def launch_process(cls, index, opt, model_agent, process_queue):
import torch
torch.set_num_threads(1) # prevent threads from spawning in this worker
logging.info(f"Launching background on Index {index}")
opt = copy.deepcopy(opt)
opt['background_index'] = index
try:
world = cls(opt, model_agent=model_agent, process_queue=process_queue)
while True:
world.parley()
except Exception:
import traceback
error = traceback.format_exc()
logging.critical(
f'Exception on background preprocesser index {index}!\n' + error
)
raise
def multiprocess_train(rank,
opt,
port=61337,
rank_offset=0,
gpu=None,
hostname='localhost'):
init_method = f"tcp://{hostname}:{port}"
with distributed_utils.distributed_context(rank,
opt,
rank_offset,
gpu,
init_method=init_method) as opt:
# Run the actual training
opt['multiprocessing'] = True
try:
return single_train.TrainLoop(opt).train()
except Exception:
import parlai.utils.logging as logging
logging.critical(traceback.format_exc())
logging.critical(
f"Got the above exception on worker {rank + rank_offset}. "
"This may cause hangs requiring manual killing of processes.")
raise
