Local Models integration

[TheWhiteWord]

Hey Devs,
let me start by saying that this programme is great. Well done on your work, and thanks for sharing it.

My question is: is there any plan to allow for the integration of local models?
Even a just section in the documentation would be great.

Have a good day
theWW

[andraz]

+1 to this question

It makes no sense to shovel money into some closed source while we have a powerful GPU that can run 13b Llama with no problem with some of the other open source projects.

[thedualspace]

I'd also be very eager to use local models with ChatDev, Llama based models show great promise

[j-loquat]

Local model use and perhaps as a more advanced feature, assign different models to different agents in the company - so could use a local python-optimized model for an engineer, and a llama2 model for the CEO, etc.

[TheWhiteWord]

@j-loquat I love that idea. That is a thing i was considering more and more. Ai becoming more and more like greek gods, each with its characther and function that complete each other.. it was the original vision of Altman too, kind of, but they lost their way

[andraz]

No need to have 1 "god AGI" (which can not be ran locally as it demands crazy hardware) if we can have 20 agents with 20 different local narrow AI models that can be loaded one after another.

[TheWhiteWord]

Oh god, sorry Devs but this conversation is too interesting. You may need to turn notifications off XD

I was trained as an artist, and the first thing to know is that limitations are the generator of creativity. A big Ai with all the knowledge of the world may just become the most boring thing to touch the planet. And this may be controversial, but I think that bad qualities are needed too...everything has its meaning and use in order to create balance. Just my opinion

[hemangjoshi37a]

This has been referenced in #33

[starkdmi]

1. Install LocalAI - OpenAI compatible server.
2. Create new model config file named gpt-3.5-turbo-16k.yaml and set the model name to gpt-3.5-turbo-16k-0613.
3. Start LocalAI server locally and run:

OPENAI_API_BASE=http://127.0.0.1:8000/v1 OPENAI_API_KEY="dummy" python run.py --task "Snake game in pure html" --name "WebSnake"

[andraz]

The command above did not work in Anaconda Prompt, but this version did:

(chatdev_conda_env) C:\chatdev>set OPENAI_API_BASE=http://127.0.0.1:5001/v1

(chatdev_conda_env) C:\chatdev>set OPENAI_API_KEY=123456

(chatdev_conda_env) C:\chatdev>python run.py --task "Hello world in python" --name "HelloWorld"
**[Preprocessing]**

**ChatDev Starts** (20230913191808)

**Timestamp**: 20230913191808

**config_path**: C:\chatdev\CompanyConfig\Default\ChatChainConfig.json

**config_phase_path**: C:\chatdev\CompanyConfig\Default\PhaseConfig.json

**config_role_path**: C:\chatdev\CompanyConfig\Default\RoleConfig.json

**task_prompt**: Hello world in python

**project_name**: HelloWorld

**Log File**: C:\chatdev\WareHouse\HelloWorld_DefaultOrganization_20230913191808.log

**ChatDevConfig**:
 ChatEnvConfig.clear_structure: True
ChatEnvConfig.brainstorming: False


**ChatGPTConfig**:
 ChatGPTConfig(temperature=0.2, top_p=1.0, n=1, stream=False, stop=None, max_tokens=None, presence_penalty=0.0, frequency_penalty=0.0, logit_bias={}, user='')

I am having a problem using it with local api:

It looks like all that the API returns is 1 token:

Text-generation-webui side:

llm_load_print_meta: model size     = 13.02 B
llm_load_print_meta: general.name   = openassistant_llama2-13b-orca-8k-3319
llm_load_print_meta: BOS token = 1 '<s>'
llm_load_print_meta: EOS token = 2 '</s>'
llm_load_print_meta: UNK token = 0 '<unk>'
llm_load_print_meta: LF token  = 13 '<0x0A>'
llm_load_tensors: ggml ctx size =    0.12 MB
llm_load_tensors: using CUDA for GPU acceleration
llm_load_tensors: mem required  =  128.35 MB (+ 1600.00 MB per state)
llm_load_tensors: offloading 40 repeating layers to GPU
llm_load_tensors: offloading non-repeating layers to GPU
llm_load_tensors: offloading v cache to GPU
llm_load_tensors: offloading k cache to GPU
llm_load_tensors: offloaded 43/43 layers to GPU
llm_load_tensors: VRAM used: 11656 MB
...................................................................................................
llama_new_context_with_model: kv self size  = 1600.00 MB
llama_new_context_with_model: compute buffer total size =  191.47 MB
llama_new_context_with_model: VRAM scratch buffer: 190.00 MB
AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 0 | VSX = 0 |
2023-09-13 19:11:27 INFO:Loaded the model in 7.52 seconds.

Warning: $This model maximum context length is 2048 tokens. However, your messages resulted in over 498 tokens and max_tokens is 15937.

llama_print_timings:        load time =   955.37 ms
llama_print_timings:      sample time =     0.23 ms /     1 runs   (    0.23 ms per token,  4424.78 tokens per second)
llama_print_timings: prompt eval time =   955.31 ms /   498 tokens (    1.92 ms per token,   521.30 tokens per second)
llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time =   957.88 ms
Output generated in 1.41 seconds (0.00 tokens/s, 0 tokens, context 498, seed 1828391196)
127.0.0.1 - - [13/Sep/2023 19:11:42] "POST /v1/chat/completions HTTP/1.1" 200 -
Warning: $This model maximum context length is 2048 tokens. However, your messages resulted in over 551 tokens and max_tokens is 15885.
Llama.generate: prefix-match hit

llama_print_timings:        load time =   955.37 ms
llama_print_timings:      sample time =     0.16 ms /     1 runs   (    0.16 ms per token,  6410.26 tokens per second)
llama_print_timings: prompt eval time =   835.88 ms /   489 tokens (    1.71 ms per token,   585.01 tokens per second)
llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time =   836.73 ms
Output generated in 1.24 seconds (0.00 tokens/s, 0 tokens, context 551, seed 192786861)
127.0.0.1 - - [13/Sep/2023 19:11:46] "POST /v1/chat/completions HTTP/1.1" 200 -
Warning: $This model maximum context length is 2048 tokens. However, your messages resulted in over 521 tokens and max_tokens is 15907.
Llama.generate: prefix-match hit

llama_print_timings:        load time =   955.37 ms
llama_print_timings:      sample time =     0.13 ms /     1 runs   (    0.13 ms per token,  7633.59 tokens per second)
llama_print_timings: prompt eval time =   884.39 ms /   459 tokens (    1.93 ms per token,   519.00 tokens per second)
llama_print_timings:        eval time =     0.00 ms /     1 runs   (    0.00 ms per token,      inf tokens per second)
llama_print_timings:       total time =   885.63 ms
Output generated in 1.26 seconds (0.00 tokens/s, 0 tokens, context 521, seed 1288396660)
127.0.0.1 - - [13/Sep/2023 19:11:53] "POST /v1/chat/completions HTTP/1.1" 200 -
Warning: $This model maximum context length is 2048 tokens. However, your messages resulted in over 574 tokens and max_tokens is 15854.
Llama.generate: prefix-match hit

ChatDev side

Chief Executive Officer: **Chief Product Officer<->Chief Executive Officer on : DemandAnalysis, turn 0**

[ChatDev is a software company powered by multiple intelligent agents, such as chief executive officer, chief human resources officer, chief product officer, chief technology officer, etc, with a multi-agent organizational structure and the mission of "changing the digital world through programming".
You are Chief Product Officer. we are both working at ChatDev. We share a common interest in collaborating to successfully complete a task assigned by a new customer.
You are responsible for all product-related matters in ChatDev. Usually includes product design, product strategy, product vision, product innovation, project management and product marketing.
Here is a new customer's task: Hello world in python.
To complete the task, you must write a response that appropriately solves the requested instruction based on your expertise and customer's needs.]



**[OpenAI_Usage_Info Receive]**
prompt_tokens: 521
completion_tokens: 1
total_tokens: 522


**[OpenAI_Usage_Info Receive]**
prompt_tokens: 574
completion_tokens: 1
total_tokens: 575


Chief Product Officer: **Chief Product Officer<->Chief Executive Officer on : DemandAnalysis, turn 1**

[ChatDev is a software company powered by multiple intelligent agents, such as chief executive officer, chief human resources officer, chief product officer, chief technology officer, etc, with a multi-agent organizational structure and the mission of "changing the digital world through programming".
You are Chief Executive Officer. Now, we are both working at ChatDev and we share a common interest in collaborating to successfully complete a task assigned by a new customer.
Your main responsibilities include being an active decision-maker on users' demands and other key policy issues, leader, manager, and executor. Your decision-making role involves high-level decisions about policy and strategy; and your communicator role can involve speaking to the organization's management and employees.
Here is a new customer's task: Hello world in python.
To complete the task, I will give you one or more instructions, and you must help me to write a specific solution that appropriately solves the requested instruction based on your expertise and my needs.]



Chief Executive Officer: **Chief Product Officer<->Chief Executive Officer on : DemandAnalysis, turn 1**

[ChatDev is a software company powered by multiple intelligent agents, such as chief executive officer, chief human resources officer, chief product officer, chief technology officer, etc, with a multi-agent organizational structure and the mission of "changing the digital world through programming".
You are Chief Product Officer. we are both working at ChatDev. We share a common interest in collaborating to successfully complete a task assigned by a new customer.
You are responsible for all product-related matters in ChatDev. Usually includes product design, product strategy, product vision, product innovation, project management and product marketing.
Here is a new customer's task: Hello world in python.
To complete the task, you must write a response that appropriately solves the requested instruction based on your expertise and customer's needs.]



**[OpenAI_Usage_Info Receive]**
prompt_tokens: 544
completion_tokens: 1
total_tokens: 545

[starkdmi]

Yeah, the command above was for macOS, no troubles with conda environment here.

@andraz, why don't you increase the context to 4K or 8K tokens? Based on your model name it support context up to 8K tokens.

Warning: $This model maximum context length is 2048 tokens. However, your messages resulted in over 521 tokens and max_tokens is 15907.

As for one token response I guess it's streaming feature, so you don't need to wait for a full response.

[xkaraman]

Hello there,
I am trying to use llama-2-7B version as described above.

I created a new yaml file with name gpt-3.5-turbo-16k.yaml and set the model name to gpt-3.5-turbo-16k-0613. Then on model used I downloaded and use one of the hugginface model library `llama-2*.bin' models.

I can successfully run it and receive answers to my questions as part of the returning object via curl but also says that
"usage":"prompt_tokens":0,"completion_tokens":0,"total_tokens":0}}

When then I try to run chatdev on a simple task ie python run.py --task "Hello world in python" --name "HelloWorld", chatdev prints the start up prompts and then receives no objects from the local llm with continuous empty usage logs

...
Note that we must ONLY discuss the product modality and do not discuss anything else! Once we all have expressed our opinion(s) and agree with the results of the discussion unanimously, any of us must actively terminate the discussion by replying with only one line, which starts with a single word <INFO>, followed by our final product modality without any other words, e.g., "<INFO> PowerPoint".

**[OpenAI_Usage_Info Receive]**
prompt_tokens: 0
completion_tokens: 0
total_tokens: 0


**[OpenAI_Usage_Info Receive]**
prompt_tokens: 0
completion_tokens: 0
total_tokens: 0


**[OpenAI_Usage_Info Receive]**
prompt_tokens: 0
completion_tokens: 0
total_tokens: 0

After 3 retries it crashes with the following KeyError.

Traceback (most recent call last):
  File "/home/**/miniconda3/envs/chatdev/lib/python3.9/site-packages/tenacity/__init__.py", line 382, in __call__
    result = fn(*args, **kwargs)
  File "/media/**/4TB_DATA/git/ChatDev/camel/utils.py", line 145, in wrapper
    return func(self, *args, **kwargs)
  File "/media/**/4TB_DATA/git/ChatDev/camel/agents/chat_agent.py", line 200, in step
    response["id"],
KeyError: 'id'

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/media/**/4TB_DATA/git/ChatDev/run.py", line 111, in <module>
    chat_chain.execute_chain()
  File "/media/**/4TB_DATA/git/ChatDev/chatdev/chat_chain.py", line 160, in execute_chain
    self.execute_step(phase_item)
  File "/media/**/4TB_DATA/git/ChatDev/chatdev/chat_chain.py", line 130, in execute_step
    self.chat_env = self.phases[phase].execute(self.chat_env,
  File "/media/**/4TB_DATA/git/ChatDev/chatdev/phase.py", line 292, in execute
    self.chatting(chat_env=chat_env,
  File "/media/**/4TB_DATA/git/ChatDev/chatdev/utils.py", line 77, in wrapper
    return func(*args, **kwargs)
  File "/media/**/4TB_DATA/git/ChatDev/chatdev/phase.py", line 131, in chatting
    assistant_response, user_response = role_play_session.step(input_user_msg, chat_turn_limit == 1)
  File "/media/**/4TB_DATA/git/ChatDev/camel/agents/role_playing.py", line 242, in step
    assistant_response = self.assistant_agent.step(user_msg_rst)
  File "/home/**/miniconda3/envs/chatdev/lib/python3.9/site-packages/tenacity/__init__.py", line 289, in wrapped_f
    return self(f, *args, **kw)
  File "/home/**/miniconda3/envs/chatdev/lib/python3.9/site-packages/tenacity/__init__.py", line 379, in __call__
    do = self.iter(retry_state=retry_state)
  File "/home/**/miniconda3/envs/chatdev/lib/python3.9/site-packages/tenacity/__init__.py", line 326, in iter
    raise retry_exc from fut.exception()
tenacity.RetryError: RetryError[<Future at 0x7ff680181ac0 state=finished raised KeyError>]

I have already exported OPENAI_API_BASE and OPENAI_API_KEY to the localhost otherwise it crashed.

What can I do to successfully use the local LLM?

Thanks for any help and sorry if this is the wrong place to ask it!

[GitSimply]

@starkdmi

2. Create new model config file named gpt-3.5-turbo-16k.yaml and set the model name to gpt-3.5-turbo-16k-0613.

What model are you using?