[AMD][AITER] Dynamically define max_num_total_tokens to avoid OOMs in AITER attention backend buffers allocation
#18263
Loading