[hotfix] fix lora load (#6231)

* [hotfix] fix lora load

* [hotfix] fix hp load

* accelerate deepseek loading
This commit is contained in:
Hongxin Liu
2025-03-01 19:04:14 +08:00
committed by GitHub
parent f32861ccc5
commit 56fe130b15
10 changed files with 146 additions and 38 deletions

View File

@@ -257,7 +257,7 @@ def train(args) -> None:
)
torch.set_default_dtype(torch.float)
booster.load_model(model, args.pretrained)
booster.load_model(model, args.pretrained, low_cpu_mem_mode=False, num_threads=8)
coordinator.print_on_master(
f"Booster init max device memory: {accelerator.max_memory_allocated() / 1024 ** 2:.2f} MB"