Buckets:
| import os,sys,json,math,time | |
| from pathlib import Path | |
| root=Path('/marimo/shuiwu') | |
| os.environ['HF_HOME']=str(root/'.cache/huggingface') | |
| os.environ['TORCHINDUCTOR_CACHE_DIR']=str(root/'.cache/torchinductor') | |
| os.environ['TRITON_CACHE_DIR']=str(root/'.cache/triton') | |
| os.environ['TOKENIZERS_PARALLELISM']='false' | |
| sys.path.insert(0,str(root/'LoRA-Forge/lora_forge/trainers/moss_image')) | |
| import torch | |
| # The original DiT remains unquantized bf16; LoRA and optimizer state are fp32. | |
| torch.set_num_threads(8) | |
| from moss_train_network import Trainer,read_config | |
| class RecordedTrainer(Trainer): | |
| def step_logs(self,loss,average,scheduler): | |
| if not math.isfinite(loss):raise RuntimeError('Non-finite training loss') | |
| result=super().step_logs(loss,average,scheduler) | |
| with (root/'logs/metrics.jsonl').open('a') as f:f.write(json.dumps({'time':time.time(),**result})+chr(10)) | |
| return result | |
| config=read_config(['--config_file',str(root/'train.toml')]) | |
| RecordedTrainer(config).train() | |
Xet Storage Details
- Size:
- 972 Bytes
- Xet hash:
- f666e3cf4f229b16e5156f1548861e85570ed97cc686b240970e34185ddc8bd9
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.