transformers

a0baf549 - FIX CIS: gate Mamba2IntegrationTest on require_torch_large_accelerator

FIX CIS: gate Mamba2IntegrationTest on require_torch_large_accelerator Mamba-Codestral-7B is ~13GB in bf16; with its conv + ssm cache + activations the integration suite OOMs on the 22GB CI runners. The tests are not testing something subtle about Mamba behaviour that needs the 7B checkpoint — they're basic 'does the model run' checks — but until they're rewritten on a tiny checkpoint, gate them on 24GB+ accelerators. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>