Vyomakesh Dundigalla
1 year ago
committed by
GitHub
commit
ad3077f67d
@ -0,0 +1,86 @@
|
||||
import os
|
||||
|
||||
from dotenv import load_dotenv
|
||||
from swarms import Agent, OpenAIChat
|
||||
from swarms.agents.multion_agent import MultiOnAgent
|
||||
from swarms.memory.chroma_db import ChromaDB
|
||||
from swarms.tools.tool import tool
|
||||
from swarms.utils.code_interpreter import SubprocessCodeInterpreter
|
||||
|
||||
# Load the environment variables
|
||||
load_dotenv()
|
||||
|
||||
|
||||
# Memory
|
||||
chroma_db = ChromaDB()
|
||||
|
||||
|
||||
# MultiOntool
|
||||
@tool
|
||||
def multion_tool(
|
||||
task: str,
|
||||
api_key: str = os.environ.get("MULTION_API_KEY"),
|
||||
):
|
||||
"""
|
||||
Executes a task using the MultiOnAgent.
|
||||
|
||||
Args:
|
||||
task (str): The task to be executed.
|
||||
api_key (str, optional): The API key for the MultiOnAgent. Defaults to the value of the MULTION_API_KEY environment variable.
|
||||
|
||||
Returns:
|
||||
The result of the task execution.
|
||||
"""
|
||||
multion = MultiOnAgent(multion_api_key=api_key)
|
||||
return multion(task)
|
||||
|
||||
|
||||
# Execute the interpreter tool
|
||||
@tool
|
||||
def execute_interpreter_tool(
|
||||
code: str,
|
||||
):
|
||||
"""
|
||||
Executes a single command using the interpreter.
|
||||
|
||||
Args:
|
||||
task (str): The command to be executed.
|
||||
|
||||
Returns:
|
||||
None
|
||||
"""
|
||||
out = SubprocessCodeInterpreter(debug_mode=True)
|
||||
out = out.run(code)
|
||||
return code
|
||||
|
||||
|
||||
# Get the API key from the environment
|
||||
api_key = os.environ.get("OPENAI_API_KEY")
|
||||
|
||||
# Initialize the language model
|
||||
llm = OpenAIChat(
|
||||
temperature=0.5,
|
||||
openai_api_key=api_key,
|
||||
)
|
||||
|
||||
|
||||
# Initialize the workflow
|
||||
agent = Agent(
|
||||
agent_name="Research Agent",
|
||||
agent_description="An agent that performs research tasks.",
|
||||
system_prompt="Perform a research task.",
|
||||
llm=llm,
|
||||
max_loops=1,
|
||||
dashboard=True,
|
||||
# tools=[multion_tool, execute_interpreter_tool],
|
||||
verbose=True,
|
||||
long_term_memory=chroma_db,
|
||||
stopping_token="done",
|
||||
)
|
||||
|
||||
# Run the workflow on a task
|
||||
out = agent.run(
|
||||
"Generate a 10,000 word blog on health and wellness, and say done"
|
||||
" when you are done"
|
||||
)
|
||||
print(out)
|
||||
@ -0,0 +1,19 @@
|
||||
from swarms.structs.agent import Agent
|
||||
from swarms.structs.message_pool import MessagePool
|
||||
from swarms import OpenAIChat
|
||||
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
agent2 = Agent(llm=OpenAIChat(), agent_name="agent2")
|
||||
agent3 = Agent(llm=OpenAIChat(), agent_name="agent3")
|
||||
|
||||
moderator = Agent(agent_name="moderator")
|
||||
agents = [agent1, agent2, agent3]
|
||||
message_pool = MessagePool(
|
||||
agents=agents, moderator=moderator, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, agent2!", turn=1)
|
||||
message_pool.add(agent=agent2, content="Hello, agent1!", turn=1)
|
||||
message_pool.add(agent=agent3, content="Hello, agent1!", turn=1)
|
||||
message_pool.get_all_messages()
|
||||
message_pool.get_visible_messages(agent=agent1, turn=1)
|
||||
message_pool.get_visible_messages(agent=agent2, turn=1)
|
||||
@ -0,0 +1,69 @@
|
||||
import os
|
||||
import multion
|
||||
|
||||
from swarms.models.base_llm import AbstractLLM
|
||||
from dotenv import load_dotenv
|
||||
|
||||
# Load environment variables
|
||||
load_dotenv()
|
||||
|
||||
# Muliton key
|
||||
MULTION_API_KEY = os.getenv("MULTION_API_KEY")
|
||||
|
||||
|
||||
class MultiOnAgent(AbstractLLM):
|
||||
"""
|
||||
Represents a multi-on agent that performs browsing tasks.
|
||||
|
||||
Args:
|
||||
max_steps (int): The maximum number of steps to perform during browsing.
|
||||
starting_url (str): The starting URL for browsing.
|
||||
|
||||
Attributes:
|
||||
max_steps (int): The maximum number of steps to perform during browsing.
|
||||
starting_url (str): The starting URL for browsing.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
multion_api_key: str = MULTION_API_KEY,
|
||||
max_steps: int = 4,
|
||||
starting_url: str = "https://www.google.com",
|
||||
*args,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(*args, **kwargs)
|
||||
self.multion_api_key = multion_api_key
|
||||
self.max_steps = max_steps
|
||||
self.starting_url = starting_url
|
||||
|
||||
self.multion = multion.login(
|
||||
use_api=True,
|
||||
multion_api_key=str(multion_api_key),
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
def run(self, task: str, *args, **kwargs):
|
||||
"""
|
||||
Runs a browsing task.
|
||||
|
||||
Args:
|
||||
task (str): The task to perform during browsing.
|
||||
*args: Additional positional arguments.
|
||||
**kwargs: Additional keyword arguments.
|
||||
|
||||
Returns:
|
||||
dict: The response from the browsing task.
|
||||
"""
|
||||
response = self.multion.browse(
|
||||
{
|
||||
"cmd": task,
|
||||
"url": self.starting_url,
|
||||
"maxSteps": self.max_steps,
|
||||
},
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
return response.result, response.status, response.lastUrl
|
||||
@ -0,0 +1,528 @@
|
||||
import base64
|
||||
import os
|
||||
import time
|
||||
from io import BytesIO
|
||||
from typing import List, Literal, Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
from PIL import Image
|
||||
from pydantic import BaseModel, Field
|
||||
from transformers import (
|
||||
AutoModelForCausalLM,
|
||||
LlamaTokenizer,
|
||||
TextIteratorStreamer,
|
||||
)
|
||||
|
||||
from swarms.models.base_multimodal_model import BaseMultiModalModel
|
||||
from swarms.utils.logger import logger
|
||||
|
||||
MODEL_PATH = "THUDM/cogvlm-chat-hf"
|
||||
TOKENIZER_PATH = "lmsys/vicuna-7b-v1.5"
|
||||
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
||||
QUANT_ENABLED = False
|
||||
|
||||
|
||||
class ImageUrl(BaseModel):
|
||||
url: str
|
||||
|
||||
|
||||
class TextContent(BaseModel):
|
||||
type: Literal["text"]
|
||||
text: str
|
||||
|
||||
|
||||
class ImageUrlContent(BaseModel):
|
||||
type: Literal["image_url"]
|
||||
image_url: ImageUrl
|
||||
|
||||
|
||||
ContentItem = Union[TextContent, ImageUrlContent]
|
||||
|
||||
|
||||
class ChatMessageInput(BaseModel):
|
||||
role: Literal["user", "assistant", "system"]
|
||||
content: Union[str, List[ContentItem]]
|
||||
name: Optional[str] = None
|
||||
|
||||
|
||||
class ChatMessageResponse(BaseModel):
|
||||
role: Literal["assistant"]
|
||||
content: str = None
|
||||
name: Optional[str] = None
|
||||
|
||||
|
||||
class DeltaMessage(BaseModel):
|
||||
role: Optional[Literal["user", "assistant", "system"]] = None
|
||||
content: Optional[str] = None
|
||||
|
||||
|
||||
class ChatCompletionRequest(BaseModel):
|
||||
model: str
|
||||
messages: List[ChatMessageInput]
|
||||
temperature: Optional[float] = 0.8
|
||||
top_p: Optional[float] = 0.8
|
||||
max_tokens: Optional[int] = None
|
||||
stream: Optional[bool] = False
|
||||
# Additional parameters
|
||||
repetition_penalty: Optional[float] = 1.0
|
||||
|
||||
|
||||
class ChatCompletionResponseChoice(BaseModel):
|
||||
index: int
|
||||
message: ChatMessageResponse
|
||||
|
||||
|
||||
class ChatCompletionResponseStreamChoice(BaseModel):
|
||||
index: int
|
||||
delta: DeltaMessage
|
||||
|
||||
|
||||
class UsageInfo(BaseModel):
|
||||
prompt_tokens: int = 0
|
||||
total_tokens: int = 0
|
||||
completion_tokens: Optional[int] = 0
|
||||
|
||||
|
||||
class ChatCompletionResponse(BaseModel):
|
||||
model: str
|
||||
object: Literal["chat.completion", "chat.completion.chunk"]
|
||||
choices: List[
|
||||
Union[
|
||||
ChatCompletionResponseChoice,
|
||||
ChatCompletionResponseStreamChoice,
|
||||
]
|
||||
]
|
||||
created: Optional[int] = Field(
|
||||
default_factory=lambda: int(time.time())
|
||||
)
|
||||
usage: Optional[UsageInfo] = None
|
||||
|
||||
|
||||
# async def create_chat_completion(request: ChatCompletionRequest):
|
||||
# global model, tokenizer
|
||||
|
||||
# gen_params = dict(
|
||||
# messages=request.messages,
|
||||
# temperature=request.temperature,
|
||||
# top_p=request.top_p,
|
||||
# max_tokens=request.max_tokens or 1024,
|
||||
# echo=False,
|
||||
# stream=request.stream,
|
||||
# )
|
||||
|
||||
# # if request.stream:
|
||||
# # predict(request.model, gen_params)
|
||||
# # response = generate_cogvlm(model, tokenizer, gen_params)
|
||||
|
||||
# usage = UsageInfo()
|
||||
|
||||
# message = ChatMessageResponse(
|
||||
# role="assistant",
|
||||
# content=response["text"],
|
||||
# )
|
||||
# logger.debug(f"==== message ====\n{message}")
|
||||
# choice_data = ChatCompletionResponseChoice(
|
||||
# index=0,
|
||||
# message=message,
|
||||
# )
|
||||
# task_usage = UsageInfo.model_validate(response["usage"])
|
||||
# for usage_key, usage_value in task_usage.model_dump().items():
|
||||
# setattr(
|
||||
# usage, usage_key, getattr(usage, usage_key) + usage_value
|
||||
# )
|
||||
# return ChatCompletionResponse(
|
||||
# model=request.model,
|
||||
# choices=[choice_data],
|
||||
# object="chat.completion",
|
||||
# usage=usage,
|
||||
# )
|
||||
|
||||
|
||||
class CogVLMMultiModal(BaseMultiModalModel):
|
||||
"""
|
||||
Initializes the CogVLM model.
|
||||
|
||||
Args:
|
||||
model_name (str): The path or name of the pre-trained model.
|
||||
tokenizer (str): The path or name of the tokenizer.
|
||||
device (str): The device to run the model on.
|
||||
quantize (bool): Whether to enable quantization.
|
||||
torch_type (str): The torch data type to use.
|
||||
temperature (float): The temperature for sampling.
|
||||
top_p (float): The top-p value for sampling.
|
||||
max_tokens (int): The maximum number of tokens to generate.
|
||||
echo (bool): Whether to echo the input text.
|
||||
stream (bool): Whether to stream the output.
|
||||
repetition_penalty (float): The repetition penalty for sampling.
|
||||
do_sample (bool): Whether to use sampling during generation.
|
||||
*args: Additional positional arguments.
|
||||
**kwargs: Additional keyword arguments.
|
||||
|
||||
Methods:
|
||||
run: Generates a response using the CogVLM model.
|
||||
generate_stream_cogvlm: Generates a stream of responses using the CogVLM model in inference mode.
|
||||
process_history_and_images: Processes history messages to extract text, identify the last user query, and convert base64 encoded image URLs to PIL images.
|
||||
|
||||
Example:
|
||||
>>> model = CogVLMMultiModal()
|
||||
>>> response = model("Describe this image with meticlous details.", "https://example.com/image.jpg")
|
||||
>>> print(response)
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
model_name: str = MODEL_PATH,
|
||||
tokenizer: str = TOKENIZER_PATH,
|
||||
device: str = DEVICE,
|
||||
quantize: bool = QUANT_ENABLED,
|
||||
torch_type: str = "float16",
|
||||
temperature: float = 0.5,
|
||||
top_p: float = 0.9,
|
||||
max_tokens: int = 3500,
|
||||
echo: bool = False,
|
||||
stream: bool = False,
|
||||
repetition_penalty: float = 1.0,
|
||||
do_sample: bool = True,
|
||||
*args,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__()
|
||||
self.model_name = model_name
|
||||
self.device = device
|
||||
self.tokenizer = tokenizer
|
||||
self.device = device
|
||||
self.quantize = quantize
|
||||
self.torch_type = torch_type
|
||||
self.temperature = temperature
|
||||
self.top_p = top_p
|
||||
self.max_tokens = max_tokens
|
||||
self.echo = echo
|
||||
self.stream = stream
|
||||
self.repetition_penalty = repetition_penalty
|
||||
self.do_sample = do_sample
|
||||
|
||||
if os.environ.get("QUANT_ENABLED"):
|
||||
pass
|
||||
else:
|
||||
with torch.cuda.device(device):
|
||||
__, total_bytes = torch.cuda.mem_get_info()
|
||||
total_gb = total_bytes / (1 << 30)
|
||||
if total_gb < 40:
|
||||
pass
|
||||
else:
|
||||
pass
|
||||
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
self.tokenizer = LlamaTokenizer.from_pretrained(
|
||||
tokenizer, trust_remote_code=True
|
||||
)
|
||||
|
||||
if (
|
||||
torch.cuda.is_available()
|
||||
and torch.cuda.get_device_capability()[0] >= 8
|
||||
):
|
||||
torch_type = torch.bfloat16
|
||||
else:
|
||||
torch_type = torch.float16
|
||||
|
||||
print(
|
||||
f"========Use torch type as:{torch_type} with"
|
||||
f" device:{device}========\n\n"
|
||||
)
|
||||
|
||||
if "cuda" in device:
|
||||
if QUANT_ENABLED:
|
||||
self.model = AutoModelForCausalLM.from_pretrained(
|
||||
model_name,
|
||||
load_in_4bit=True,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch_type,
|
||||
low_cpu_mem_usage=True,
|
||||
*args,
|
||||
**kwargs,
|
||||
).eval()
|
||||
else:
|
||||
self.model = (
|
||||
AutoModelForCausalLM.from_pretrained(
|
||||
model_name,
|
||||
load_in_4bit=False,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch_type,
|
||||
low_cpu_mem_usage=True,
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
.to(device)
|
||||
.eval()
|
||||
)
|
||||
|
||||
else:
|
||||
self.model = (
|
||||
AutoModelForCausalLM.from_pretrained(
|
||||
model_name,
|
||||
trust_remote_code=True,
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
.float()
|
||||
.to(device)
|
||||
.eval()
|
||||
)
|
||||
|
||||
def run(self, task: str, img: str, *args, **kwargs):
|
||||
"""
|
||||
Generates a response using the CogVLM model. It processes the chat history and image data, if any,
|
||||
and then invokes the model to generate a response.
|
||||
"""
|
||||
messages = [task]
|
||||
|
||||
params = dict(
|
||||
messages=messages,
|
||||
temperature=self.temperature,
|
||||
repitition_penalty=self.repetition_penalty,
|
||||
top_p=self.top_p,
|
||||
max_new_tokens=self.max_tokens,
|
||||
)
|
||||
|
||||
for response in self.generate_stream_cogvlm(params):
|
||||
pass
|
||||
|
||||
return response
|
||||
|
||||
@torch.inference_mode()
|
||||
def generate_stream_cogvlm(
|
||||
self,
|
||||
params: dict,
|
||||
):
|
||||
"""
|
||||
Generates a stream of responses using the CogVLM model in inference mode.
|
||||
It's optimized to handle continuous input-output interactions with the model in a streaming manner.
|
||||
"""
|
||||
messages = params["messages"]
|
||||
temperature = float(params.get("temperature", 1.0))
|
||||
repetition_penalty = float(
|
||||
params.get("repetition_penalty", 1.0)
|
||||
)
|
||||
top_p = float(params.get("top_p", 1.0))
|
||||
max_new_tokens = int(params.get("max_tokens", 256))
|
||||
query, history, image_list = self.process_history_and_images(
|
||||
messages
|
||||
)
|
||||
|
||||
logger.debug(f"==== request ====\n{query}")
|
||||
|
||||
input_by_model = self.model.build_conversation_input_ids(
|
||||
self.tokenizer,
|
||||
query=query,
|
||||
history=history,
|
||||
images=[image_list[-1]],
|
||||
)
|
||||
inputs = {
|
||||
"input_ids": (
|
||||
input_by_model["input_ids"]
|
||||
.unsqueeze(0)
|
||||
.to(self.device)
|
||||
),
|
||||
"token_type_ids": (
|
||||
input_by_model["token_type_ids"]
|
||||
.unsqueeze(0)
|
||||
.to(self.device)
|
||||
),
|
||||
"attention_mask": (
|
||||
input_by_model["attention_mask"]
|
||||
.unsqueeze(0)
|
||||
.to(self.device)
|
||||
),
|
||||
"images": [
|
||||
[
|
||||
input_by_model["images"][0]
|
||||
.to(self.device)
|
||||
.to(self.torch_type)
|
||||
]
|
||||
],
|
||||
}
|
||||
if (
|
||||
"cross_images" in input_by_model
|
||||
and input_by_model["cross_images"]
|
||||
):
|
||||
inputs["cross_images"] = [
|
||||
[
|
||||
input_by_model["cross_images"][0]
|
||||
.to(self.device)
|
||||
.to(self.torch_type)
|
||||
]
|
||||
]
|
||||
|
||||
input_echo_len = len(inputs["input_ids"][0])
|
||||
streamer = TextIteratorStreamer(
|
||||
tokenizer=self.tokenizer,
|
||||
timeout=60.0,
|
||||
skip_promptb=True,
|
||||
skip_special_tokens=True,
|
||||
)
|
||||
gen_kwargs = {
|
||||
"repetition_penalty": repetition_penalty,
|
||||
"max_new_tokens": max_new_tokens,
|
||||
"do_sample": True if temperature > 1e-5 else False,
|
||||
"top_p": top_p if temperature > 1e-5 else 0,
|
||||
"streamer": streamer,
|
||||
}
|
||||
if temperature > 1e-5:
|
||||
gen_kwargs["temperature"] = temperature
|
||||
|
||||
total_len = 0
|
||||
generated_text = ""
|
||||
with torch.no_grad():
|
||||
self.model.generate(**inputs, **gen_kwargs)
|
||||
for next_text in streamer:
|
||||
generated_text += next_text
|
||||
yield {
|
||||
"text": generated_text,
|
||||
"usage": {
|
||||
"prompt_tokens": input_echo_len,
|
||||
"completion_tokens": (
|
||||
total_len - input_echo_len
|
||||
),
|
||||
"total_tokens": total_len,
|
||||
},
|
||||
}
|
||||
ret = {
|
||||
"text": generated_text,
|
||||
"usage": {
|
||||
"prompt_tokens": input_echo_len,
|
||||
"completion_tokens": total_len - input_echo_len,
|
||||
"total_tokens": total_len,
|
||||
},
|
||||
}
|
||||
yield ret
|
||||
|
||||
def process_history_and_images(
|
||||
self,
|
||||
messages: List[ChatMessageInput],
|
||||
) -> Tuple[
|
||||
Optional[str],
|
||||
Optional[List[Tuple[str, str]]],
|
||||
Optional[List[Image.Image]],
|
||||
]:
|
||||
"""
|
||||
Process history messages to extract text, identify the last user query,
|
||||
and convert base64 encoded image URLs to PIL images.
|
||||
|
||||
Args:
|
||||
messages(List[ChatMessageInput]): List of ChatMessageInput objects.
|
||||
return: A tuple of three elements:
|
||||
- The last user query as a string.
|
||||
- Text history formatted as a list of tuples for the model.
|
||||
- List of PIL Image objects extracted from the messages.
|
||||
"""
|
||||
formatted_history = []
|
||||
image_list = []
|
||||
last_user_query = ""
|
||||
|
||||
for i, message in enumerate(messages):
|
||||
role = message.role
|
||||
content = message.content
|
||||
|
||||
# Extract text content
|
||||
if isinstance(content, list): # text
|
||||
text_content = " ".join(
|
||||
item.text
|
||||
for item in content
|
||||
if isinstance(item, TextContent)
|
||||
)
|
||||
else:
|
||||
text_content = content
|
||||
|
||||
# Extract image data
|
||||
if isinstance(content, list): # image
|
||||
for item in content:
|
||||
if isinstance(item, ImageUrlContent):
|
||||
image_url = item.image_url.url
|
||||
if image_url.startswith(
|
||||
"data:image/jpeg;base64,"
|
||||
):
|
||||
base64_encoded_image = image_url.split(
|
||||
"data:image/jpeg;base64,"
|
||||
)[1]
|
||||
image_data = base64.b64decode(
|
||||
base64_encoded_image
|
||||
)
|
||||
image = Image.open(
|
||||
BytesIO(image_data)
|
||||
).convert("RGB")
|
||||
image_list.append(image)
|
||||
|
||||
# Format history
|
||||
if role == "user":
|
||||
if i == len(messages) - 1:
|
||||
last_user_query = text_content
|
||||
else:
|
||||
formatted_history.append((text_content, ""))
|
||||
elif role == "assistant":
|
||||
if formatted_history:
|
||||
if formatted_history[-1][1] != "":
|
||||
assert False, (
|
||||
"the last query is answered. answer"
|
||||
f" again. {formatted_history[-1][0]},"
|
||||
f" {formatted_history[-1][1]},"
|
||||
f" {text_content}"
|
||||
)
|
||||
formatted_history[-1] = (
|
||||
formatted_history[-1][0],
|
||||
text_content,
|
||||
)
|
||||
else:
|
||||
assert False, "assistant reply before user"
|
||||
else:
|
||||
assert False, f"unrecognized role: {role}"
|
||||
|
||||
return last_user_query, formatted_history, image_list
|
||||
|
||||
async def predict(self, params: dict):
|
||||
"""
|
||||
Handle streaming predictions. It continuously generates responses for a given input stream.
|
||||
This is particularly useful for real-time, continuous interactions with the model.
|
||||
"""
|
||||
|
||||
choice_data = ChatCompletionResponseStreamChoice(
|
||||
index=0,
|
||||
delta=DeltaMessage(role="assistant"),
|
||||
finish_reason=None,
|
||||
)
|
||||
chunk = ChatCompletionResponse(
|
||||
model=self.model_name,
|
||||
choices=[choice_data],
|
||||
object="chat.completion.chunk",
|
||||
)
|
||||
yield f"{chunk.model_dump_json(exclude_unset=True)}"
|
||||
|
||||
previous_text = ""
|
||||
for new_response in self.generate_stream_cogvlm(params):
|
||||
decoded_unicode = new_response["text"]
|
||||
delta_text = decoded_unicode[len(previous_text) :]
|
||||
previous_text = decoded_unicode
|
||||
delta = DeltaMessage(
|
||||
content=delta_text,
|
||||
role="assistant",
|
||||
)
|
||||
choice_data = ChatCompletionResponseStreamChoice(
|
||||
index=0,
|
||||
delta=delta,
|
||||
)
|
||||
chunk = ChatCompletionResponse(
|
||||
model=self.model_name,
|
||||
choices=[choice_data],
|
||||
object="chat.completion.chunk",
|
||||
)
|
||||
yield f"{chunk.model_dump_json(exclude_unset=True)}"
|
||||
choice_data = ChatCompletionResponseStreamChoice(
|
||||
index=0,
|
||||
delta=DeltaMessage(),
|
||||
)
|
||||
chunk = ChatCompletionResponse(
|
||||
model=self.model_name,
|
||||
choices=[choice_data],
|
||||
object="chat.completion.chunk",
|
||||
)
|
||||
yield f"{chunk.model_dump_json(exclude_unset=True)}"
|
||||
@ -0,0 +1,87 @@
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
import json
|
||||
from swarms.models.base_llm import AbstractLLM
|
||||
from typing import Any
|
||||
|
||||
|
||||
class FireFunctionCaller(AbstractLLM):
|
||||
"""
|
||||
A class that represents a caller for the FireFunction model.
|
||||
|
||||
Args:
|
||||
model_name (str): The name of the model to be used.
|
||||
device (str): The device to be used.
|
||||
function_spec (Any): The specification of the function.
|
||||
max_tokens (int): The maximum number of tokens.
|
||||
system_prompt (str): The system prompt.
|
||||
*args: Variable length argument list.
|
||||
**kwargs: Arbitrary keyword arguments.
|
||||
|
||||
Methods:
|
||||
run(self, task: str, *args, **kwargs) -> None: Run the function with the given task and arguments.
|
||||
|
||||
Examples:
|
||||
>>> fire_function_caller = FireFunctionCaller()
|
||||
>>> fire_function_caller.run("Add 2 and 3")
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
model_name: str = "fireworks-ai/firefunction-v1",
|
||||
device: str = "cuda",
|
||||
function_spec: Any = None,
|
||||
max_tokens: int = 3000,
|
||||
system_prompt: str = "You are a helpful assistant with access to functions. Use them if required.",
|
||||
*args,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__(model_name, device)
|
||||
self.model_name = model_name
|
||||
self.device = device
|
||||
self.fucntion_spec = function_spec
|
||||
self.max_tokens = max_tokens
|
||||
self.system_prompt = system_prompt
|
||||
|
||||
self.model = AutoModelForCausalLM.from_pretrained(
|
||||
model_name, device_map="auto", *args, **kwargs
|
||||
)
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
|
||||
self.functions = json.dumps(function_spec, indent=4)
|
||||
|
||||
def run(self, task: str, *args, **kwargs):
|
||||
"""
|
||||
Run the function with the given task and arguments.
|
||||
|
||||
Args:
|
||||
task (str): The task to be performed.
|
||||
*args: Variable length argument list.
|
||||
**kwargs: Arbitrary keyword arguments.
|
||||
|
||||
Returns:
|
||||
None
|
||||
"""
|
||||
messages = [
|
||||
{"role": "functions", "content": self.functions},
|
||||
{
|
||||
"role": "system",
|
||||
"content": self.system_prompt,
|
||||
},
|
||||
{
|
||||
"role": "user",
|
||||
"content": task,
|
||||
},
|
||||
]
|
||||
|
||||
model_inputs = self.tokenizer.apply_chat_template(
|
||||
messages, return_tensors="pt"
|
||||
).to(self.model.device)
|
||||
|
||||
generated_ids = self.model.generate(
|
||||
model_inputs,
|
||||
max_new_tokens=self.max_tokens,
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
decoded = self.tokenizer.batch_decode(generated_ids)
|
||||
print(decoded[0])
|
||||
@ -0,0 +1,43 @@
|
||||
from unittest.mock import MagicMock
|
||||
from swarms.models.fire_function import FireFunctionCaller
|
||||
|
||||
|
||||
def test_fire_function_caller_run(mocker):
|
||||
# Create mock model and tokenizer
|
||||
model = MagicMock()
|
||||
tokenizer = MagicMock()
|
||||
mocker.patch.object(FireFunctionCaller, "model", model)
|
||||
mocker.patch.object(FireFunctionCaller, "tokenizer", tokenizer)
|
||||
|
||||
# Create mock task and arguments
|
||||
task = "Add 2 and 3"
|
||||
args = (2, 3)
|
||||
kwargs = {}
|
||||
|
||||
# Create mock generated_ids and decoded output
|
||||
generated_ids = [1, 2, 3]
|
||||
decoded_output = "5"
|
||||
model.generate.return_value = generated_ids
|
||||
tokenizer.batch_decode.return_value = [decoded_output]
|
||||
|
||||
# Create FireFunctionCaller instance
|
||||
fire_function_caller = FireFunctionCaller()
|
||||
|
||||
# Run the function
|
||||
fire_function_caller.run(task, *args, **kwargs)
|
||||
|
||||
# Assert model.generate was called with the correct inputs
|
||||
model.generate.assert_called_once_with(
|
||||
tokenizer.apply_chat_template.return_value,
|
||||
max_new_tokens=fire_function_caller.max_tokens,
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
# Assert tokenizer.batch_decode was called with the correct inputs
|
||||
tokenizer.batch_decode.assert_called_once_with(generated_ids)
|
||||
|
||||
# Assert the decoded output is printed
|
||||
assert decoded_output in mocker.patch.object(
|
||||
print, "call_args_list"
|
||||
)
|
||||
@ -0,0 +1,214 @@
|
||||
import hashlib
|
||||
from time import time_ns
|
||||
from typing import Callable, List, Optional, Sequence, Union
|
||||
|
||||
from swarms.structs.agent import Agent
|
||||
from swarms.structs.base_swarm import BaseSwarm
|
||||
from swarms.utils.loguru_logger import logger
|
||||
|
||||
|
||||
def _hash(input: str):
|
||||
"""
|
||||
Hashes the input string using SHA256 algorithm.
|
||||
|
||||
Args:
|
||||
input (str): The string to be hashed.
|
||||
|
||||
Returns:
|
||||
str: The hexadecimal representation of the hash value.
|
||||
"""
|
||||
hex_dig = hashlib.sha256(input.encode("utf-8")).hexdigest()
|
||||
return hex_dig
|
||||
|
||||
|
||||
def msg_hash(
|
||||
agent: Agent, content: str, turn: int, msg_type: str = "text"
|
||||
):
|
||||
"""
|
||||
Generate a hash value for a message.
|
||||
|
||||
Args:
|
||||
agent (Agent): The agent sending the message.
|
||||
content (str): The content of the message.
|
||||
turn (int): The turn number of the message.
|
||||
msg_type (str, optional): The type of the message. Defaults to "text".
|
||||
|
||||
Returns:
|
||||
int: The hash value of the message.
|
||||
"""
|
||||
time = time_ns()
|
||||
return _hash(
|
||||
f"agent: {agent.agent_name}\ncontent: {content}\ntimestamp:"
|
||||
f" {str(time)}\nturn: {turn}\nmsg_type: {msg_type}"
|
||||
)
|
||||
|
||||
|
||||
class MessagePool(BaseSwarm):
|
||||
"""
|
||||
A class representing a message pool for agents in a swarm.
|
||||
|
||||
Attributes:
|
||||
agents (Optional[Sequence[Agent]]): The list of agents in the swarm.
|
||||
moderator (Optional[Agent]): The moderator agent.
|
||||
turns (Optional[int]): The number of turns.
|
||||
routing_function (Optional[Callable]): The routing function for message distribution.
|
||||
show_names (Optional[bool]): Flag indicating whether to show agent names.
|
||||
messages (List[Dict]): The list of messages in the pool.
|
||||
|
||||
Examples:
|
||||
>>> from swarms.structs.agent import Agent
|
||||
>>> from swarms.structs.message_pool import MessagePool
|
||||
>>> agent1 = Agent(agent_name="agent1")
|
||||
>>> agent2 = Agent(agent_name="agent2")
|
||||
>>> agent3 = Agent(agent_name="agent3")
|
||||
>>> moderator = Agent(agent_name="moderator")
|
||||
>>> agents = [agent1, agent2, agent3]
|
||||
>>> message_pool = MessagePool(agents=agents, moderator=moderator, turns=5)
|
||||
>>> message_pool.add(agent=agent1, content="Hello, agent2!", turn=1)
|
||||
>>> message_pool.add(agent=agent2, content="Hello, agent1!", turn=1)
|
||||
>>> message_pool.add(agent=agent3, content="Hello, agent1!", turn=1)
|
||||
>>> message_pool.get_all_messages()
|
||||
[{'agent': Agent(agent_name='agent1'), 'content': 'Hello, agent2!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent2'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent3'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}]
|
||||
>>> message_pool.get_visible_messages(agent=agent1, turn=1)
|
||||
[{'agent': Agent(agent_name='agent1'), 'content': 'Hello, agent2!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent2'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent3'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}]
|
||||
>>> message_pool.get_visible_messages(agent=agent2, turn=1)
|
||||
[{'agent': Agent(agent_name='agent1'), 'content': 'Hello, agent2!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent2'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}, {'agent': Agent(agent_name='agent3'), 'content': 'Hello, agent1!', 'turn': 1, 'visible_to': 'all', 'logged': True}]
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
agents: Optional[Sequence[Agent]] = None,
|
||||
moderator: Optional[Agent] = None,
|
||||
turns: Optional[int] = 5,
|
||||
routing_function: Optional[Callable] = None,
|
||||
show_names: Optional[bool] = False,
|
||||
autosave: Optional[bool] = False,
|
||||
*args,
|
||||
**kwargs,
|
||||
):
|
||||
super().__init__()
|
||||
|
||||
self.agent = agents
|
||||
self.moderator = moderator
|
||||
self.turns = turns
|
||||
self.routing_function = routing_function
|
||||
self.show_names = show_names
|
||||
self.autosave = autosave
|
||||
|
||||
self.messages = []
|
||||
|
||||
logger.info("MessagePool initialized")
|
||||
logger.info(f"Number of agents: {len(agents)}")
|
||||
logger.info(
|
||||
f"Agents: {[agent.agent_name for agent in agents]}"
|
||||
)
|
||||
logger.info(f"moderator: {moderator.agent_name} is available")
|
||||
logger.info(f"Number of turns: {turns}")
|
||||
|
||||
def add(
|
||||
self,
|
||||
agent: Agent,
|
||||
content: str,
|
||||
turn: int,
|
||||
visible_to: Union[str, List[str]] = "all",
|
||||
logged: bool = True,
|
||||
):
|
||||
"""
|
||||
Add a message to the pool.
|
||||
|
||||
Args:
|
||||
agent (Agent): The agent sending the message.
|
||||
content (str): The content of the message.
|
||||
turn (int): The turn number.
|
||||
visible_to (Union[str, List[str]], optional): The agents who can see the message. Defaults to "all".
|
||||
logged (bool, optional): Flag indicating whether the message should be logged. Defaults to True.
|
||||
"""
|
||||
|
||||
self.messages.append(
|
||||
{
|
||||
"agent": agent,
|
||||
"content": content,
|
||||
"turn": turn,
|
||||
"visible_to": visible_to,
|
||||
"logged": logged,
|
||||
}
|
||||
)
|
||||
logger.info(f"Message added: {content}")
|
||||
|
||||
def reset(self):
|
||||
"""
|
||||
Reset the message pool.
|
||||
"""
|
||||
self.messages = []
|
||||
logger.info("MessagePool reset")
|
||||
|
||||
def last_turn(self):
|
||||
"""
|
||||
Get the last turn number.
|
||||
|
||||
Returns:
|
||||
int: The last turn number.
|
||||
"""
|
||||
if len(self.messages) == 0:
|
||||
return 0
|
||||
else:
|
||||
return self.messages[-1]["turn"]
|
||||
|
||||
@property
|
||||
def last_message(self):
|
||||
"""
|
||||
Get the last message in the pool.
|
||||
|
||||
Returns:
|
||||
dict: The last message.
|
||||
"""
|
||||
if len(self.messages) == 0:
|
||||
return None
|
||||
else:
|
||||
return self.messages[-1]
|
||||
|
||||
def get_all_messages(self):
|
||||
"""
|
||||
Get all messages in the pool.
|
||||
|
||||
Returns:
|
||||
List[Dict]: The list of all messages.
|
||||
"""
|
||||
return self.messages
|
||||
|
||||
def get_visible_messages(self, agent: Agent, turn: int):
|
||||
"""
|
||||
Get the visible messages for a given agent and turn.
|
||||
|
||||
Args:
|
||||
agent (Agent): The agent.
|
||||
turn (int): The turn number.
|
||||
|
||||
Returns:
|
||||
List[Dict]: The list of visible messages.
|
||||
"""
|
||||
# Get the messages before the current turn
|
||||
prev_messages = [
|
||||
message
|
||||
for message in self.messages
|
||||
if message["turn"] < turn
|
||||
]
|
||||
|
||||
visible_messages = []
|
||||
for message in prev_messages:
|
||||
if (
|
||||
message["visible_to"] == "all"
|
||||
or agent.agent_name in message["visible_to"]
|
||||
):
|
||||
visible_messages.append(message)
|
||||
return visible_messages
|
||||
|
||||
def query(self, query: str):
|
||||
"""
|
||||
Query a message from the messages list and then pass it to the moderator
|
||||
"""
|
||||
return [
|
||||
(mod, content)
|
||||
for mod, content in self.messages
|
||||
if mod == self.moderator
|
||||
]
|
||||
@ -0,0 +1,10 @@
|
||||
from loguru import logger
|
||||
|
||||
logger = logger.add(
|
||||
"MessagePool.log",
|
||||
level="INFO",
|
||||
colorize=True,
|
||||
format="<green>{time}</green> <level>{message}</level>",
|
||||
backtrace=True,
|
||||
diagnose=True,
|
||||
)
|
||||
@ -0,0 +1,57 @@
|
||||
import pytest
|
||||
from unittest.mock import patch, MagicMock
|
||||
from swarms.agents.multion_agent import MultiOnAgent
|
||||
|
||||
|
||||
@patch("swarms.agents.multion_agent.multion")
|
||||
def test_multion_agent_run(mock_multion):
|
||||
mock_response = MagicMock()
|
||||
mock_response.result = "result"
|
||||
mock_response.status = "status"
|
||||
mock_response.lastUrl = "lastUrl"
|
||||
mock_multion.browse.return_value = mock_response
|
||||
|
||||
agent = MultiOnAgent(
|
||||
multion_api_key="test_key",
|
||||
max_steps=5,
|
||||
starting_url="https://www.example.com",
|
||||
)
|
||||
result, status, last_url = agent.run("task")
|
||||
|
||||
assert result == "result"
|
||||
assert status == "status"
|
||||
assert last_url == "lastUrl"
|
||||
mock_multion.browse.assert_called_once_with(
|
||||
{
|
||||
"cmd": "task",
|
||||
"url": "https://www.example.com",
|
||||
"maxSteps": 5,
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
# Additional tests for different tasks
|
||||
@pytest.mark.parametrize(
|
||||
"task", ["task1", "task2", "task3", "task4", "task5"]
|
||||
)
|
||||
@patch("swarms.agents.multion_agent.multion")
|
||||
def test_multion_agent_run_different_tasks(mock_multion, task):
|
||||
mock_response = MagicMock()
|
||||
mock_response.result = "result"
|
||||
mock_response.status = "status"
|
||||
mock_response.lastUrl = "lastUrl"
|
||||
mock_multion.browse.return_value = mock_response
|
||||
|
||||
agent = MultiOnAgent(
|
||||
multion_api_key="test_key",
|
||||
max_steps=5,
|
||||
starting_url="https://www.example.com",
|
||||
)
|
||||
result, status, last_url = agent.run(task)
|
||||
|
||||
assert result == "result"
|
||||
assert status == "status"
|
||||
assert last_url == "lastUrl"
|
||||
mock_multion.browse.assert_called_once_with(
|
||||
{"cmd": task, "url": "https://www.example.com", "maxSteps": 5}
|
||||
)
|
||||
@ -0,0 +1,45 @@
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
|
||||
from swarms.models.fire_function import FireFunctionCaller
|
||||
|
||||
|
||||
def test_fire_function_caller_run(mocker):
|
||||
# Create mock model and tokenizer
|
||||
model = MagicMock()
|
||||
tokenizer = MagicMock()
|
||||
mocker.patch.object(FireFunctionCaller, "model", model)
|
||||
mocker.patch.object(FireFunctionCaller, "tokenizer", tokenizer)
|
||||
|
||||
# Create mock task and arguments
|
||||
task = "Add 2 and 3"
|
||||
args = (2, 3)
|
||||
kwargs = {}
|
||||
|
||||
# Create mock generated_ids and decoded output
|
||||
generated_ids = [1, 2, 3]
|
||||
decoded_output = "5"
|
||||
model.generate.return_value = generated_ids
|
||||
tokenizer.batch_decode.return_value = [decoded_output]
|
||||
|
||||
# Create FireFunctionCaller instance
|
||||
fire_function_caller = FireFunctionCaller()
|
||||
|
||||
# Run the function
|
||||
fire_function_caller.run(task, *args, **kwargs)
|
||||
|
||||
# Assert model.generate was called with the correct inputs
|
||||
model.generate.assert_called_once_with(
|
||||
tokenizer.apply_chat_template.return_value,
|
||||
max_new_tokens=fire_function_caller.max_tokens,
|
||||
*args,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
# Assert tokenizer.batch_decode was called with the correct inputs
|
||||
tokenizer.batch_decode.assert_called_once_with(generated_ids)
|
||||
|
||||
# Assert the decoded output is printed
|
||||
assert decoded_output in mocker.patch.object(
|
||||
print, "call_args_list"
|
||||
)
|
||||
@ -0,0 +1,117 @@
|
||||
from swarms.structs.agent import Agent
|
||||
from swarms.structs.message_pool import MessagePool
|
||||
from swarms import OpenAIChat
|
||||
|
||||
|
||||
def test_message_pool_initialization():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
agent2 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
moderator = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
agents = [agent1, agent2]
|
||||
message_pool = MessagePool(
|
||||
agents=agents, moderator=moderator, turns=5
|
||||
)
|
||||
|
||||
assert message_pool.agent == agents
|
||||
assert message_pool.moderator == moderator
|
||||
assert message_pool.turns == 5
|
||||
assert message_pool.messages == []
|
||||
|
||||
|
||||
def test_message_pool_add():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, world!", turn=1)
|
||||
|
||||
assert message_pool.messages == [
|
||||
{
|
||||
"agent": agent1,
|
||||
"content": "Hello, world!",
|
||||
"turn": 1,
|
||||
"visible_to": "all",
|
||||
"logged": True,
|
||||
}
|
||||
]
|
||||
|
||||
|
||||
def test_message_pool_reset():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, world!", turn=1)
|
||||
message_pool.reset()
|
||||
|
||||
assert message_pool.messages == []
|
||||
|
||||
|
||||
def test_message_pool_last_turn():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, world!", turn=1)
|
||||
|
||||
assert message_pool.last_turn() == 1
|
||||
|
||||
|
||||
def test_message_pool_last_message():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, world!", turn=1)
|
||||
|
||||
assert message_pool.last_message == {
|
||||
"agent": agent1,
|
||||
"content": "Hello, world!",
|
||||
"turn": 1,
|
||||
"visible_to": "all",
|
||||
"logged": True,
|
||||
}
|
||||
|
||||
|
||||
def test_message_pool_get_all_messages():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(agent=agent1, content="Hello, world!", turn=1)
|
||||
|
||||
assert message_pool.get_all_messages() == [
|
||||
{
|
||||
"agent": agent1,
|
||||
"content": "Hello, world!",
|
||||
"turn": 1,
|
||||
"visible_to": "all",
|
||||
"logged": True,
|
||||
}
|
||||
]
|
||||
|
||||
|
||||
def test_message_pool_get_visible_messages():
|
||||
agent1 = Agent(llm=OpenAIChat(), agent_name="agent1")
|
||||
agent2 = Agent(agent_name="agent2")
|
||||
message_pool = MessagePool(
|
||||
agents=[agent1, agent2], moderator=agent1, turns=5
|
||||
)
|
||||
message_pool.add(
|
||||
agent=agent1,
|
||||
content="Hello, agent2!",
|
||||
turn=1,
|
||||
visible_to=[agent2.agent_name],
|
||||
)
|
||||
|
||||
assert message_pool.get_visible_messages(
|
||||
agent=agent2, turn=2
|
||||
) == [
|
||||
{
|
||||
"agent": agent1,
|
||||
"content": "Hello, agent2!",
|
||||
"turn": 1,
|
||||
"visible_to": [agent2.agent_name],
|
||||
"logged": True,
|
||||
}
|
||||
]
|
||||
Loading…
Reference in new issue