Is it possible that node's name can change during torch.compile?

torch.compile
1

I encountered a fatal error when the function get_kernel_metadata executes:

Traceback (most recent call last):
  ...
  File "/home/.pyenv/versions/3.11.9/lib/python3.11/site-packages/torch/_inductor/utils.py", line 597, in get_kernel_metadata
    inductor_nodes.sort(
  File "/home/.pyenv/versions/3.11.9/lib/python3.11/site-packages/torch/_inductor/utils.py", line 598, in <lambda>
    key=lambda n: single_graph._inductor_kernel_metadata_node_to_idx_map[n]
                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
KeyError: slice_scatter_default_5

I printed out the graph single_graph whenever it is not None, and I can see that with the same id(single_graph), the single_graph changes during the whole inductor compilation. Yet, the changes are quite minor as only some nodes’ name are modified somehow.

image
image1370×486 127 KB

My questions are:

  1. Is it possible that a node’s name can change during torch.compile? If so, which method or procedure should I look into?
  2. I’m now using PyTorch 2.6, and I see that the torch/_inductor/utils.py file is not updated wrt this get_kernel_metadata. Is there any consumption behind this method, or is there any bug?