I am developing an asynchronous I/O framework for Windows LLM inference. Suppose I have 8 threads executing computation tasks. When a computation task finds that the required neurons for the next stage are not in the memory pool, the computation thread places the task closure into the I/O queue. An I/O thread then retrieves it, allocates a buffer from the memory pool using a lock-free LRU, and generates a series of I/O closures (usually hundreds for each task) containing the buffer, size, and offset, which are then placed into the I/O execution queue.