feat: add swanlab logger

[relic-yuexi]

Add SwanLab Logger Support

Summary

This PR adds support for SwanLab as a logging backend for LitGPT training and finetuning workflows. SwanLab is an experiment tracking and visualization tool that provides an alternative to existing loggers like TensorBoard, W&B, and MLFlow.

Based on Lightning-AI/pytorch-lightning#21506

Changes

Core Implementation

• litgpt/utils.py:
  • Added SwanLabLogger import from lightning.pytorch.loggers
  • Extended choose_logger() function to support logger_name="swanlab"
  • Implemented SwanLab logger initialization with support for:
    • project: Project name (defaults to experiment name)
    • experiment_name: Experiment/run name
    • description: Experiment description
    • logdir: Log directory (defaults to out_dir/logs/swanlab)
    • mode: SwanLab mode (defaults to "cloud")
    • config: Additional configuration
    • version: Version (only used when resuming training)
  • Updated error message to include "swanlab" in valid options

Type Annotations

Updated logger_name parameter type hints across all training scripts to include "swanlab":

• litgpt/pretrain.py
• extensions/thunder/pretrain.py
• litgpt/finetune/adapter.py
• litgpt/finetune/adapter_v2.py
• litgpt/finetune/full.py
• litgpt/finetune/lora.py
• litgpt/finetune/lora_legacy.py

Tests

• tests/test_utils.py:
  • Added SwanLabLogger import
  • Added test case for choose_logger("swanlab", ...) with RequirementCache("swanlab") check
  • Verified that invalid logger names raise appropriate ValueError

Usage

Users can now use SwanLab for logging by specifying --logger_name swanlab:

litgpt pretrain --logger_name swanlab --log_args.project my_project pythia-14m

Additional SwanLab options can be passed via --log_args:

litgpt pretrain \
  --logger_name swanlab \
  --log_args.project my_project \
  --log_args.mode disabled \
  --log_args.description "My experiment" \
  pythia-14m

Testing

Tested locally with litgpt pretrain pythia-14m --config config_hub/pretrain/debug.yaml to ensure basic functionality works correctly.

Compatibility

• Requires swanlab package to be installed
• Compatible with all existing training and finetuning workflows
• Follows the same pattern as other loggers (TensorBoard, W&B, MLFlow)