smk/graph-of-thoughts
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial commit

Browse Source
This commit is contained in:
Nils Blach
2023-08-21 03:33:46 +02:00
commit 41a0a6e67c
54 changed files with 25134 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
graph_of_thoughts/controller/README.md
+117
View File
@@ -0,0 +1,117 @@
# Controller
The Controller class is responsible for traversing the Graph of Operations (GoO), which is a static structure that is constructed once, before the execution starts.
GoO prescribes the execution plan of thought operations and the Controller invokes their execution, generating the Graph Reasoning State (GRS).
In order for a GoO to be executed, an instance of Large Language Model (LLM) must be supplied to the controller.
Currently, the framework supports the following LLMs:
- GPT-4 / GPT-3.5 (Remote - OpenAI API)
- Llama-2 (Local - HuggingFace Transformers)
The following section describes how to instantiate individual LLMs and the Controller to run a defined GoO.
Furthermore, process of adding new LLM into the framework is outlined at the end.
## LLM Instantiation
- Create a copy of `config_template.json` named `config.json`.
- Fill configuration details based on the used model (below).
### GPT-4 / GPT-3.5
- Adjust predefined `chatgpt`, `chatgpt4` or create new configuration with unique key.
| Key | Value |
|---------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| model_id | Model name based on [OpenAI model overview](https://platform.openai.com/docs/models/overview). |
| prompt_token_cost | Price per 1000 prompt tokens based on [OpenAI pricing](https://openai.com/pricing), used for calculating cumulative price per LLM instance. |
| response_token_cost | Price per 1000 response tokens based on [OpenAI pricing](https://openai.com/pricing), used for calculating cumulative price per LLM instance. |
| temperature | Parameter of OpenAI models that controls randomness and the creativity of the responses (higher temperature = more diverse and unexpected responses). Value between 0.0 and 2.0, default is 1.0. More information can be found in the [OpenAI API reference](https://platform.openai.com/docs/api-reference/completions/create#completions/create-temperature). |
| max_tokens | The maximum number of tokens to generate in the chat completion. Value depends on the maximum context size of the model specified in the [OpenAI model overview](https://platform.openai.com/docs/models/overview). More information can be found in the [OpenAI API reference](https://platform.openai.com/docs/api-reference/chat/create#chat/create-max_tokens). |
| stop | String or array of strings specifying sequence of characters which if detected, stops further generation of tokens. More information can be found in the [OpenAI API reference](https://platform.openai.com/docs/api-reference/chat/create#chat/create-stop). |
| organization | Organization to use for the API requests (may be empty). |
| api_key | Personal API key that will be used to access OpenAI API. |
- Instantiate the language model based on the selected configuration key (predefined / custom).
```
lm = controller.ChatGPT(
"path/to/config.json",
model_name=<configuration key>
)
```
### Llama-2
- Requires local hardware to run inference and HuggingFace account.
- Adjust predefined `llama7b-hf`, `llama13b-hf`, `llama70b-hf` or create new configuration with unique key.
| Key | Value |
|---------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| model_id | Specifies HuggingFace Llama 2 model identifier (`meta-llama/<model_id>`). |
| cache_dir | Local directory where model will be downloaded and accessed. |
| prompt_token_cost | Price per 1000 prompt tokens (currently not used - local model = no cost). |
| response_token_cost | Price per 1000 response tokens (currently not used - local model = no cost). |
| temperature | Parameter that controls randomness and the creativity of the responses (higher temperature = more diverse and unexpected responses). Value between 0.0 and 1.0, default is 0.6. |
| top_k | Top-K sampling method described in [Transformers tutorial](https://huggingface.co/blog/how-to-generate). Default value is set to 10. |
| max_tokens | The maximum number of tokens to generate in the chat completion. More tokens require more memory. |
- Instantiate the language model based on the selected configuration key (predefined / custom).
```
lm = controller.Llama2HF(
"path/to/config.json",
model_name=<configuration key>
)
```
- Request access to Llama-2 via [Meta form](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) using same email address as for the HuggingFace account.
- After the access is granted, go to [HuggingFace Llama-2 model card](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf), log in and accept licence (_"You have been granted access to this model"_ message should appear).
- Generate HuggingFace access token.
- Log in from CLI with: `huggingface-cli login --token <your token>`.
Note: 4-bit quantization is used to reduce the model size for inference. During instantiation, the model is downloaded from HuggingFace into the cache directory specified in the `config.json`. Running queries using larger models will require multiple GPUs (splitting across many GPUs is done automatically by the Transformers library).
## Controller Instantiation
- Requires custom `Prompter`, `Parser` and instantiated `GraphOfOperations` - creation of these is described separately.
- Use instantiated `lm` from above.
- Prepare initial state (thought) as dictionary - this can be used in the initial prompts by the operations.
```
graph_of_operations = ...create
executor = controller.Controller(
lm,
graph_of_operations,
<CustomPrompter()>,
<CustomParser()>,
<initial state>,
)
executor.run()
executor.output_graph("path/to/output.json")
```
- After the run the graph is written to an output file, which contains individual operations, their thoughts, information about scores and validity and total amount of used tokens / cost.
## Adding LLMs
More LLMs can be added by following these steps:
- Create new class as a subclass of `AbstractLanguageModel`.
- Use the constructor for loading configuration and instantiating the language model (if needed).
```
class CustomLanguageModel(AbstractLanguageModel):
def __init__(
self,
config_path: str = "",
model_name: str = "llama7b-hf",
cache: bool = False
) -> None:
super().__init__(config_path, model_name, cache)
self.config: Dict = self.config[model_name]
# Load data from configuration into variables if needed
# Instantiate LLM if needed
```
- Implement `query` abstract method that is used to get a list of responses from the LLM (call to remote API or local model inference).
```
def query(self, query: str, num_responses: int = 1) -> Any:
# Support caching
# Call LLM and retrieve list of responses - based on num_responses
# Return LLM response structure (not only raw strings)
```
- Implement `get_response_texts` abstract method that is used to get a list of raw texts from the LLM response structure produced by `query`.
```
def get_response_texts(self, query_response: Union[List[Dict], Dict]) -> List[str]:
# Retrieve list of raw strings from the LLM response structure
```
graph_of_thoughts/controller/__init__.py
+4
View File
@@ -0,0 +1,4 @@
from .chatgpt import ChatGPT
from .llamachat_hf import Llama2HF
from .abstract_language_model import AbstractLanguageModel
from .controller import Controller
graph_of_thoughts/controller/abstract_language_model.py
+92
View File
@@ -0,0 +1,92 @@
# Copyright (c) 2023 ETH Zurich.
# All rights reserved.
#
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
#
# main author: Nils Blach
from abc import ABC, abstractmethod
from typing import List, Dict, Union, Any
import json
import os
import logging
class AbstractLanguageModel(ABC):
"""
Abstract base class that defines the interface for all language models.
"""
def __init__(
self, config_path: str = "", model_name: str = "", cache: bool = False
) -> None:
"""
Initialize the AbstractLanguageModel instance with configuration, model details, and caching options.
:param config_path: Path to the config file. Defaults to "".
:type config_path: str
:param model_name: Name of the language model. Defaults to "".
:type model_name: str
:param cache: Flag to determine whether to cache responses. Defaults to False.
:type cache: bool
"""
self.logger = logging.getLogger(self.__class__.__name__)
self.config: Dict = None
self.model_name: str = model_name
self.cache = cache
if self.cache:
self.respone_cache: Dict[str, List[Any]] = {}
self.load_config(config_path)
self.prompt_tokens: int = 0
self.completion_tokens: int = 0
self.cost: float = 0.0
def load_config(self, path: str) -> None:
"""
Load configuration from a specified path.
:param path: Path to the config file. If an empty path provided,
default is `config.json` in the current directory.
:type path: str
"""
if path == "":
current_dir = os.path.dirname(os.path.abspath(__file__))
path = os.path.join(current_dir, "config.json")
with open(path, "r") as f:
self.config = json.load(f)
self.logger.debug(f"Loaded config from {path} for {self.model_name}")
def clear_cache(self) -> None:
"""
Clear the response cache.
"""
self.respone_cache.clear()
@abstractmethod
def query(self, query: str, num_responses: int = 1) -> Any:
"""
Abstract method to query the language model.
:param query: The query to be posed to the language model.
:type query: str
:param num_responses: The number of desired responses.
:type num_responses: int
:return: The language model's response(s).
:rtype: Any
"""
pass
@abstractmethod
def get_response_texts(self, query_responses: Union[List[Dict], Dict]) -> List[str]:
"""
Abstract method to extract response texts from the language model's response(s).
:param query_responses: The responses returned from the language model.
:type query_responses: Union[List[Dict], Dict]
:return: List of textual responses.
:rtype: List[str]
"""
pass
graph_of_thoughts/controller/chatgpt.py
+156
View File
@@ -0,0 +1,156 @@
# Copyright (c) 2023 ETH Zurich.
# All rights reserved.
#
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
#
# main author: Nils Blach
import backoff
import openai
import os
import random
import time
from typing import List, Dict, Union
from .abstract_language_model import AbstractLanguageModel
class ChatGPT(AbstractLanguageModel):
"""
The ChatGPT class handles interactions with the OpenAI models using the provided configuration.
Inherits from the AbstractLanguageModel and implements its abstract methods.
"""
def __init__(
self, config_path: str = "", model_name: str = "chatgpt", cache: bool = False
) -> None:
"""
Initialize the ChatGPT instance with configuration, model details, and caching options.
:param config_path: Path to the configuration file. Defaults to "".
:type config_path: str
:param model_name: Name of the model, default is 'chatgpt'. Used to select the correct configuration.
:type model_name: str
:param cache: Flag to determine whether to cache responses. Defaults to False.
:type cache: bool
"""
super().__init__(config_path, model_name, cache)
self.config: Dict = self.config[model_name]
# The model_id is the id of the model that is used for chatgpt, i.e. gpt-4, gpt-3.5-turbo, etc.
self.model_id: str = self.config["model_id"]
# The prompt_token_cost and response_token_cost are the costs for 1000 prompt tokens and 1000 response tokens respectively.
self.prompt_token_cost: float = self.config["prompt_token_cost"]
self.response_token_cost: float = self.config["response_token_cost"]
# The temperature of a model is defined as the randomness of the model's output.
self.temperature: float = self.config["temperature"]
# The maximum number of tokens to generate in the chat completion.
self.max_tokens: int = self.config["max_tokens"]
# The stop sequence is a sequence of tokens that the model will stop generating at (it will not generate the stop sequence).
self.stop: Union[str, List[str]] = self.config["stop"]
# The account organization is the organization that is used for chatgpt.
self.organization: str = self.config["organization"]
if self.organization == "":
self.logger.warning("OPENAI_ORGANIZATION is not set")
else:
openai.organization = self.organization
# The api key is the api key that is used for chatgpt. Env variable OPENAI_API_KEY takes precedence over config.
self.api_key: str = os.getenv("OPENAI_API_KEY", self.config["api_key"])
if self.api_key == "":
raise ValueError("OPENAI_API_KEY is not set")
openai.api_key = self.api_key
def query(self, query: str, num_responses: int = 1) -> Dict:
"""
Query the OpenAI model for responses.
:param query: The query to be posed to the language model.
:type query: str
:param num_responses: Number of desired responses, default is 1.
:type num_responses: int
:return: Response(s) from the OpenAI model.
:rtype: Dict
"""
if self.cache and query in self.respone_cache:
return self.respone_cache[query]
if num_responses == 1:
response = self.chat([{"role": "user", "content": query}], num_responses)
else:
response = []
next_try = num_responses
total_num_attempts = num_responses
while num_responses > 0 and total_num_attempts > 0:
try:
assert next_try > 0
res = self.chat([{"role": "user", "content": query}], next_try)
response.append(res)
num_responses -= next_try
next_try = min(num_responses, next_try)
except Exception as e:
next_try = (next_try + 1) // 2
self.logger.warning(
f"Error in chatgpt: {e}, trying again with {next_try} samples"
)
time.sleep(random.randint(1, 3))
total_num_attempts -= 1
if self.cache:
self.respone_cache[query] = response
return response
@backoff.on_exception(
backoff.expo, openai.error.OpenAIError, max_time=10, max_tries=6
)
def chat(self, messages: List[Dict], num_responses: int = 1) -> Dict:
"""
Send chat messages to the OpenAI model and retrieves the model's response.
Implements backoff on OpenAI error.
:param messages: A list of message dictionaries for the chat.
:type messages: List[Dict]
:param num_responses: Number of desired responses, default is 1.
:type num_responses: int
:return: The OpenAI model's response.
:rtype: Dict
"""
response = openai.ChatCompletion.create(
model=self.model_id,
messages=messages,
temperature=self.temperature,
max_tokens=self.max_tokens,
n=num_responses,
stop=self.stop,
)
self.prompt_tokens += response["usage"]["prompt_tokens"]
self.completion_tokens += response["usage"]["completion_tokens"]
prompt_tokens_k = float(self.prompt_tokens) / 1000.0
completion_tokens_k = float(self.completion_tokens) / 1000.0
self.cost = (
self.prompt_token_cost * prompt_tokens_k
+ self.response_token_cost * completion_tokens_k
)
self.logger.info(
f"This is the response from chatgpt: {response}"
f"\nThis is the cost of the response: {self.cost}"
)
return response
def get_response_texts(self, query_response: Union[List[Dict], Dict]) -> List[str]:
"""
Extract the response texts from the query response.
:param query_response: The response dictionary (or list of dictionaries) from the OpenAI model.
:type query_response: Union[List[Dict], Dict]
:return: List of response strings.
:rtype: List[str]
"""
if isinstance(query_response, Dict):
query_response = [query_response]
return [
choice["message"]["content"]
for response in query_response
for choice in response["choices"]
]
graph_of_thoughts/controller/config_template.json
+49
View File
@@ -0,0 +1,49 @@
{
"chatgpt" : {
"model_id": "gpt-3.5-turbo",
"prompt_token_cost": 0.0015,
"response_token_cost": 0.002,
"temperature": 1.0,
"max_tokens": 1536,
"stop": null,
"organization": "",
"api_key": ""
},
"chatgpt4" : {
"model_id": "gpt-4",
"prompt_token_cost": 0.03,
"response_token_cost": 0.06,
"temperature": 1.0,
"max_tokens": 4096,
"stop": null,
"organization": "",
"api_key": ""
},
"llama7b-hf" : {
"model_id": "Llama-2-7b-chat-hf",
"cache_dir": "/llama",
"prompt_token_cost": 0.0,
"response_token_cost": 0.0,
"temperature": 0.6,
"top_k": 10,
"max_tokens": 4096
},
"llama13b-hf" : {
"model_id": "Llama-2-13b-chat-hf",
"cache_dir": "/llama",
"prompt_token_cost": 0.0,
"response_token_cost": 0.0,
"temperature": 0.6,
"top_k": 10,
"max_tokens": 4096
},
"llama70b-hf" : {
"model_id": "Llama-2-70b-chat-hf",
"cache_dir": "/llama",
"prompt_token_cost": 0.0,
"response_token_cost": 0.0,
"temperature": 0.6,
"top_k": 10,
"max_tokens": 4096
}
}
graph_of_thoughts/controller/controller.py
+152
View File
@@ -0,0 +1,152 @@
# Copyright (c) 2023 ETH Zurich.
# All rights reserved.
#
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
#
# main author: Nils Blach
import json
import logging
from typing import List
from .abstract_language_model import AbstractLanguageModel
from graph_of_thoughts.operations import GraphOfOperations, Thought
from graph_of_thoughts.prompter import Prompter
from graph_of_thoughts.parser import Parser
class Controller:
"""
Controller class to manage the execution flow of the Graph of Operations,
generating the Graph Reasoning State.
This involves language models, graph operations, prompting, and parsing.
"""
def __init__(
self,
lm: AbstractLanguageModel,
graph: GraphOfOperations,
prompter: Prompter,
parser: Parser,
problem_parameters: dict,
) -> None:
"""
Initialize the Controller instance with the language model,
operations graph, prompter, parser, and problem parameters.
:param lm: An instance of the AbstractLanguageModel.
:type lm: AbstractLanguageModel
:param graph: The Graph of Operations to be executed.
:type graph: OperationsGraph
:param prompter: An instance of the Prompter class, used to generate prompts.
:type prompter: Prompter
:param parser: An instance of the Parser class, used to parse responses.
:type parser: Parser
:param problem_parameters: Initial parameters/state of the problem.
:type problem_parameters: dict
"""
self.logger = logging.getLogger(self.__class__.__module__)
self.lm = lm
self.graph = graph
self.prompter = prompter
self.parser = parser
self.problem_parameters = problem_parameters
self.run_executed = False
def run(self) -> None:
"""
Run the controller and execute the operations from the Graph of
Operations based on their readiness.
Ensures the program is in a valid state before execution.
:raises AssertionError: If the Graph of Operation has no roots.
:raises AssertionError: If the successor of an operation is not in the Graph of Operations.
"""
self.logger.debug("Checking that the program is in a valid state")
assert self.graph.roots is not None, "The operations graph has no root"
self.logger.debug("The program is in a valid state")
execution_queue = [
operation
for operation in self.graph.operations
if operation.can_be_executed()
]
while len(execution_queue) > 0:
current_operation = execution_queue.pop(0)
self.logger.info("Executing operation %s", current_operation.operation_type)
current_operation.execute(
self.lm, self.prompter, self.parser, **self.problem_parameters
)
self.logger.info("Operation %s executed", current_operation.operation_type)
for operation in current_operation.successors:
assert (
operation in self.graph.operations
), "The successor of an operation is not in the operations graph"
if operation.can_be_executed():
execution_queue.append(operation)
self.logger.info("All operations executed")
self.run_executed = True
def get_final_thoughts(self) -> List[List[Thought]]:
"""
Retrieve the final thoughts after all operations have been executed.
:return: List of thoughts for each operation in the graph's leaves.
:rtype: List[List[Thought]]
:raises AssertionError: If the `run` method hasn't been executed yet.
"""
assert self.run_executed, "The run method has not been executed"
return [operation.get_thoughts() for operation in self.graph.leaves]
def output_graph(self, path: str) -> None:
"""
Serialize the state and results of the operations graph to a JSON file.
:param path: The path to the output file.
:type path: str
"""
output = []
for operation in self.graph.operations:
operation_serialized = {
"operation": operation.operation_type.name,
"thoughts": [thought.state for thought in operation.get_thoughts()],
}
if any([thought.scored for thought in operation.get_thoughts()]):
operation_serialized["scored"] = [
thought.scored for thought in operation.get_thoughts()
]
operation_serialized["scores"] = [
thought.score for thought in operation.get_thoughts()
]
if any([thought.validated for thought in operation.get_thoughts()]):
operation_serialized["validated"] = [
thought.validated for thought in operation.get_thoughts()
]
operation_serialized["validity"] = [
thought.valid for thought in operation.get_thoughts()
]
if any(
[
thought.compared_to_ground_truth
for thought in operation.get_thoughts()
]
):
operation_serialized["compared_to_ground_truth"] = [
thought.compared_to_ground_truth
for thought in operation.get_thoughts()
]
operation_serialized["problem_solved"] = [
thought.solved for thought in operation.get_thoughts()
]
output.append(operation_serialized)
output.append(
{
"prompt_tokens": self.lm.prompt_tokens,
"completion_tokens": self.lm.completion_tokens,
"cost": self.lm.cost,
}
)
with open(path, "w") as file:
file.write(json.dumps(output, indent=2))
graph_of_thoughts/controller/llamachat_hf.py
+119
View File
@@ -0,0 +1,119 @@
# Copyright (c) 2023 ETH Zurich.
# All rights reserved.
#
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
#
# main author: Ales Kubicek
import os
import torch
import transformers
from typing import List, Dict, Union
from .abstract_language_model import AbstractLanguageModel
class Llama2HF(AbstractLanguageModel):
"""
An interface to use LLaMA 2 models through the HuggingFace library.
"""
def __init__(
self, config_path: str = "", model_name: str = "llama7b-hf", cache: bool = False
) -> None:
"""
Initialize an instance of the Llama2HF class with configuration, model details, and caching options.
:param config_path: Path to the configuration file. Defaults to an empty string.
:type config_path: str
:param model_name: Specifies the name of the LLaMA model variant. Defaults to "llama7b-hf".
Used to select the correct configuration.
:type model_name: str
:param cache: Flag to determine whether to cache responses. Defaults to False.
:type cache: bool
"""
super().__init__(config_path, model_name, cache)
self.config: Dict = self.config[model_name]
# Detailed id of the used model.
self.model_id: str = self.config["model_id"]
# Costs for 1000 tokens.
self.prompt_token_cost: float = self.config["prompt_token_cost"]
self.response_token_cost: float = self.config["response_token_cost"]
# The temperature is defined as the randomness of the model's output.
self.temperature: float = self.config["temperature"]
# Top K sampling.
self.top_k: int = self.config["top_k"]
# The maximum number of tokens to generate in the chat completion.
self.max_tokens: int = self.config["max_tokens"]
# Important: must be done before importing transformers
os.environ["TRANSFORMERS_CACHE"] = self.config["cache_dir"]
hf_model_id = f"meta-llama/{self.model_id}"
model_config = transformers.AutoConfig.from_pretrained(hf_model_id)
bnb_config = transformers.BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16,
)
self.tokenizer = transformers.AutoTokenizer.from_pretrained(hf_model_id)
self.model = transformers.AutoModelForCausalLM.from_pretrained(
hf_model_id,
trust_remote_code=True,
config=model_config,
quantization_config=bnb_config,
device_map="auto",
)
self.model.eval()
torch.no_grad()
self.generate_text = transformers.pipeline(
model=self.model, tokenizer=self.tokenizer, task="text-generation"
)
def query(self, query: str, num_responses: int = 1) -> List[Dict]:
"""
Query the LLaMA 2 model for responses.
:param query: The query to be posed to the language model.
:type query: str
:param num_responses: Number of desired responses, default is 1.
:type num_responses: int
:return: Response(s) from the LLaMA 2 model.
:rtype: List[Dict]
"""
if self.cache and query in self.respone_cache:
return self.respone_cache[query]
sequences = []
query = f"<s><<SYS>>You are a helpful assistant. Always follow the intstructions precisely and output the response exactly in the requested format.<</SYS>>\n\n[INST] {query} [/INST]"
for _ in range(num_responses):
sequences.extend(
self.generate_text(
query,
do_sample=True,
top_k=self.top_k,
num_return_sequences=1,
eos_token_id=self.tokenizer.eos_token_id,
max_length=self.max_tokens,
)
)
response = [
{"generated_text": sequence["generated_text"][len(query) :].strip()}
for sequence in sequences
]
if self.cache:
self.respone_cache[query] = response
return response
def get_response_texts(self, query_responses: List[Dict]) -> List[str]:
"""
Extract the response texts from the query response.
:param query_responses: The response list of dictionaries generated from the `query` method.
:type query_responses: List[Dict]
:return: List of response strings.
:rtype: List[str]
"""
return [query_response["generated_text"] for query_response in query_responses]