Under torch 2.12.0, vLLM's vllm::sparse_attn_indexer custom operator (used by DeepSeek-V3.2's sparse attention indexer) hits PyTorch's custom-op aliasing check. The op returns a tensor that aliases one of its inputs (or another return).

Behavior depends on environment:

• In CI (CI env var set, e.g., GitHub Actions): hard RuntimeError, engine fails to start.
• Outside CI (regular user-land): only a UserWarning, execution continues.

The check is gated by:

# torch/_functorch/config.py
check_custom_op_aliasing = True
error_on_custom_op_aliasing = bool(os.getenv("CI"))

In 2.12, the plan is to flip error_on_custom_op_aliasing to True by default (i.e., everywhere, not only in CI). At that point user-land will see the same hard error.

The error message:

RuntimeError: vllm::sparse_attn_indexer (with implementation in ???): The output of this custom operator (1) must not also be an input to this custom operator and (2) may not alias any inputs to this custom operator or other returns. The most common way to trigger this error is if we have y = custom_op(x) and y and x are the same Tensor. Please instead return a clone of the offending output tensor(s) (e.g. return x.clone()) or refactor the custom operator to not return y.

The user-land warning is the same text prefixed with UserWarning: and ending with: This is deprecated and will become an error in PyTorch 2.12.

Environment

• torch: 2.12.0+cu130 (test channel)
• triton: 3.7.0
• CUDA: 13.0
• Python: 3.12
• GPU: NVIDIA B200 (linux.dgx.b200.8, TP=8)
• vLLM model: deepseek-ai/DeepSeek-V3.2

Reproduction

CI=1 vllm serve deepseek-ai/DeepSeek-V3.2 --tensor-parallel-size 8 --disable-log-stats  --load-format dummy --compilation-config '{"cudagraph_mode":"PIECEWISE","use_inductor_graph_partition":true}'

(Worker_TP5 pid=47907) ERROR [multiproc_executor.py:962]
  RuntimeError: vllm::sparse_attn_indexer (with implementation in ???):
  The output of this custom operator (1) must not also be an input to this custom operator
  and (2) may not alias any inputs to this custom operator or other returns.
  ...
(APIServer pid=47461) RuntimeError: Engine core initialization failed.

Outside CI, the same model loads — only the UserWarning is emitted (per #173844's "warn by default" rollout).

Failing job: https://github.com/pytorch/pytorch/actions/runs/24839322153/job/72733144857

Why this is a 2.12-surfaced regression

The aliasing check is a deliberate PyTorch tightening for 2.12:

• PR #166545 (2026-01-02) — added _check_custom_op_aliasing runtime check, gated by config flags.
• PR #173844 (2026-02-03, "Warn wrong custom ops by default") — flipped on by default. PR description: "By 2.12, we want to error on custom ops that have wrong schemas."

So in 2.11 + GitHub Actions (CI=1) this is already a hard error; in 2.11 user-land it's only a warning. In 2.12 the default flips to error everywhere.

Where the offending op is registered

vllm::sparse_attn_indexer was introduced in vLLM PR vllm-project/vllm#29287 ([ROCm][Deepseekv3.2] Refactor Sparse Indexer as CustomOp, merged 2025-11-24). The CustomOp wraps the heavy DSA kernels (fp8_mqa_logits / fp8_paged_mqa_logits, etc.) for the DeepSeek-V3.2 sparse MLA path. Its current return tensor either is, or aliases, one of the input tensors.

cc @ezyang @gchanan @seemethere @malfet @pytorch/pytorch-dev-infra @chauhang @penguinwu @bdhirsh @bobrenjc93 @aorenste @atalman @zou3519 @Lucaskabela @angelayi @tugsbayasgalan

Fixed

• Fixed by PR: Make custom op aliasing check warn (not error) in CI (https://github.com/pytorch/pytorch/pull/182068)