Improvements to MIRA model architecture and dataset handling

[healthkowshik]

Summary

• Cache handling: Check past_key_values before use; use get_seq_length().
• Rotary embedding: Refactor sequence length calculations after cache updates.
• Time quantization: Pre-quantize sequences during dataset initialization.
• Time values: Add time_values handling in window datasets for CT-RoPE.
• Dictionary inputs: Support dictionary inputs with error handling.
• Interface compliance: Implement get_sequence_length_by_idx() method.
• Package structure: Add MIRA/init.py for package initialization.
• Naming: Standardize evaluation strategy parameter naming.

Notes

• Had to create a symbolic link from lowercase mira to uppercase MIRA.

cd MIRA
ln -s MIRA mira

• Handle mismatch in package versions.

pip install --quiet -r requirements.txt
pip install --quiet torchdiffeq
pip install --quiet wandb -qU

python3 -m pip install --upgrade --quiet accelerate peft transformers

[healthkowshik]

@microsoft-github-policy-service agree