src/utils/patch_bugs.py

import torch
import lightning.pytorch.loggers.wandb as wandb

setattr(wandb, '_WANDB_AVAILABLE', True)
torch.set_float32_matmul_precision('medium')

# torch.compile / dynamo fallback: don't kill training when dynamo fails to
# trace a graph (e.g. symbolic shape mismatches in compiled HF text encoders);
# just run that node in eager mode instead.
torch._dynamo.config.suppress_errors = True
# also try not to recompile forever when shapes keep changing on us
torch._dynamo.config.cache_size_limit = 64

import logging
logger = logging.getLogger("wandb")
logger.setLevel(logging.WARNING)

import os
os.environ["NCCL_DEBUG"] = "WARN"
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)
warnings.simplefilter(action='ignore', category=UserWarning)