Some quick thoughts: I would have imagined that instead of saving the params in the BundledAOTAutogradCacheEntry, we would want to have some similar logic to lift them at runtime into extra graph arguments. Also, is there a repro that uses inline_inbuilt_nn_modules=True where we see the params/buffers that AOT has to lift?

(1) the params themselves can be large tensors, so won't reading/writing them directly to the cache cause problems? (longer read/write latency, also blowing up the size of the cache)

(2) The background here is that... AOTDispatcher used to lift all params/buffers into graph inputs a while ago, but after @anijain2305 landed inline_inbuilt_nn_modules=True, most params/buffers started getting lifted by dynamo. I don't actually have a good sense of which params/buffers are not handled by this case and still need to be lifted by AOT. Maybe @anijain2305 knows? If we want to avoid saving these params to the cache directly, we probably need to either have dynamo properly lift all of these params/buffers, or figure out what their dynamo source is so we can lift them properly on a cache hit.

Ah, it's possible that this is just legacy code, and we don't need to care about it as long as inline_inbuilt_nn_modules=True. I will double check!