fix: adapter checkpoint loading on resume

[geoffreyangus]

This PR ensures that DeepSpeed does not throw an error on resume when loading adapter checkpoints.

Adapter checkpoints only contain a subset of the model's weights. This means that strict checkpoint loading throws a RuntimeError that may look like the following:

> /home/ray/anaconda3/lib/python3.8/site-packages/ludwig/trainers/trainer.py(839)train()
-> self.resume_weights_and_optimizer(training_checkpoints_path, checkpoint)
  /home/ray/anaconda3/lib/python3.8/site-packages/ludwig/trainers/trainer.py(1466)resume_weights_and_optimizer()
-> CheckpointManager.load_latest_checkpoint(checkpoint, model_weights_progress_path, self.device)
  /home/ray/anaconda3/lib/python3.8/site-packages/ludwig/utils/checkpoint_utils.py(332)load_latest_checkpoint()
-> checkpoint.load(last_ckpt, device)
  /home/ray/anaconda3/lib/python3.8/site-packages/ludwig/distributed/deepspeed.py(224)load()
-> _, client_state = self.model.load_checkpoint(save_path, load_lr_scheduler_states=False)
  /home/ray/anaconda3/lib/python3.8/site-packages/deepspeed/runtime/engine.py(2705)load_checkpoint()
-> load_path, client_states = self._load_checkpoint(load_dir,
  /home/ray/anaconda3/lib/python3.8/site-packages/deepspeed/runtime/engine.py(2773)_load_checkpoint()
-> self.load_module_state_dict(checkpoint=checkpoint,
  /home/ray/anaconda3/lib/python3.8/site-packages/deepspeed/runtime/engine.py(2568)load_module_state_dict()
-> self.module.load_state_dict(
  /home/ray/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py(2041)load_state_dict()
-> raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for LLM:
Missing key(s) in state_dict: "model.base_model.model.model.embed_tokens.weight", "model.base_model.model.model.layers.0.self_attn.q_proj.weight", "model.base_model.model.model.layers.0.self_attn.k_proj.weight", "model.base_model.model.model.layers.0.self_attn.v_proj.weight", "model.base_model.model.model.layers.0.self_attn.o_proj.weight",
...

since the base model weights are not stored in the checkpoint. This PR ensures that this strict constraint is relaxed.

[arnavgarg1]

Makes sense! Thanks for this PR @geoffreyangus

[github-actions]

Unit Test Results

  6 files  ±0    6 suites  ±0   19m 47s ⏱️ +6s
12 tests ±0    9 ✔️ ±0    3 💤 ±0  0 ❌ ±0 
60 runs  ±0  42 ✔️ ±0  18 💤 ±0  0 ❌ ±0 

Results for commit 8993a33. ± Comparison against base commit 9992f32.

♻️ This comment has been updated with latest results.