Transformers v1.7.0 regional compilation in TorchDynamoPlugin, layerwise casting, FSDP2 FULL_STATE_DICT, and QLoRA support

AI Impact Summary

v1.7.0 introduces regional compilation in TorchDynamoPlugin, accelerating first-inference by caching and reusing optimized code for repeated blocks such as decoder layers. It adds per-layer casting hooks to enable storage and compute dtypes (e.g., FP8) per layer, reducing peak memory and enabling larger models within the same hardware. The release also broadens deployment with FULL_STATE_DICT support for FSDP2, QLoRA training, and CPU offload memory fixes; teams should verify their model pipelines (enable use_regional_compilation, attach_layerwise_casting_hooks, and test QLoRA paths) to realize the gains and avoid regressions.