[laifenxiawucha]

Summary

Fixes an AttributeError in torch.nn.utils.prune.PruningContainer when a method is added to an empty container.

Root Cause

PruningContainer.add_pruning_method assumed _tensor_name had already been initialized. For PruningContainer() followed by add_pruning_method(...), that attribute was missing, so the first add raised AttributeError instead of adopting the method's tensor name.

Changes

• make PruningContainer adopt _tensor_name from the first added pruning method
• add a regression test covering empty-container incremental construction

Verification

Used the issue reproducer with the patched torch.nn.utils.prune module loaded from this branch in the provided XPU Python environment:

/home/jyz/cuda-fix-alignment/.venv/bin/python <inline repro using /tmp/opencode/pytorch-issue-1895/torch/nn/utils/prune.py>

Observed:

torch 2.14.0.dev20260705+xpu
xpu_available True
container tensor weight xpu:0
RESULT: not-reproduced

Issue: https://github.com/intel-sandbox/torch-xpu-ops-exp/issues/1895