apunkt/MetaGPT
1
0
Fork 0
mirror of https://github.com/FoundationAgents/MetaGPT.git synced 2026-05-05 13:52:38 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat: merge geekan:main

Browse source
This commit is contained in:
莘权 马 2024-03-05 10:59:34 +08:00
commit e22a28215d
102 changed files with 1766 additions and 756 deletions
Download patch file Download diff file Expand all files Collapse all files

52
metagpt/utils/cost_manager.py
View file

@ -6,12 +6,13 @@
@Desc : mashenquan, 2023/8/28. Separate the `CostManager` class to support user-level cost accounting.
"""
import re
from typing import NamedTuple
from pydantic import BaseModel
from metagpt.logs import logger
from metagpt.utils.token_counter import TOKEN_COSTS
from metagpt.utils.token_counter import FIREWORKS_GRADE_TOKEN_COSTS, TOKEN_COSTS
class Costs(NamedTuple):
@ -29,6 +30,7 @@ class CostManager(BaseModel):
total_budget: float = 0
max_budget: float = 10.0
total_cost: float = 0
token_costs: dict[str, dict[str, float]] = TOKEN_COSTS # different model's token cost
def update_cost(self, prompt_tokens, completion_tokens, model):
"""
@ -43,8 +45,13 @@ class CostManager(BaseModel):
return
self.total_prompt_tokens += prompt_tokens
self.total_completion_tokens += completion_tokens
if model not in self.token_costs:
logger.warning(f"Model {model} not found in TOKEN_COSTS.")
return
cost = (
prompt_tokens * TOKEN_COSTS[model]["prompt"] + completion_tokens * TOKEN_COSTS[model]["completion"]
prompt_tokens * self.token_costs[model]["prompt"]
+ completion_tokens * self.token_costs[model]["completion"]
) / 1000
self.total_cost += cost
logger.info(
@ -99,3 +106,44 @@ class TokenCostManager(CostManager):
self.total_prompt_tokens += prompt_tokens
self.total_completion_tokens += completion_tokens
logger.info(f"prompt_tokens: {prompt_tokens}, completion_tokens: {completion_tokens}")
class FireworksCostManager(CostManager):
def model_grade_token_costs(self, model: str) -> dict[str, float]:
def _get_model_size(model: str) -> float:
size = re.findall(".*-([0-9.]+)b", model)
size = float(size[0]) if len(size) > 0 else -1
return size
if "mixtral-8x7b" in model:
token_costs = FIREWORKS_GRADE_TOKEN_COSTS["mixtral-8x7b"]
else:
model_size = _get_model_size(model)
if 0 < model_size <= 16:
token_costs = FIREWORKS_GRADE_TOKEN_COSTS["16"]
elif 16 < model_size <= 80:
token_costs = FIREWORKS_GRADE_TOKEN_COSTS["80"]
else:
token_costs = FIREWORKS_GRADE_TOKEN_COSTS["-1"]
return token_costs
def update_cost(self, prompt_tokens: int, completion_tokens: int, model: str):
"""
Refs to `https://app.fireworks.ai/pricing` **Developer pricing**
Update the total cost, prompt tokens, and completion tokens.
Args:
prompt_tokens (int): The number of tokens used in the prompt.
completion_tokens (int): The number of tokens used in the completion.
model (str): The model used for the API call.
"""
self.total_prompt_tokens += prompt_tokens
self.total_completion_tokens += completion_tokens
token_costs = self.model_grade_token_costs(model)
cost = (prompt_tokens * token_costs["prompt"] + completion_tokens * token_costs["completion"]) / 1000000
self.total_cost += cost
logger.info(
f"Total running cost: ${self.total_cost:.4f}"
f"Current cost: ${cost:.4f}, prompt_tokens: {prompt_tokens}, completion_tokens: {completion_tokens}"
)

12
metagpt/utils/repair_llm_raw_output.py
View file

@ -119,6 +119,7 @@ def repair_json_format(output: str) -> str:
logger.info(f"repair_json_format: {'}]'}")
elif output.startswith("{") and output.endswith("]"):
output = output[:-1] + "}"
# remove comments in output json string, after json value content, maybe start with #, maybe start with //
arr = output.split("\n")
new_arr = []
@ -208,6 +209,17 @@ def repair_invalid_json(output: str, error: str) -> str:
elif (rline[col_no] in ["'", '"']) and (line.startswith('"') or line.startswith("'")) and "," not in line:
# problem, `"""` or `'''` without `,`
new_line = f",{line}"
elif col_no - 1 >= 0 and rline[col_no - 1] in ['"', "'"]:
# backslash problem like \" in the output
char = rline[col_no - 1]
nearest_char_idx = rline[col_no:].find(char)
new_line = (
rline[: col_no - 1]
+ "\\"
+ rline[col_no - 1 : col_no + nearest_char_idx]
+ "\\"
+ rline[col_no + nearest_char_idx :]
)
elif '",' not in line and "," not in line and '"' not in line:
new_line = f'{line}",'
elif not line.endswith(","):

4
metagpt/utils/text.py
View file

@ -25,7 +25,7 @@ def reduce_message_length(
"""
max_token = TOKEN_MAX.get(model_name, 2048) - count_string_tokens(system_text, model_name) - reserved
for msg in msgs:
if count_string_tokens(msg, model_name) < max_token:
if count_string_tokens(msg, model_name) < max_token or model_name not in TOKEN_MAX:
return msg
raise RuntimeError("fail to reduce message length")
@ -93,7 +93,7 @@ def split_paragraph(paragraph: str, sep: str = ".,", count: int = 2) -> list[str
continue
ret = ["".join(j) for j in _split_by_count(sentences, count)]
return ret
return _split_by_count(paragraph, count)
return list(_split_by_count(paragraph, count))
def decode_unicode_escape(text: str) -> str:

113
metagpt/utils/token_counter.py
View file

@ -32,9 +32,107 @@ TOKEN_COSTS = {
"gpt-4-vision-preview": {"prompt": 0.01, "completion": 0.03}, # TODO add extra image price calculator
"gpt-4-1106-vision-preview": {"prompt": 0.01, "completion": 0.03},
"text-embedding-ada-002": {"prompt": 0.0004, "completion": 0.0},
"glm-3-turbo": {"prompt": 0.0, "completion": 0.0007}, # 128k version, prompt + completion tokens=0.005￥/k-tokens
"glm-4": {"prompt": 0.0, "completion": 0.014}, # 128k version, prompt + completion tokens=0.1￥/k-tokens
"glm-3-turbo": {"prompt": 0.0007, "completion": 0.0007}, # 128k version, prompt + completion tokens=0.005￥/k-tokens
"glm-4": {"prompt": 0.014, "completion": 0.014}, # 128k version, prompt + completion tokens=0.1￥/k-tokens
"gemini-pro": {"prompt": 0.00025, "completion": 0.0005},
"moonshot-v1-8k": {"prompt": 0.012, "completion": 0.012}, # prompt + completion tokens=0.012￥/k-tokens
"moonshot-v1-32k": {"prompt": 0.024, "completion": 0.024},
"moonshot-v1-128k": {"prompt": 0.06, "completion": 0.06},
"open-mistral-7b": {"prompt": 0.00025, "completion": 0.00025},
"open-mixtral-8x7b": {"prompt": 0.0007, "completion": 0.0007},
"mistral-small-latest": {"prompt": 0.002, "completion": 0.006},
"mistral-medium-latest": {"prompt": 0.0027, "completion": 0.0081},
"mistral-large-latest": {"prompt": 0.008, "completion": 0.024},
}
"""
QianFan Token Price https://cloud.baidu.com/doc/WENXINWORKSHOP/s/hlrk4akp7#tokens%E5%90%8E%E4%BB%98%E8%B4%B9
Due to QianFan has multi price strategies, we unify `Tokens post-payment` as a statistical method.
"""
QIANFAN_MODEL_TOKEN_COSTS = {
"ERNIE-Bot-4": {"prompt": 0.017, "completion": 0.017},
"ERNIE-Bot-8k": {"prompt": 0.0034, "completion": 0.0067},
"ERNIE-Bot": {"prompt": 0.0017, "completion": 0.0017},
"ERNIE-Bot-turbo": {"prompt": 0.0011, "completion": 0.0011},
"EB-turbo-AppBuilder": {"prompt": 0.0011, "completion": 0.0011},
"ERNIE-Speed": {"prompt": 0.00056, "completion": 0.0011},
"BLOOMZ-7B": {"prompt": 0.00056, "completion": 0.00056},
"Llama-2-7B-Chat": {"prompt": 0.00056, "completion": 0.00056},
"Llama-2-13B-Chat": {"prompt": 0.00084, "completion": 0.00084},
"Llama-2-70B-Chat": {"prompt": 0.0049, "completion": 0.0049},
"ChatGLM2-6B-32K": {"prompt": 0.00056, "completion": 0.00056},
"AquilaChat-7B": {"prompt": 0.00056, "completion": 0.00056},
"Mixtral-8x7B-Instruct": {"prompt": 0.0049, "completion": 0.0049},
"SQLCoder-7B": {"prompt": 0.00056, "completion": 0.00056},
"CodeLlama-7B-Instruct": {"prompt": 0.00056, "completion": 0.00056},
"XuanYuan-70B-Chat-4bit": {"prompt": 0.0049, "completion": 0.0049},
"Qianfan-BLOOMZ-7B-compressed": {"prompt": 0.00056, "completion": 0.00056},
"Qianfan-Chinese-Llama-2-7B": {"prompt": 0.00056, "completion": 0.00056},
"Qianfan-Chinese-Llama-2-13B": {"prompt": 0.00084, "completion": 0.00084},
"ChatLaw": {"prompt": 0.0011, "completion": 0.0011},
"Yi-34B-Chat": {"prompt": 0.0, "completion": 0.0},
}
QIANFAN_ENDPOINT_TOKEN_COSTS = {
"completions_pro": QIANFAN_MODEL_TOKEN_COSTS["ERNIE-Bot-4"],
"ernie_bot_8k": QIANFAN_MODEL_TOKEN_COSTS["ERNIE-Bot-8k"],
"completions": QIANFAN_MODEL_TOKEN_COSTS["ERNIE-Bot"],
"eb-instant": QIANFAN_MODEL_TOKEN_COSTS["ERNIE-Bot-turbo"],
"ai_apaas": QIANFAN_MODEL_TOKEN_COSTS["EB-turbo-AppBuilder"],
"ernie_speed": QIANFAN_MODEL_TOKEN_COSTS["ERNIE-Speed"],
"bloomz_7b1": QIANFAN_MODEL_TOKEN_COSTS["BLOOMZ-7B"],
"llama_2_7b": QIANFAN_MODEL_TOKEN_COSTS["Llama-2-7B-Chat"],
"llama_2_13b": QIANFAN_MODEL_TOKEN_COSTS["Llama-2-13B-Chat"],
"llama_2_70b": QIANFAN_MODEL_TOKEN_COSTS["Llama-2-70B-Chat"],
"chatglm2_6b_32k": QIANFAN_MODEL_TOKEN_COSTS["ChatGLM2-6B-32K"],
"aquilachat_7b": QIANFAN_MODEL_TOKEN_COSTS["AquilaChat-7B"],
"mixtral_8x7b_instruct": QIANFAN_MODEL_TOKEN_COSTS["Mixtral-8x7B-Instruct"],
"sqlcoder_7b": QIANFAN_MODEL_TOKEN_COSTS["SQLCoder-7B"],
"codellama_7b_instruct": QIANFAN_MODEL_TOKEN_COSTS["CodeLlama-7B-Instruct"],
"xuanyuan_70b_chat": QIANFAN_MODEL_TOKEN_COSTS["XuanYuan-70B-Chat-4bit"],
"qianfan_bloomz_7b_compressed": QIANFAN_MODEL_TOKEN_COSTS["Qianfan-BLOOMZ-7B-compressed"],
"qianfan_chinese_llama_2_7b": QIANFAN_MODEL_TOKEN_COSTS["Qianfan-Chinese-Llama-2-7B"],
"qianfan_chinese_llama_2_13b": QIANFAN_MODEL_TOKEN_COSTS["Qianfan-Chinese-Llama-2-13B"],
"chatlaw": QIANFAN_MODEL_TOKEN_COSTS["ChatLaw"],
"yi_34b_chat": QIANFAN_MODEL_TOKEN_COSTS["Yi-34B-Chat"],
}
"""
DashScope Token price https://help.aliyun.com/zh/dashscope/developer-reference/tongyi-thousand-questions-metering-and-billing
Different model has different detail page. Attention, some model are free for a limited time.
"""
DASHSCOPE_TOKEN_COSTS = {
"qwen-turbo": {"prompt": 0.0011, "completion": 0.0011},
"qwen-plus": {"prompt": 0.0028, "completion": 0.0028},
"qwen-max": {"prompt": 0.0, "completion": 0.0},
"qwen-max-1201": {"prompt": 0.0, "completion": 0.0},
"qwen-max-longcontext": {"prompt": 0.0, "completion": 0.0},
"llama2-7b-chat-v2": {"prompt": 0.0, "completion": 0.0},
"llama2-13b-chat-v2": {"prompt": 0.0, "completion": 0.0},
"qwen-72b-chat": {"prompt": 0.0, "completion": 0.0},
"qwen-14b-chat": {"prompt": 0.0011, "completion": 0.0011},
"qwen-7b-chat": {"prompt": 0.00084, "completion": 0.00084},
"qwen-1.8b-chat": {"prompt": 0.0, "completion": 0.0},
"baichuan2-13b-chat-v1": {"prompt": 0.0011, "completion": 0.0011},
"baichuan2-7b-chat-v1": {"prompt": 0.00084, "completion": 0.00084},
"baichuan-7b-v1": {"prompt": 0.0, "completion": 0.0},
"chatglm-6b-v2": {"prompt": 0.0011, "completion": 0.0011},
"chatglm3-6b": {"prompt": 0.0, "completion": 0.0},
"ziya-llama-13b-v1": {"prompt": 0.0, "completion": 0.0}, # no price page, judge it as free
"dolly-12b-v2": {"prompt": 0.0, "completion": 0.0},
"belle-llama-13b-2m-v1": {"prompt": 0.0, "completion": 0.0},
"moss-moon-003-sft-v1": {"prompt": 0.0, "completion": 0.0},
"chatyuan-large-v2": {"prompt": 0.0, "completion": 0.0},
"billa-7b-sft-v1": {"prompt": 0.0, "completion": 0.0},
}
FIREWORKS_GRADE_TOKEN_COSTS = {
"-1": {"prompt": 0.0, "completion": 0.0}, # abnormal condition
"16": {"prompt": 0.2, "completion": 0.8}, # 16 means model size <= 16B; 0.2 means $0.2/1M tokens
"80": {"prompt": 0.7, "completion": 2.8}, # 80 means 16B < model size <= 80B
"mixtral-8x7b": {"prompt": 0.4, "completion": 1.6},
}
TOKEN_MAX = {
@ -57,8 +155,17 @@ TOKEN_MAX = {
"gpt-4-vision-preview": 128000,
"gpt-4-1106-vision-preview": 128000,
"text-embedding-ada-002": 8192,
"chatglm_turbo": 32768,
"glm-3-turbo": 128000,
"glm-4": 128000,
"gemini-pro": 32768,
"moonshot-v1-8k": 8192,
"moonshot-v1-32k": 32768,
"moonshot-v1-128k": 128000,
"open-mistral-7b": 8192,
"open-mixtral-8x7b": 32768,
"mistral-small-latest": 32768,
"mistral-medium-latest": 32768,
"mistral-large-latest": 32768,
}