PyTorch Lightning

FileNotFoundError in pytorch-lightning often arises when file paths used for saving checkpoints, configurations, or logs are invalid or the destination directory doesn't exist. To fix this, ensure all specified directories exist before writing to them, using `os.makedirs(path, exist_ok=True)` to create them if needed; also, validate that file paths are correctly formed, especially when dealing with absolute paths or paths with special characters on different operating systems. Consider using `os.path.join` for robust path construction.

The "NotImplementedError" in PyTorch Lightning usually arises when a required method (like `training_step`, `configure_optimizers`, or callbacks expecting specific hooks) is not defined in your LightningModule or Callback. Resolve this by ensuring you've overridden all necessary methods in your LightningModule or Callback classes with your custom logic, paying close attention to the expected inputs and outputs for each method as defined in the PyTorch Lightning documentation. If using callbacks, verify existing inherited methods.

ProcessExitedException in PyTorch Lightning tests often arises from unexpected process termination during multi-processing scenarios such as `ddp_fork`. This commonly stems from resource exhaustion, unhandled exceptions within child processes, or conflicts with system-level libraries. To resolve it, ensure adequate system resources (RAM, CPU), implement robust error handling within the child processes, and check for library incompatibilities, especially with multiprocessing on MacOS.

OutOfMemoryError in PyTorch Lightning typically occurs when the GPU runs out of memory during training. Fix this by reducing the `batch_size` in your DataLoader, using gradient accumulation with `accumulate_grad_batches` in the Trainer, or offloading computations to the CPU using `torch.cuda.empty_cache()` periodically and enabling `model.half()` for mixed precision training if applicable. Consider larger GPUs or distributed training for further memory relief.

DistNetworkError in PyTorch Lightning distributed tests often arises from address conflicts, specifically the EADDRINUSE error, indicating a port is already in use. To fix, specify available ports by setting the MASTER_PORT environment variable using os.environ["MASTER_PORT"] = str(find_free_port()), or configure the TCP store to find an open port automatically, mitigating address collisions during distributed initialization.

WandbAttachFailedError in pytorch-lightning often arises when Wandb is initialized outside of the main process in a distributed training setting, especially with TPUs, which interferes with proper experiment tracking. To fix this, ensure Wandb is only initialized within the main process (rank 0) by using a conditional check like `if self.trainer.global_rank == 0: wandb.init(...)` or utilize pytorch-lightning's built-in WandbLogger to handle this automatically.