transformers

Speedup model init on CPU (by 10x+ for llama-3-8B as one example)

#31771

Merged