Do something with describe_tensor/describe_storage outputs

[ezyang]

Sample:

V0925 13:05:39.630000 713625 /data/users/grigorysizov/fbsource/fbcode/caffe2/torch/_subclasses/meta_utils.py:204] {"describe_storage": {"id": 1, "describer_id": 0, "size": 40}, "rank": 0, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
V0925 13:05:39.631000 713625 /data/users/grigorysizov/fbsource/fbcode/caffe2/torch/_subclasses/meta_utils.py:417] {"describe_tensor": {"id": 1, "ndim": 2, "dtype": "torch.int64", "device": "device(type='cuda', index=0)", "size": [1, 4], "dynamo_dynamic_indices": [0], "is_inference": true, "is_leaf": true, "stride": [5, 1], "storage_offset": 1, "storage": 1, "view_func": "<built-in method _view_func_unsafe of Tensor object at 0x7ecbb75a1080>", "describer_id": 0}, "rank": 0, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
V0925 13:05:39.631000 713625 /data/users/grigorysizov/fbsource/fbcode/caffe2/torch/_subclasses/meta_utils.py:1640] {"describe_source": {"describer_id": 0, "id": 1, "source": "L['draft_tokens']"}, "rank": 0, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}V0925 13:05:39.660000 713625 /data/users/grigorysizov/fbsource/fbcode/caffe2/torch/_subclasses/meta_utils.py:204] {"describe_storage": {"id": 2, "describer_id": 0, "size": 2052096}, "rank": 0, "frame_id": 0, "frame_compile_id": 0, "attempt": 0}

These give a detailed description of all the input tensors into compile. I recently used this to aid me in constructing a minimal test case where input size/stride details were important. Maybe a good way to render these is to transform them into equivalent PyTorch code that would construct them, perhaps by using MetaConverter directly (which is what these are fed into). Note that view_func was not successfully serialized, so you need to do something funny instead for that...