Skip to content

[Qwen-Image 2.1 / Z-Image] Complex-valued RoPE hard-fails on backends without complex64 kernels (Ascend NPU: aclnnIndex EZ1001) #15338

Open

Description

@cyzlmh

Describe the bug

Two pipeline families in the Qwen-Image lineage still build their RoPE table with torch.polar and then index/multiply it as complex64 on the device, with no real-valued fallback:

  • transformer_qwenimage21.py — QwenImage21Rope.forward hardcodes use_real=False
  • transformer_z_image.py — torch.polar + view_as_complex

transformer_qwenimage.py (Qwen-Image 1.0) already has ROPE_PER_DEVICE from the AWS Neuron work in #13718, but only "neuron" is registered, so other accelerators keep taking the complex path.

On backends without complex64 kernels this is a hard failure rather than a slowdown, and it affects three targets, not one:

  1. Ascend NPU (torch_npu) — the op-plugin has no complex64 index/mul. QwenImage21Rope.forward raises aclnnIndex EZ1001 as soon as the complex frequency table is indexed on device. Qwen-Image 2.1 and Z-Image currently cannot run on Ascend NPUs without an out-of-tree monkey-patch.
  2. CoreML export — torch.polar has no MIL equivalent (RoPE implemetation, Torch.polar op not available in MIL requesting support apple/coremltools#2258).
  3. torch.compile — inductor cannot codegen complex ops (Complex numbers in Qwen/Z-Image Image pipeline incompatible with torch.compile #12668), which matches this same code path.

CUDA eager is unaffected, which is presumably why this has stayed low-pressure upstream.

I understand #12668 already tracks the complex-op problem from the torch.compile angle; opening this as a separate issue because the backend failure mode and the fix scope are different (see "Describe the solution you'd like"). Happy to close this if you would rather consolidate into #12668.

Reproduction

The failure is device-specific, but the code path is easy to isolate. On an Ascend NPU with a from-source diffusers checkout:

import torch
from diffusers.models.transformers.transformer_qwenimage21 import QwenImage21Rope

# Any small instantiation of the rope module is enough; the failure happens at the
# first on-device indexing of the complex frequency table.
rope = QwenImage21Rope(...).to("npu")          # or .to(torch.device("npu"))
rope(x, freqs_cis)                              # -> aclnnIndex EZ1001

A backend-agnostic sanity check for the same assumption (runs anywhere, no model weights needed):

import torch
freqs = torch.polar(torch.ones(8, 4), torch.rand(8, 4)).to("npu")   # complex64 on device
freqs[torch.arange(8).to("npu")]                                    # index -> unsupported on NPU

Describe the solution you'd like

Extend the existing #13718 pattern instead of adding a new one:

  • port ROPE_PER_DEVICE + the real-valued apply_rotary_emb_qwen_neuron path into transformer_qwenimage21.py and transformer_z_image.py;
  • keep the frequency table and its indexing on CPU, crossing the device boundary as reals (_get_device_freqs already does this for Neuron);
  • register "npu" in ROPE_PER_DEVICE for Qwen-Image 1.0.

Context

I hit this while bringing up image models on a single Ascend 910B2; Qwen-Image 2.1 was the only pipeline in the set that could not run at all without patching.

Disclosure, and a note on scope: I opened PR #14905 for exactly this in September and it was auto-closed because it did not link an issue — that is on me, not on the change. Before re-opening anything I would like an explicit ack here, as the AI-assisted contribution section asks. I am also happy to split the work into one PR per pipeline instead of one size/L PR covering both, if that is easier to review.

Logs

RuntimeError: The inner error is reported as above. The process exits for this inner error. (Inner error: EZ1001: ... aclnnIndex ...)
  File "diffusers/models/transformers/transformer_qwenimage21.py", line ..., in forward
    ...

(The exact op-level message from aclnnIndex is the one that names the unsupported complex64 dtype.)

System Info

- diffusers: main (from-source checkout at the time of report; PR #14905 was opened 2026-09-29)
- platform: Linux aarch64
- Python: 3.12.13
- torch: 2.10.0+cpu
- torch_npu: 2.10.0.post4
- CANN: 9.1.0
- accelerator: 1x Ascend 910B2 (HBM 64 GB)

Who can help?

Anyone with familiarity with the Qwen-Image RoPE implementations, or with Apple/Neuron backends where the same real-valued path already exists.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions