Hugging Face RecurrentGemma comparison

Pinned source: huggingface/transformers commit 42ca97014c85d71a88ad60d55f08cb9fb4d26e2c, under src/transformers/models/recurrent_gemma/.

The implementation supplies model integration, generation cache handling and a readable PyTorch recurrence. At the audited commit, RG-LRU prefill still uses a Python time loop and decode updates the cache with ordinary PyTorch operations. No custom CUDA RG-LRU scan, Triton training kernel or custom backward is present in that path.

HF is therefore an integration/reference baseline, not a candidate for the strongest optimized GPU ranking. Small correctness/integration observations can be reported, but running its Python loop on H800 would not answer the public optimized-kernel question and is excluded from performance claims.