Run LLMs locally on Cloud Workstations. Uses:
- Quantized models from 🤗
- llama-cpp-python's webserver
In this guide:
Running as a Cloud Workstation
This repository includes a Dockerfile that can be used to create a custom base image
for a Cloud Workstation environment that includes the llm tool.
To get started, you'll need to have a GCP Project and have the gcloud CLI installed.
-
Set environment variables
-
Set the
PROJECT_IDandPROJECT_NUMenvironment variables from your GCP project. You must modify the values.export PROJECT_ID=<project-id> export PROJECT_NUM=<project-num>
-
Set other needed environment variables. You can modify the values.
export REGION=us-central1 export LOCALLLM_REGISTRY=localllm-registry export LOCALLLM_IMAGE_NAME=localllm export LOCALLLM_CLUSTER=localllm-cluster export LOCALLLM_WORKSTATION=localllm-workstation export LOCALLLM_PORT=8000
-
-
Set the default project.
gcloud config set project $PROJECT_ID
-
Enable needed services.
gcloud services enable \ cloudbuild.googleapis.com \ workstations.googleapis.com \ container.googleapis.com \ containeranalysis.googleapis.com \ containerscanning.googleapis.com \ artifactregistry.googleapis.com -
Create an Artifact Registry repository for docker images.
gcloud artifacts repositories create $LOCALLLM_REGISTRY \ --location=$REGION \ --repository-format=docker
-
Build and push the image to Artifact Registry using Cloud Build. Details are in cloudbuild.yaml.
gcloud builds submit . \ --substitutions=_IMAGE_REGISTRY=$LOCALLLM_REGISTRY,_IMAGE_NAME=$LOCALLLM_IMAGE_NAME
-
Configure a Cloud Workstation cluster.
gcloud workstations clusters create $LOCALLLM_CLUSTER \ --region=$REGION
-
Create a Cloud Workstation configuration. We suggest using a machine type of e2-standard-32 which has 32 vCPU, 16 core and 128 GB memory.
gcloud workstations configs create $LOCALLLM_WORKSTATION \ --region=$REGION \ --cluster=$LOCALLLM_CLUSTER \ --machine-type=e2-standard-32 \ --container-custom-image=us-central1-docker.pkg.dev/${PROJECT_ID}/${LOCALLLM_REGISTRY}/${LOCALLLM_IMAGE_NAME}:latest
-
Create a Cloud Workstation.
gcloud workstations create $LOCALLLM_WORKSTATION \ --cluster=$LOCALLLM_CLUSTER \ --config=$LOCALLLM_WORKSTATION \ --region=$REGION
-
Grant access to the default Cloud Workstation service account.
gcloud artifacts repositories add-iam-policy-binding $LOCALLLM_REGISTRY \ --location=$REGION \ --member=serviceAccount:service-$PROJECT_NUM@gcp-sa-workstationsvm.iam.gserviceaccount.com \ --role=roles/artifactregistry.reader
-
Start the workstation.
gcloud workstations start $LOCALLLM_WORKSTATION \ --cluster=$LOCALLLM_CLUSTER \ --config=$LOCALLLM_WORKSTATION \ --region=$REGION
-
Connect to the workstation using ssh. Alternatively, you can connect to the workstation interactively in the browser.
gcloud workstations ssh $LOCALLLM_WORKSTATION \ --cluster=$LOCALLLM_CLUSTER \ --config=$LOCALLLM_WORKSTATION \ --region=$REGION
-
Start serving the default model from the repo.
llm run TheBloke/Llama-2-13B-Ensemble-v5-GGUF $LOCALLLM_PORT -
Get the hostname of the workstation using:
gcloud workstations describe $LOCALLLM_WORKSTATION \ --cluster=$LOCALLLM_CLUSTER \ --config=$LOCALLLM_WORKSTATION \ --region=$REGION
-
Interact with the model by visiting the live OpenAPI documentation page:
https://$LOCALLLM_PORT-$LLM_HOSTNAME/docs.
Assumes that models are downloaded to ~/.cache/huggingface/hub/. This is the default cache
path used by Hugging Face Hub library and only supports .gguf files.
If you're using models from TheBloke and you don't specify a filename, we'll attempt to use the model with 4 bit medium quantization, or you can specify a filename explicitly.
-
List downloaded models.
llm list
-
List running models.
llm ps
-
Start serving models.
-
Start serving the default model from the repo. Download if not present.
llm run TheBloke/Llama-2-13B-Ensemble-v5-GGUF 8000
-
Start serving a specific model. Download if not present.
llm run TheBloke/Llama-2-13B-Ensemble-v5-GGUF --filename llama-2-13b-ensemble-v5.Q4_K_S.gguf 8000
-
-
Stop serving models.
-
Stop serving all models from the repo.
llm kill TheBloke/Llama-2-13B-Ensemble-v5-GGUF -
Stop serving a specific model.
llm kill TheBloke/Llama-2-13B-Ensemble-v5-GGUF --filename llama-2-13b-ensemble-v5.Q4_K_S.gguf
-
-
Download models.
- Download the default model from the repo.
llm pull TheBloke/Llama-2-13B-Ensemble-v5-GGUF
- Download a specific model from the repo.
llm pull TheBloke/Llama-2-13B-Ensemble-v5-GGUF --filename llama-2-13b-ensemble-v5.Q4_K_S.gguf
- Download the default model from the repo.
-
Remove models.
-
Remove all models downloaded from the repo.
llm rm TheBloke/Llama-2-13B-Ensemble-v5-GGUF
-
Remove a specific model from the repo.
llm rm TheBloke/Llama-2-13B-Ensemble-v5-GGUF --filename llama-2-13b-ensemble-v5.Q4_K_S.gguf
-
-
Install the tools.
# Install the tools pip3 install openai pip3 install ./llm-tool/. -
Download and run a model.
llm run TheBloke/Llama-2-13B-Ensemble-v5-GGUF 8000
-
Try out a query. The default query is for a haiku about cats.
python3 querylocal.py
-
Interact with the Open API interface via the
/docsextension. For the above, visit http://localhost:8000/docs.
To assist with debugging, you can configure model startup to write logs to a log file by providing a yaml python logging configuration file (for example see llm_log_config.yaml):
llm run TheBloke/Llama-2-13B-Ensemble-v5-GGUF 8000 --log-config llm_log_config.yaml
If running from Cloud Workstations, logs from running models will be written to /var/log/locallm.log
(llm_log_config.yaml is provided by default via the environment variable
LOG_CONFIG within the image). You can follow the logs with:
tail -f /var/log/localllm.log
To run locally using the bundled log config:
sudo touch /var/log/localllm.log
sudo chown user:user /var/log/localllm.log # use your user and group
export LOG_CONFIG=$(pip show llm | grep Location | awk '{print $2}')/llm_log_config.yaml
llm run ...
If you are running multiple models, the logs from each will be written to the same file and interleaved.
This project imports freely available LLMs and makes them available from Cloud Workstations. We recommend independently verifying any content generated by the models. We do not assume any responsibility or liability for the use or interpretation of generated content.