initialise process group for tp

This commit is contained in:
bursteratom
2024-12-11 11:37:21 -05:00
parent 85381b6b15
commit b17b1aada7

View File

@@ -827,7 +827,7 @@ class ModelLoader:
_ = _configure_zero3_memory_efficient_loading()
if self.cfg.tensor_parallel == "auto":
rank = int(os.environ["RANK"])
rank = int(os.environ.get("LOCAL_RANK", 0))
device = torch.device(f"cuda:{rank}")
torch.distributed.init_process_group("nccl", device_id=device)