423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-22 13:42:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix #4077

Browse Source 
Former-commit-id: 90ed3cae92c624cb53e7349bb7edeaf402d030e6
This commit is contained in:
MengqingCao 2024-06-05 08:03:30 +00:00
parent 51e3229528
commit 55815ab1ff
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
src/llamafactory/cli.py
View File

@ -70,7 +70,11 @@ def main():
elif command == Command.EXPORT: elif command == Command.EXPORT:
export_model() export_model()
elif command == Command.TRAIN: elif command == Command.TRAIN:
if get_device_count() > 1: if get_device_count() > 0:
# NOTE (MengqingCao): why use torchrun when only one accelerator is available?
# DeepSpeed only warp model with DeepSpeedEngine when launching by distributed launcher,
# e.g., torchrun, causing some feature missing
# sa: https://github.com/huggingface/transformers/issues/24309
master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1") master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999))) master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port)) logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))