[Wan2.2 VAE, Qwen-Image-2.1 VAE] Drop redundant input clone in residual up/down blocks

[luca-888]

What does this PR do?

Fixes #15040

WanResidualDownBlock and WanResidualUpBlock (and their QwenImage21 copies) do x_copy = x.clone() and only use x_copy for the avg_shortcut at the end. Nothing in between modifies x in place, so the clone is unnecessary. It only keeps an extra activation-sized buffer alive for the whole block. This PR replaces it with x_copy = x.

Outputs are bitwise identical, peak memory goes down, and speed is unchanged:

A100-80GB, released weights, torch 2.10	peak before	peak after
Qwen-Image-2.1 decode 1024², bf16	7.20 GiB	6.63 GiB
Qwen-Image-2.1 decode 2048², bf16	26.87 GiB	24.62 GiB
Wan2.2 decode 121×704×1280, fp32	21.94 GiB	20.22 GiB
Wan2.2 decode 121×704×1280, bf16	12.71 GiB	11.85 GiB

Tests

Existing VAE tests, on CPU (torch 2.14.1), same result as on the parent commit 08d6977:

$ pytest tests/models/autoencoders/test_models_autoencoder_wan.py tests/models/autoencoders/test_models_autoencoder_kl_qwenimage21.py
35 passed, 45 skipped, 5 warnings in 14.29s

The skips are the usual CPU ones (accelerator-only, nightly, auth token, unsupported features).

Copies and style:

$ python utils/check_copies.py
(no output, exit 0)
$ ruff check src/diffusers/models/autoencoders/autoencoder_kl_wan.py src/diffusers/models/autoencoders/autoencoder_kl_qwenimage21.py
All checks passed!

Before/after on released weights (A100-80GB, torch 2.10.0+cu130), equal is torch.equal on the outputs:

qi21_bf16_enc_1024            equal=True  peak  2.90 ->  2.72 GiB  t 0.07 -> 0.06 s
qi21_bf16_dec_1024            equal=True  peak  7.20 ->  6.63 GiB  t 0.21 -> 0.21 s
qi21_bf16_enc_2048            equal=True  peak  9.70 ->  8.95 GiB  t 0.24 -> 0.23 s
qi21_bf16_dec_2048            equal=True  peak 26.87 -> 24.62 GiB  t 0.84 -> 0.84 s
wan22_fp32_enc_1f_704x1280    equal=True  peak  4.28 ->  4.15 GiB  t 0.59 -> 0.59 s
wan22_fp32_dec_121f_704x1280  equal=True  peak 21.94 -> 20.22 GiB  t 248.32 -> 248.22 s
wan22_bf16_dec_121f_704x1280  equal=True  peak 12.71 -> 11.85 GiB  t 14.00 -> 13.95 s
wan22_fp32_enc_121f_704x1280  equal=True  peak 11.12 -> 10.58 GiB  t 53.11 -> 53.07 s

Self-review notes

Posted as a comment below.

Before submitting

• Used an AI agent (Claude Code); ran self-review, notes in the comment below.
• Read the contributor guideline.

Who can review?

@yiyixuxu @dg845

[github-actions]

Hi @luca-888, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.

Please note that PRs without a linked issue are likely to be automatically closed 10 days after this notice.

Once the PR links an issue (or gets the no-issue-needed label), you can ignore this message — it stays here as a comment, but it no longer applies.

[luca-888]

Self-review notes

READY, no blocking issues. Nothing in these blocks writes to x in place, so x_copy = x is safe.

Left for the reviewer:

1. No new test: outputs don't change, and existing VAE tests cover both blocks.
2. The feat_cache[idx] = x.clone() calls are left alone; they are a separate case.