perf(qwen-image-2.1): skip unused single-frame VAE cache and add residual-gate reuse opt-in

[Asthenia0412]

Background

Qwen-Image-2.1 VAE decode on vLLM-Omni allocates temporal feature-cache copies even for single-frame decode where they contribute nothing to the output, and re-recreates residual gates per block instead of reusing them.

Problem

• Single-frame untitled VAE decode: temporal feature-cache allocations are materialized but unused
• Residual gates in VAE forward: recreated per residual block

Solution

1. Add opt-in pipeline config qwen21_skip_single_frame_vae_cache (default off)
2. Conditionally skip temporal cache materialization in untitled single-frame VAE decode path
3. Add opt-in qwen21_reuse_vae_residual_gates kwarg threading through VAE blocks

Testing

• Config default off + opt-in configurable
• Single-frame decode: byte-identical output with/without opt-in
• Multi-frame decode: unchanged (opt-in no-ops)
• Tiled path: unchanged
• Distributed VAE: inheritance preserved
• 569-line test module covering all paths
• Memory benchmark script (57 lines)

Relates to #8776 and #8775.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-15T05:36:04.043000Z	6c013c2	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/index.md.

Module owners: @princepride @wtomin @david6666666

Routing: @princepride via module of the changed files, CODEOWNERS; @wtomin via module of the changed files, CODEOWNERS; @david6666666 via module of the changed files

@Asthenia0412, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Resolved as of 4e804ae60fe9: the high-priority or low-quality signal noted on an earlier commit no longer applies.

[hsliuustc0106]

@Asthenia0412 Current-head merge checklist for 6c013c2: GitHub currently reports conflicts with the target branch.

• Contributor: update against the target branch and resolve the conflicts; identify the resulting fix/head commit.
• Contributor: include the local affected VAE test commands, environment, results and log/screenshot evidence for that updated head. For this performance change, preserve the paired baseline/candidate memory and output-parity evidence with exact SHAs.
• Reviewer: inspect the conflict resolution and evidence, then verify the new-head checks and outstanding review concerns.
• Maintainer: confirm whether a CUDA CI run is needed after that verification, avoiding duplicate runs.
• hs: final squash-and-merge confirmation.

This is conflict/evidence triage, not a completed substantive code review.

[BLANKETusers]

The qwen21_reuse_vae_residual_gates opt-in is threaded through the forward signatures, but the cache-producing paths still unconditionally allocate fresh clones (cache_x = ...clone()) and assign them into feat_cache; the _reuse_gates value is never consulted here. As a result, enabling the flag does not reuse any destination buffer or reduce the per-layer allocator churn described in the PR. Please either implement the slot allocation/copy path for this flag or remove the opt-in until it has an observable effect, and add a test that verifies allocation/reuse rather than only output equality.

[vllm-omni-review-bot]

Omni ReviewBot: superseded

The CI failure noted on b7eb32569879 refers to an earlier head; the pull request now points at 2bfeb83fb730.

[vllm-omni-review-bot]

Omni ReviewBot: superseded

The CI failure noted on 2bfeb83fb730 refers to an earlier head; the pull request now points at cc230fdf3e1d.

[Asthenia0412]

@hsliuustc0106 @BLANKETusers Hello could you review it thanks~