- Fix block_size mismatch causing "params, grads, exp_avgs, and
exp_avg_sqs must have same dtype" RuntimeError when finetuning with
block_size different from checkpoint. Set finetuning config
block_size=1024 to match pretrained model_updated.pt.
- Add warning when config model parameters differ from checkpoint to
help users diagnose shape mismatch issues.
- Fix GradScaler logic: enabled=(dtype=='float32') was inverted,
corrected to enabled=(dtype=='float16').
- Add pyproject.toml with complete dependency list (transformers, tqdm,
tiktoken were missing) and project metadata.
