Document how to use specific LLMs

[rbren]

What problem or use case are you trying to solve? Lots of folks are struggling to get OpenDevin working with non-OpenAI models. Local ollama seems to be particularly hard

Describe the UX of the solution you'd like We should have a doc that lists out 3-4 major providers, explains how to get an API key, and how to configure OpenDevin

Do you have thoughts on the technical implementation? Just a Models.md that we can link to from README.md

[gaborkukucska]

+1 for ollama support and documentation.

I get ModuleNotFoundError: No module named 'llama_index.embeddings.ollama' when I use LLM_EMBEDDING_MODEL="llama2" and ModuleNotFoundError: No module named 'llama_index.embeddings.huggingface' when I use LLM_EMBEDDING_MODEL="local"

[enyst]

Do you have thoughts on the technical implementation? Just a Models.md that we can link to from README.md

I see github still has the wiki feature. I don't see it on opendevin, but IIRC that's because it would need enabled for the project. What if we use that for a few / several model-specific pages? I don't remember if access can be enabled widely but it's a wiki, I'd assume so.

[rbren]

That seems reasonable to me

[DanRegalia]

Can you update the supported tags for the .env with commands to connect to open llms, openAI/OllamaWebUI/Oogabooga/ETC?

[Dampfinchen]

IMO, Ollama can be a nuisance for people who already have gguf files because it requires GGUFs to be converted. It also has no GUI, in contrast to koboldcpp and Ooba.

I recommend using llama.cpp server, koboldcpp and Ooba instead. Those are really easy to use inference programs.

[DGdev91]

Agreed. I also made a PR for setting LLM_BASE_URL with make setup-config. Should be enough for any OpenAI-compatible https://github.com/OpenDevin/OpenDevin/pull/616

For example, it works with Koboldcpp by setting it like this: LLM_API_KEY="." LLM_MODEL="gpt-4-0125-preview" LLM_BASE_URL="http://localhost:5001/v1/" WORKSPACE_DIR="./workspace"

[JayQuimby]

I made a markdown file for Ollama #615.

[feuler]

I use oobabooga with the openai compatible api. And currently trying OpenCodeInterpreter-DS-33B (exl2) as model first. I'm not aware which non-openai/claude model works best yet...

config.toml: WORKSPACE_DIR="./workspace" LLM_API_KEY="sk-111111111111111111111111111111111111111111111111" LLM_BASE_URL="http://127.0.0.1:5000" LLM_EMBEDDING_MODEL="local" LLM_MODEL="oobabooga/OpenCodeInterpreter-DS-33B-4.65bpw-h6-exl2"

After starting a development i currently get these kind of errors sometimes during the process:

ERROR: 'str' object has no attribute 'copy' Traceback (most recent call last): File "/opt/llm/OpenDevin/opendevin/controller/agent_controller.py", line 113, in step action = self.agent.step(self.state) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/agent.py", line 166, in step action = prompts.parse_action_response(action_resp) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/utils/prompts.py", line 135, in parse_action_response return action_from_dict(action_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/opendevin/action/init.py", line 24, in action_from_dict action = action.copy() ^^^^^^^^^^^ AttributeError: 'str' object has no attribute 'copy'

############### ERROR: "'action' key is not found in action={}" Traceback (most recent call last): File "/opt/llm/OpenDevin/opendevin/controller/agent_controller.py", line 113, in step action = self.agent.step(self.state) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/agent.py", line 166, in step action = prompts.parse_action_response(action_resp) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/utils/prompts.py", line 135, in parse_action_response return action_from_dict(action_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/opendevin/action/init.py", line 26, in action_from_dict raise KeyError(f"'action' key is not found in {action=}") KeyError: "'action' key is not found in action={}"

[stratte89]

Can you update the supported tags for the .env with commands to connect to open llms, openAI/OllamaWebUI/Oogabooga/ETC?

here is a guide for oobabooga webui https://github.com/OpenDevin/OpenDevin/commit/08a2dfb01af1aec6743f5e4c23507d63980726c0#commitcomment-140559598

[rmortes]

I use oobabooga with the openai compatible api. And currently trying OpenCodeInterpreter-DS-33B (exl2) as model first. I'm not aware which non-openai/claude model works best yet...

config.toml: WORKSPACE_DIR="./workspace" LLM_API_KEY="sk-111111111111111111111111111111111111111111111111" LLM_BASE_URL="http://127.0.0.1:5000" LLM_EMBEDDING_MODEL="local" LLM_MODEL="oobabooga/OpenCodeInterpreter-DS-33B-4.65bpw-h6-exl2"

After starting a development i currently get these kind of errors sometimes during the process:

ERROR: 'str' object has no attribute 'copy' Traceback (most recent call last): File "/opt/llm/OpenDevin/opendevin/controller/agent_controller.py", line 113, in step action = self.agent.step(self.state) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/agent.py", line 166, in step action = prompts.parse_action_response(action_resp) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/utils/prompts.py", line 135, in parse_action_response return action_from_dict(action_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/opendevin/action/init.py", line 24, in action_from_dict action = action.copy() ^^^^^^^^^^^ AttributeError: 'str' object has no attribute 'copy'

############### ERROR: "'action' key is not found in action={}" Traceback (most recent call last): File "/opt/llm/OpenDevin/opendevin/controller/agent_controller.py", line 113, in step action = self.agent.step(self.state) ^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/agent.py", line 166, in step action = prompts.parse_action_response(action_resp) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/agenthub/monologue_agent/utils/prompts.py", line 135, in parse_action_response return action_from_dict(action_dict) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/llm/OpenDevin/opendevin/action/init.py", line 26, in action_from_dict raise KeyError(f"'action' key is not found in {action=}") KeyError: "'action' key is not found in action={}"

Getting a similar error using the following config in MacOS

config.toml

LLM_MODEL="ollama/deepseek-coder:instruct"
LLM_API_KEY="ollama"
LLM_EMBEDDING_MODEL="local"
LLM_BASE_URL="http://localhost:11434"
WORKSPACE_DIR="./workspace"

Oops. Something went wrong: 'str' object has no attribute 'copy'
Oops. Something went wrong: "'action['action']=None' is not defined. Available actions: dict_keys(['kill', 'run', 'browse', 'read', 'write', 'recall', 'think', 'finish', 'add_task', 'modify_task'])"

[stratte89]

hey, use openai/modelname like

LLM_API_KEY="na" LLM_BASE_URL="http://0.0.0.0:5000/v1" LLM_MODEL="openai/alpindale_Mistral-7B-v0.2-hf" LLM_EMBEDDING_MODEL="lokal" WORKSPACE_DIR="./workspace" MAX_ITERATIONS=20000

although i dont think that the model name really matter for oobabooga as long as its openai/ it can be anything, at least it works for me, i can switch model without changing the config and it's working.

[rbren]

@Rags100 this isn't really the right place for your question. Please follow the README, and search our issues for related problems (there's an existing one for uvloop--you'll need WSL to make it work).

If you continue to have trouble, feel free to file a new issue with the template filled out. Thanks!

[rbren]

Hey all--lots of unrelated comments in this thread. Please try to keep this about LLM Model documentation.

I'm going to delete the unrelated comments--feel free to open new issues if you're having trouble

[enyst]

Documentation for Azure: https://github.com/OpenDevin/OpenDevin/pull/1035

[larkinwc]

Added documentation for using Google's Gemini model through AI studio as well as VertexAI through GCP #1321

[rbren]

We've made a lot of progress on this one, so I'm going to close it. More docs welcome!