feat&fix: enhance error handling during environment initialization and VM allocation

This commit is contained in:
adlsdztony
2025-06-03 13:38:47 +00:00
parent e363da2fd7
commit 8d54d4302f
3 changed files with 277 additions and 95 deletions

View File

@@ -2,12 +2,15 @@
Utils and basic architecture credit to https://github.com/web-arena-x/webarena/blob/main/run.py.
"""
from __future__ import annotations
import argparse
import datetime
import json
import logging
import os
import sys
import signal
import time
from typing import List, Dict
import math
from tqdm import tqdm
@@ -16,6 +19,11 @@ import lib_run_single
from desktop_env.desktop_env import DesktopEnv
from mm_agents.openai_cua_agent import OpenAICUAAgent
# Global variables for signal handling
active_environments = []
processes = []
is_terminating = False
# import wandb
# load the environment variables from .env file
@@ -147,8 +155,39 @@ def distribute_tasks(test_all_meta: dict, num_envs: int) -> List[Dict]:
return distributed_tasks
def process_signal_handler(signum, frame, env_idx):
"""Signal handler for child processes to gracefully shut down their environments."""
logger.info(f"Process {env_idx + 1} received signal {signum}. Shutting down...")
# Get the active_environments from the caller's frame
local_vars = frame.f_locals
print(f"Local variables in process {env_idx + 1}: {local_vars}")
active_environments = local_vars.get('active_environments', [])
# Close environment in the current process context
for env in active_environments:
if env is not None:
try:
logger.info(f"Process {env_idx + 1} closing environment...")
env.close()
logger.info(f"Process {env_idx + 1} environment closed successfully")
except Exception as e:
logger.error(f"Process {env_idx + 1} error closing environment: {e}")
logger.info(f"Process {env_idx + 1} shutdown complete. Exiting.")
sys.exit(0)
def run_env_tasks(env_idx: int, env_tasks: dict, args: argparse.Namespace, shared_scores: list):
"""Run tasks for a single environment."""
# Each process has its own list of active environments
active_environments = []
env = None
# Setup signal handlers for this process too
signal.signal(signal.SIGINT, lambda signum, frame: process_signal_handler(signum, frame, env_idx))
signal.signal(signal.SIGTERM, lambda signum, frame: process_signal_handler(signum, frame, env_idx))
# ami-05e7d7bd279ea4f14
env = DesktopEnv(
path_to_vm=args.path_to_vm,
@@ -163,6 +202,7 @@ def run_env_tasks(env_idx: int, env_tasks: dict, args: argparse.Namespace, share
os_type="Ubuntu",
require_a11y_tree=args.observation_type in ["a11y_tree", "screenshot_a11y_tree", "som"],
)
active_environments.append(env)
agent = OpenAICUAAgent(
env=env,
model=args.model,
@@ -175,56 +215,114 @@ def run_env_tasks(env_idx: int, env_tasks: dict, args: argparse.Namespace, share
)
logger.info(f"Executing tasks in environment {env_idx + 1}/{args.num_envs}")
for domain in tqdm(env_tasks, desc=f"Env{env_idx+1}-Domain"):
for example_id in tqdm(env_tasks[domain], desc="Example", leave=False):
config_file = os.path.join(
args.test_config_base_dir, f"examples/{domain}/{example_id}.json"
)
with open(config_file, "r", encoding="utf-8") as f:
example = json.load(f)
logger.info(f"[Env {env_idx+1}][Domain]: {domain}")
logger.info(f"[Env {env_idx+1}][Example ID]: {example_id}")
logger.info(f"[Env {env_idx+1}][Instruction]: {example['instruction']}")
example_result_dir = os.path.join(
args.result_dir,
args.action_space,
args.observation_type,
args.model,
domain,
example_id,
)
os.makedirs(example_result_dir, exist_ok=True)
try:
lib_run_single.run_single_example_openaicua(
agent,
env,
example,
args.max_steps,
example["instruction"],
args,
example_result_dir,
shared_scores,
try:
for domain in tqdm(env_tasks, desc=f"Env{env_idx+1}-Domain"):
for example_id in tqdm(env_tasks[domain], desc="Example", leave=False):
config_file = os.path.join(
args.test_config_base_dir, f"examples/{domain}/{example_id}.json"
)
except Exception as e:
logger.error(f"Exception in Env{env_idx+1} {domain}/{example_id}: {e}")
env.controller.end_recording(
os.path.join(example_result_dir, "recording.mp4")
with open(config_file, "r", encoding="utf-8") as f:
example = json.load(f)
logger.info(f"[Env {env_idx+1}][Domain]: {domain}")
logger.info(f"[Env {env_idx+1}][Example ID]: {example_id}")
logger.info(f"[Env {env_idx+1}][Instruction]: {example['instruction']}")
example_result_dir = os.path.join(
args.result_dir,
args.action_space,
args.observation_type,
args.model,
domain,
example_id,
)
with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
f.write(
json.dumps(
{"Error": f"Time limit exceeded in {domain}/{example_id}"}
)
os.makedirs(example_result_dir, exist_ok=True)
try:
lib_run_single.run_single_example_openaicua(
agent,
env,
example,
args.max_steps,
example["instruction"],
args,
example_result_dir,
shared_scores,
)
f.write("\n")
except Exception as e:
logger.error(f"Exception in Env{env_idx+1} {domain}/{example_id}: {e}")
try:
env.controller.end_recording(
os.path.join(example_result_dir, "recording.mp4")
)
except Exception as rec_e:
logger.error(f"Failed to end recording: {rec_e}")
with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
f.write(
json.dumps(
{"Error": f"Time limit exceeded in {domain}/{example_id}"}
)
)
f.write("\n")
finally:
# This ensures the environment is closed even if there's an exception
logger.info(f"Process {env_idx + 1} cleaning up environment...")
try:
env.close()
logger.info(f"Process {env_idx + 1} environment closed successfully")
except Exception as e:
logger.error(f"Process {env_idx + 1} error during environment cleanup: {e}")
def signal_handler(signum, frame):
"""Handle termination signals (SIGINT, SIGTERM) to gracefully shutdown environments."""
global is_terminating, active_environments, processes
env.close()
# Avoid duplicate handling
if is_terminating:
return
is_terminating = True
logger.info(f"Received signal {signum}. Gracefully shutting down...")
# Close all registered environments in the main process
for env in active_environments:
try:
logger.info(f"Closing environment...")
env.close()
logger.info(f"Environment closed successfully")
except Exception as e:
logger.error(f"Error closing environment: {e}")
# Send termination signal to all child processes first
for p in processes:
if p.is_alive():
try:
logger.info(f"Sending termination signal to process {p.name}...")
p.terminate()
except Exception as e:
logger.error(f"Error sending termination signal to process: {e}")
# Allow a short time for processes to handle their own cleanup
time.sleep(1)
# Forcefully terminate any processes that didn't exit
for p in processes:
if p.is_alive():
try:
logger.info(f"Forcefully terminating process {p.name}...")
import signal
os.kill(p.pid, signal.SIGKILL)
except Exception as e:
logger.error(f"Error forcefully terminating process: {e}")
logger.info("Shutdown complete. Exiting.")
sys.exit(0)
def test(args: argparse.Namespace, test_all_meta: dict) -> None:
global processes
logger.info("Args: %s", args)
distributed_tasks = distribute_tasks(test_all_meta, args.num_envs)
@@ -244,10 +342,28 @@ def test(args: argparse.Namespace, test_all_meta: dict) -> None:
)
processes.append(p)
p.start()
logger.info(f"Started process {p.name} with PID {p.pid}")
# Wait for all processes to complete
for p in processes:
p.join()
try:
# Wait for all processes to complete
for p in processes:
p.join()
logger.info(f"Process {p.name} completed")
except KeyboardInterrupt:
logger.info("Main process received KeyboardInterrupt. Initiating graceful shutdown...")
# Let the signal handler do the cleanup
raise
except Exception as e:
logger.error(f"Unexpected error while waiting for processes: {e}", exc_info=True)
# Ensure cleanup happens
for p in processes:
if p.is_alive():
try:
logger.info(f"Terminating process {p.name} due to error...")
p.terminate()
except Exception as term_e:
logger.error(f"Error terminating process {p.name}: {term_e}")
raise
# Convert shared list to regular list
scores = list(shared_scores)
@@ -331,31 +447,76 @@ if __name__ == "__main__":
####### The complete version of the list of examples #######
os.environ["TOKENIZERS_PARALLELISM"] = "false"
args = config()
# Register signal handlers for graceful termination
signal.signal(signal.SIGINT, signal_handler) # Handle Ctrl+C
signal.signal(signal.SIGTERM, signal_handler) # Handle termination signal
try:
args = config()
with open(args.test_all_meta_path, "r", encoding="utf-8") as f:
test_all_meta = json.load(f)
with open(args.test_all_meta_path, "r", encoding="utf-8") as f:
test_all_meta = json.load(f)
if args.domain != "all":
test_all_meta = {args.domain: test_all_meta[args.domain]}
if args.domain != "all":
test_all_meta = {args.domain: test_all_meta[args.domain]}
test_file_list = get_unfinished(
args.action_space,
args.model,
args.observation_type,
args.result_dir,
test_all_meta,
)
left_info = ""
for domain in test_file_list:
left_info += f"{domain}: {len(test_file_list[domain])}\n"
logger.info(f"Left tasks:\n{left_info}")
test_file_list = get_unfinished(
args.action_space,
args.model,
args.observation_type,
args.result_dir,
test_all_meta,
)
left_info = ""
for domain in test_file_list:
left_info += f"{domain}: {len(test_file_list[domain])}\n"
logger.info(f"Left tasks:\n{left_info}")
get_result(
args.action_space,
args.model,
args.observation_type,
args.result_dir,
test_all_meta,
)
test(args, test_file_list)
get_result(
args.action_space,
args.model,
args.observation_type,
args.result_dir,
test_all_meta,
)
test(args, test_file_list)
except KeyboardInterrupt:
logger.info("Main process received KeyboardInterrupt.")
# Signal handler will take care of cleanup
except Exception as e:
logger.error(f"Unexpected error in main process: {e}", exc_info=True)
# Also trigger cleanup for unhandled exceptions
signal_handler(signal.SIGTERM, None)
finally:
# Final cleanup in case any environments or processes remain
logger.info("Main process final cleanup...")
for env in active_environments:
if env is not None:
try:
logger.info(f"Closing environment in final cleanup...")
env.close()
logger.info(f"Environment closed successfully in final cleanup")
except Exception as e:
logger.error(f"Error during final environment cleanup: {e}")
# First try gentle termination
for p in processes:
if p is not None and p.is_alive():
try:
logger.info(f"Terminating process {p.name}...")
p.terminate()
except Exception as e:
logger.error(f"Error terminating process: {e}")
# Wait a moment for processes to terminate
time.sleep(1)
# Then force kill if needed
for p in processes:
if p is not None and p.is_alive():
try:
logger.info(f"Force killing process {p.name}...")
os.kill(p.pid, signal.SIGKILL)
logger.info(f"Process {p.name} force killed")
except Exception as e:
logger.error(f"Error force killing process: {e}")