🐛 Describe the bug

Summary

_decompose_scatter_mutating in reinplace_inplaceable_ops fails when the input tensor to a scatter op has zero strides (from expand). The decomposition clones with clone_preserve_strides, preserving the aliased memory layout, then attempts copy_ on a view of the aliased clone, which fails.

  RuntimeError: more than one element of the written-to tensor refers to a single memory location

Root Cause

AOTAutograd's backward can produce valid graphs where expand with stride [0, 0] feeds into diagonal_scatter (e.g., from the backward of diagonal_scatter in the forward pass). This pattern appears when diagonal_scatter is used on a small tensor derived from a scalar-like gradient.

In _decompose_scatter_mutating (torch/_inductor/fx_passes/reinplace.py), the decomposition does:

  inp = graph_call_function(graph, aten.clone, inp)   # preserves stride [0, 0]
  tmp = graph_call_function(graph, view.target, tmp, ...)  # diagonal view of aliased memory
  graph_call_function(graph, aten.copy_.default, tmp, src)  # FAILS: aliased memory

graph_call_function runs the ops on fake tensors to compute metadata. aten.clone on a stride-[0, 0] tensor preserves the zero strides (via PyTorch's stride-preserving clone semantics). The subsequent aten.diagonal creates a view where multiple elements point to the same location. aten.copy_ then rejects this.

Suggested Fix

In _decompose_scatter_mutating, force the clone to be contiguous when the input has zero strides:

  # Current:
  inp = graph_call_function(graph, aten.clone, inp)

  # Fix:
  if any(s == 0 for s in inp.meta["val"].stride()):
      inp = graph_call_function(graph, aten.clone, inp, memory_format=torch.contiguous_format)
  else:
      inp = graph_call_function(graph, aten.clone, inp)

This materializes the aliased view into contiguous memory before the mutation, which is both correct and efficient (the clone is already being done — just with different stride semantics).

Affected Ops

Any scatter op in _SCATTER_OP_TO_VIEW with a stride-zero input:

• aten.diagonal_scatter
• aten.select_scatter
• aten.slice_scatter
• aten.as_strided_scatter

Repro

The attached fx_graph_runnable_18.txt (generated by tlparse) reproduces the issue directly: python fx_graph_runnable_18.txt

This is a backward graph produced by aot_export_joint_with_descriptors for a model using torch.diagonal_scatter with ignore_diagonal=True in its loss function. The backward decomposition produces expand(scalar_grad, [2, 2]) with stride [0, 0], followed by diagonal_scatter.

torch.compile's single-shot path avoids this because functionalization ordering prevents the pattern from reaching reinplace. The two-stage path (aot_export → transform → aot_compile, used by AutoParallel) exposes it.

fx_graph_runnable_18.txt

Versions

PyTorch nightly from 2.12.0.dev20260323+cu128

cc @ezyang @gchanan @kadeng @msaroufim @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @muchulee8 @amjames @aakhundov @coconutruben @jataylo