{ "falconsai.synthesized": true, "num_hidden_layers": 22, "falconsai.tool": "FALCONS.AI Model Surgeon V7.99", "falconsai.attn_note": "attention kernel is chosen at load time (e.g. attn_implementation='flash_attention_2' on CUDA/ROCm hosts that have it); nothing in this file selects it" }