韩宇
init
1b7e88c
raw
history blame
9.99 kB
from __future__ import annotations
import re
from abc import ABC, abstractmethod
from collections import defaultdict
from collections.abc import Hashable
from pathlib import Path
from typing import Any, ClassVar, Dict, List, Optional, TypeVar, Union
from PIL import Image
from pydantic import Field, field_validator
from tenacity import retry, stop_after_attempt, stop_after_delay
from ...base import BotBase
from ...utils.env import EnvVar
from ...utils.general import LRUCache
from ...utils.registry import registry
from .prompt.base import _OUTPUT_PARSER, StrParser
from .prompt.parser import BaseOutputParser
from .prompt.prompt import PromptTemplate
from .schemas import Message
import copy
from collections.abc import Iterator
T = TypeVar("T", str, dict, list)
class BaseLLM(BotBase, ABC):
cache: bool = False
lru_cache: LRUCache = Field(default=LRUCache(EnvVar.LLM_CACHE_NUM))
@property
def workflow_instance_id(self) -> str:
if hasattr(self, "_parent"):
return self._parent.workflow_instance_id
return None
@workflow_instance_id.setter
def workflow_instance_id(self, value: str):
if hasattr(self, "_parent"):
self._parent.workflow_instance_id = value
@abstractmethod
def _call(self, records: List[Message], **kwargs) -> str:
"""Run the LLM on the given prompt and input."""
async def _acall(self, records: List[Message], **kwargs) -> str:
"""Run the LLM on the given prompt and input."""
raise NotImplementedError("Async generation not implemented for this LLM.")
def generate(self, records: List[Message], **kwargs) -> str: # TODO: use python native lru cache
"""Run the LLM on the given prompt and input."""
if self.cache:
key = self._cache_key(records)
cached_res = self.lru_cache.get(key)
if cached_res:
return cached_res
else:
gen = self._call(records, **kwargs)
self.lru_cache.put(key, gen)
return gen
else:
return self._call(records, **kwargs)
@retry(
stop=(
stop_after_delay(EnvVar.STOP_AFTER_DELAY)
| stop_after_attempt(EnvVar.STOP_AFTER_ATTEMPT)
),
reraise=True,
)
async def agenerate(self, records: List[str], **kwargs) -> str:
"""Run the LLM on the given prompt and input."""
if self.cache:
key = self._cache_key(records)
cached_res = self.lru_cache.get(key)
if cached_res:
return cached_res
else:
gen = await self._acall(records, **kwargs)
self.lru_cache.put(key, gen)
return gen
else:
return await self._acall(records, **kwargs)
def _cache_key(self, records: List[Message]) -> int:
return str([item.model_dump() for item in records])
def dict(self, *args, **kwargs):
kwargs["exclude"] = {"lru_cache"}
return super().model_dump(*args, **kwargs)
def json(self, *args, **kwargs):
kwargs["exclude"] = {"lru_cache"}
return super().model_dump_json(*args, **kwargs)
T = TypeVar("T", str, dict, list)
class BaseLLMBackend(BotBase, ABC):
"""Prompts prepare and LLM infer"""
output_parser: Optional[BaseOutputParser] = None
prompts: List[PromptTemplate] = []
llm: BaseLLM
@property
def token_usage(self):
if not hasattr(self, 'workflow_instance_id'):
raise AttributeError("workflow_instance_id not set")
return dict(self.stm(self.workflow_instance_id).get('token_usage', defaultdict(int)))
@field_validator("output_parser", mode="before")
@classmethod
def set_output_parser(cls, output_parser: Union[BaseOutputParser, Dict, None]):
if output_parser is None:
return StrParser()
elif isinstance(output_parser, BaseOutputParser):
return output_parser
elif isinstance(output_parser, dict):
return _OUTPUT_PARSER[output_parser["name"]](**output_parser)
else:
raise ValueError
@field_validator("prompts", mode="before")
@classmethod
def set_prompts(
cls, prompts: List[Union[PromptTemplate, Dict, str]]
) -> List[PromptTemplate]:
init_prompts = []
for prompt in prompts:
prompt = copy.deepcopy(prompt)
if isinstance(prompt, Path):
if prompt.suffix == ".prompt":
init_prompts.append(PromptTemplate.from_file(prompt))
elif isinstance(prompt, str):
if prompt.endswith(".prompt"):
init_prompts.append(PromptTemplate.from_file(prompt))
init_prompts.append(PromptTemplate.from_template(prompt))
elif isinstance(prompt, dict):
init_prompts.append(PromptTemplate.from_config(prompt))
elif isinstance(prompt, PromptTemplate):
init_prompts.append(prompt)
else:
raise ValueError(
"Prompt only support str, dict and PromptTemplate object"
)
return init_prompts
@field_validator("llm", mode="before")
@classmethod
def set_llm(cls, llm: Union[BaseLLM, Dict]):
if isinstance(llm, dict):
return registry.get_llm(llm["name"])(**llm)
elif isinstance(llm, BaseLLM):
return llm
else:
raise ValueError("LLM only support dict and BaseLLM object")
def prep_prompt(
self, input_list: List[Dict[str, Any]], prompts=None, **kwargs
) -> List[List[Message]]:
"""Prepare prompts from inputs."""
if prompts is None:
prompts = self.prompts
images = []
if len(kwargs_images := kwargs.get("images", [])):
images = kwargs_images
processed_prompts = []
for inputs in input_list:
records = []
for prompt in prompts:
selected_inputs = {k: inputs.get(k, "") for k in prompt.input_variables}
prompt_str = prompt.template
parts = re.split(r"(\{\{.*?\}\})", prompt_str)
formatted_parts = []
for part in parts:
if part.startswith("{{") and part.endswith("}}"):
part = part[2:-2].strip()
value = selected_inputs[part]
if isinstance(value, (Image.Image, list)):
formatted_parts.extend(
[value] if isinstance(value, Image.Image) else value
)
else:
formatted_parts.append(str(value))
else:
formatted_parts.append(str(part))
formatted_parts = (
formatted_parts[0] if len(formatted_parts) == 1 else formatted_parts
)
if prompt.role == "system":
records.append(Message.system(formatted_parts))
elif prompt.role == "user":
records.append(Message.user(formatted_parts))
if len(images):
records.append(Message.user(images))
processed_prompts.append(records)
return processed_prompts
def infer(self, input_list: List[Dict[str, Any]], **kwargs) -> List[T]:
prompts = self.prep_prompt(input_list, **kwargs)
res = []
stm_token_usage = self.stm(self.workflow_instance_id).get('token_usage', defaultdict(int))
def process_stream(self, stream_output):
for chunk in stream_output:
if chunk.usage is not None:
for key, value in chunk.usage.dict().items():
if key in ["prompt_tokens", "completion_tokens", 'total_tokens']:
if value is not None:
stm_token_usage[key] += value
self.stm(self.workflow_instance_id)['token_usage'] = stm_token_usage
yield chunk
for prompt in prompts:
output = self.llm.generate(prompt, **kwargs)
if not isinstance(output, Iterator):
for key, value in output.get("usage", {}).items():
if key in ["prompt_tokens", "completion_tokens", 'total_tokens']:
if value is not None:
stm_token_usage[key] += value
if not self.llm.stream:
for choice in output["choices"]:
if choice.get("message"):
choice["message"]["content"] = self.output_parser.parse(
choice["message"]["content"]
)
res.append(output)
else:
res.append(process_stream(self, output))
self.stm(self.workflow_instance_id)['token_usage'] = stm_token_usage
return res
async def ainfer(self, input_list: List[Dict[str, Any]], **kwargs) -> List[T]:
prompts = self.prep_prompt(input_list)
res = []
for prompt in prompts:
output = await self.llm.agenerate(prompt, **kwargs)
for key, value in output["usage"].items():
self.token_usage[key] += value
for choice in output["choices"]:
if choice.get("message"):
choice["message"]["content"] = self.output_parser.parse(
choice["message"]["content"]
)
res.append(output)
return res
def simple_infer(self, **kwargs: Any) -> T:
return self.infer([kwargs])[0]
async def simple_ainfer(self, **kwargs: Any) -> T:
return await self.ainfer([kwargs])[0]