Harden shared prefill eligibility and cache expansion

Guard shared prefill for non-standard input expansion

Optimize initial prefill in group beam search

Add 'custom_generate/' from commit 'a86e3dab342f58d0adaba71c6229d84058f37e07'

initial commit

Support Transformers v5 cache handling (#4)

Fix typo

Update for new versions!

Update custom_generate/generate.py

downcast the scores after upcasting to prevent runtime errors (#1)

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub