Add script for continuously pulling down newly saved models and evaluating them with the vllm runner.
You can run it in the background (via nohup/screen/tmux) with python3 gcs_eval.py, and it will continuously run this loop (checks every 10s):
find new models that have been saved in GCS_MODEL_DIR
download each new model into LOCAL_MODEL_DIR
evaluate the new model and output the relevant csv file
move the model from GCS_MODEL_DIR to GCS_MODEL_EVAL_DIR
Some things that don't work:
calling run_vllm_eval directly within the loop as initializing the LLM with vllm will call ray.init(), which can only be called once per script, and will hang on subsequent calls.
Add script for continuously pulling down newly saved models and evaluating them with the vllm runner. You can run it in the background (via nohup/screen/tmux) with
python3 gcs_eval.py
, and it will continuously run this loop (checks every 10s):GCS_MODEL_DIR
LOCAL_MODEL_DIR
GCS_MODEL_DIR
toGCS_MODEL_EVAL_DIR
Some things that don't work:
run_vllm_eval
directly within the loop as initializing the LLM withvllm
will callray.init()
, which can only be called once per script, and will hang on subsequent calls.