToolCall(
id=gen_callid(),
function=FunctionCall(name=recipient, arguments=arguments)))
assert parts[-1].strip() in ('', '<|stop|>'), f'Unexpected content after tool calls: {parts[-1]}\nFull string: {s}'
content = '\n'.join(text_content).strip()
return Message(role="assistant", content=content if content else None, tool_calls=tool_calls if tool_calls else None)
return (converter.format_grammar(), parse)
else:
raise ValueError(f"Unsupported tool call style: {chat_format.tool_style}")
elif response_schema:
converter._add_rule("root", response_rule + ' ' + converter._format_literal(suffix))
@typechecked
def parse(s: str) -> Optional[Message]:
s = strip_suffix(s)
return Message(role="assistant", content=s)
return (converter.format_grammar(), parse)
else:
converter._add_rule("root", converter._format_literal(prefix) + ' ' + converter._format_literal(suffix))
@typechecked
def parse(s: str) -> Optional[Message]:
s = strip_suffix(s)
return Message(role="assistant", content=s)
return (None, parse)