launch-computation-example / process_runs.py
meg-huggingface
Addressing the need to update the requests dataset to COMPLETED, FAILED, etc. based on model *and task*.
7de9b42
import argparse
import os
from datasets import load_dataset, Dataset
from huggingface_hub import HfApi
TOKEN = os.environ.get("DEBUG")
api = HfApi(token=TOKEN)
REQUESTS_DSET = "AIEnergyScore/requests_debug"
RESULTS_DSET = "AIEnergyScore/results_debug"
PENDING = 'PENDING'
COMPLETED = 'COMPLETED'
FAILED = 'FAILED'
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--run_dir",
default="/runs",
type=str,
required=False,
help="Path to the run directory.",
)
parser.add_argument(
"--attempts",
default="/attempts.txt",
type=str,
required=False,
help="File with per-line run attempt directories. Assumes format '/runs/{task}/{model}/{timestamp}'",
)
parser.add_argument(
"--failed_attempts",
default="/failed_attempts.txt",
type=str,
required=False,
help="File with per-line failed run directories. Assumes format '/runs/{task}/{model}/{timestamp}'",
)
args = parser.parse_args()
return args
def check_for_traceback(run_dir):
# run_dir="./runs/${experiment_name}/${backend_model}/${now}"
found_error = False
error_message = ""
try:
# Read error message
with open(f"{run_dir}/error.log", 'r') as f:
# There may be a better way to do this that finds the
# index of Traceback, then prints from there : end-of-file index (the file length-1).
for line in f:
# Question: Do we even need to check for this? The presence of the
# error file, or at least a non-empty one,
# means there's been an error, no?
if 'Traceback (most recent call last):' in line:
found_error = True
if found_error:
error_message += line
except FileNotFoundError as e:
# When does this happen?
print(f"Could not find {run_dir}/error.log")
return error_message
def update_requests(requests, all_attempts, failed_attempts):
"""
Sets All PENDING requests with the given model & task to 'COMPLETED' or 'FAILED.'
Reads in the all_attempts text file and failed_attempts text file, in which
each line is a run directory run_dir="/runs/${experiment_name}/${backend_model}/${now}"
:param requests: requests Dataset
:param all_attempts: text file of the run directories of each task/model/timestamp
:param failed_attempts: text file of the run directories of each task/model/timestamp
:return:
"""
requests_df = requests.to_pandas()
# Each line is a run directory, where
# run_dir="/runs/${experiment_name}/${backend_model}/${now}", where
# ${backend_model} is ${organization}/${model_name}
for line in all_attempts:
line = line.strip()
print(f"Checking {line}")
split_run_dir = line.strip().strip("/").split("/")
print(f"Processing run directory {split_run_dir}")
task = split_run_dir[1]
print(f"Task is {task}")
# The naming of the optimum benchmark configs uses an underscore.
# The naming of the HF Api list models function uses a hyphen.
# We therefore need to adapt this task string name depending on
# which part of our pipeline we're talking to.
hyphenated_task_name = "-".join(task.split("_"))
model = "/".join([split_run_dir[2], split_run_dir[3]])
print(f"Model is {model}")
traceback_error = check_for_traceback(line)
if traceback_error != "":
print("Found a traceback error!")
print(traceback_error)
requests_df.loc[(requests_df["status"] == PENDING) & (requests_df["model"] == model) & (requests_df["task"] == hyphenated_task_name), ['status']] = FAILED
requests_df.loc[(requests_df["status"] == PENDING) & (requests_df["model"] == model) & (requests_df["task"] == hyphenated_task_name), ['error_message']] = traceback_error
elif line in failed_attempts:
print(f"Job failed, but not sure why -- didn't find a traceback in {line}.")
print(f"Setting {model}, {hyphenated_task_name}, status {PENDING} to {FAILED}.")
print(requests_df[(requests_df["status"] == PENDING) & (requests_df["model"] == model) & (requests_df["task"] == hyphenated_task_name)])
requests_df.loc[(requests_df["status"] == PENDING) & (requests_df["model"] == model) & (requests_df["task"] == hyphenated_task_name), ['status']] = FAILED
else:
requests_df.loc[(requests_df["status"] == PENDING) & (requests_df["model"] == model) & (requests_df["task"] == hyphenated_task_name), ['status']] = COMPLETED
updated_dset = Dataset.from_pandas(requests_df)
return updated_dset
if __name__ == '__main__':
args = parse_args()
# Uploads all run output to the results dataset.
print(f"Uploading {args.run_dir} to {RESULTS_DSET}")
api.upload_folder(
folder_path=args.run_dir,
repo_id=f"{RESULTS_DSET}",
repo_type="dataset",
)
# Update requests dataset based on whether things have failed or not.
print(f"Examining the run directory for each model & task to determine if it {FAILED} or {COMPLETED}.")
requests = load_dataset(f"{REQUESTS_DSET}", split="test", token=TOKEN)
all_attempts = open(f"{args.attempts}", "r+").readlines()
failed_attempts = open(f"{args.failed_attempts}", "r+").readlines()
updated_requests = update_requests(requests, all_attempts, failed_attempts)
print(f"Uploading updated {REQUESTS_DSET}.")
updated_requests.push_to_hub(f"{REQUESTS_DSET}", split="test", token=TOKEN)
print("Done.")