[avinashgola]

Summary

• count FSDP flat parameters using their unpadded unsharded size when available, avoiding padded shard inflation in parameter/FLOP accounting
• mark hf_olmo as using legacy tuple caches and accept newer generation kwargs from Transformers, including cache_position and position_ids
• document the offline OLMES path for evaluating converted local/custom checkpoints, including GSM8K-style generative evals

Fixes #695.
Addresses #898 and #854.

Testing

• python -m py_compile olmo/model.py hf_olmo/modeling_olmo.py tests/model_test.py
• ruff check olmo/model.py hf_olmo/modeling_olmo.py tests/model_test.py

Not run: pytest, because this local environment is missing torch.