Use sdpa instead of flash_attention_2 (flash_attn not built; sdpa is torch-native)

init