model_config.py

from cmdline import args

MODEL_HOST = args.model_host
# MODEL_HOST = "huggingface"
# MODEL_HOST = "openai"

# Pick an OpenAI model:
OPENAI_BASE_URL = "http://localhost:11434/v1/"
OPENAI_API_KEY = "ollama"
OPENAI_MODEL = "phind-codellama"
# OPENAI_MODEL = "gpt-4"

# Pick a base model hosted on HuggingFace:
BASE_MODEL_NAME = args.base_model_name
# BASE_MODEL_NAME = "Phind/Phind-CodeLlama-34B-v2"
# BASE_MODEL_NAME = "codellama/CodeLlama-34b-hf"
# BASE_MODEL_NAME = "codellama/CodeLlama-34b-Instruct-hf"
# BASE_MODEL_NAME = "deepseek-ai/deepseek-coder-33b-base"
# BASE_MODEL_NAME = "deepseek-ai/deepseek-coder-33b-instruct"
# BASE_MODEL_NAME = "mistralai/Mixtral-8x7B-Instruct-v0.1"
# BASE_MODEL_NAME = "./my_dpo_model"

# Provide an optional checkpoint on top of the base model:
PEFT_MODEL_PATH = args.peft_model_path
# PEFT_MODEL_PATH = None

# Or pick an entire PPO model -- overrides all the above:
PPO_MODEL_PATH = args.ppo_model_path
# PPO_MODEL_PATH = "./my_ppo_model"

# Set to None for run.py to instead use base model (and optional peft):
# PPO_MODEL_PATH = None

# Set to True to use custom stop words:
CUSTOM_STOP = args.custom_stop
# CUSTOM_STOP = True

# Sample the same whether generating one or many samples
SAME_FOR_MANY_SAMPLES = args.same_for_many_samples
# SAME_FOR_MANY_SAMPLES = True

# Set to True to use beam search instead of sampling
BEAM_SEARCH = args.beam_search
# BEAM_SEARCH = False

# Can set to None
MODEL_ARG_TOP_K = args.model_arg_topk
MODEL_ARG_TOP_P = args.model_arg_topp
MODEL_ARG_TEMP = args.model_arg_temp
# MODEL_ARG_TOP_K = 0.0
# MODEL_ARG_TOP_P = 1.0
# MODEL_ARG_TEMP = 1.0
# MODEL_ARG_TOP_K = None
# MODEL_ARG_TOP_P = None
# MODEL_ARG_TEMP = None