Closed collinarnett closed 4 months ago
One more thing. Some of the executables installed from LLM-VM don't work due to import errors.
File "/nix/store/02w0r9mqq5z19gmzxf7qgcp8m4cnlypc-python3.11-llm_vm-0.1.55/lib/python3.11/site-packages/llm_vm/agents/FLAT/models/utils/tool_picker_model/tool_picker_model_data.py", line 25, in __get_toolpicker_model
json_data: List[QuestionSplitModelJSONData] = json.load(open(file_path, "r"))["data"]
^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: [Errno 2] No such file or directory: '/nix/store/02w0r9mqq5z19gmzxf7qgcp8m4cnlypc-python3.11-llm_vm-0.1.55/lib/python3.11/site-packages/llm_vm/agents/FLAT/models/raw_data/toolpicker_data.json'
I imagine this issue is limited to the agents since they are run relative to the project root in the README.
Don't want to work on this anymore and this repo seems dead.
This PR adds Nix as a method of installing and running LLM-VM. With this PR, as long as the flake experimental feature is enabled with Nix, users can easily run the LLM-VM server with:
Users can also run the server with cuda support baked in:
(the impure is needed here in order for nixGL to find the host nvidia drivers)
Developers also benefit from this PR with the ability to start a shell with a python environment including LLM-VM.
From here tests can be run by starting the server with the first command then running the tests in the developer shell.
In total, the following outputs are provided.
I would be grateful if others could review this PR by installing nix via the The Determinate Nix Installer and running the commands I included above. I do not have a rocm enabled video card so I would appreciate for those that do to leave their feedback.
Also let me know your general experience. There may be things I can do to make things better.