feat(sdxl): run SDXL attention on SageAttention when enabled

[Pfannkuchensack]

Note

Stacked PR 6 of 8: merge after #9662. It targets main, so until the PRs before it merge, its diff also shows their commits and the merges that carry them forward. This PR's own change is one commit, 1968eb9c58.

Summary

With attention_backend: sage, SDXL denoising and SDXL tiled multi-diffusion enter the SageAttention scope from #9657, so eligible attention calls run on SageAttention, including ControlNets that run inside the loop. SD1.5 and SD2 do not enter it: SD1.5's head sizes (40, 80, 160) never qualify, and SD2 is unmeasured. With attention_type: sliced, SDXL's attention does not go through SDPA, so nothing changes there. The default auto is unchanged. The docs page lists SDXL under Supported models and notes the sliced case.

Related Issues / Discussions

Stacked on #9657 (SageAttention backend) and #9662, the PR before it.

QA Instructions

SDXL, model loaded, three arms per seed as described in #9657:

	1024²	2048²
Step, default SDPA	134 ms	631 ms
Step, cuDNN-first SDPA	135 ms	579 ms
Step, SageAttention	126 ms (−6 %)	496 ms (−21 %)
Whole generation	4.4 → 4.2 s (−5 %)	19.6 → 16.1 s (−18 %)
Peak VRAM in the denoise scope	unchanged	unchanged
LPIPS against default SDPA, median (cuDNN-first)	0.027 (0.015)	0.014 (0.006)

• At 1024², 10 of the 140 attention calls per UNet pass qualify: the 4096-token self-attention of the outer blocks. The 1024-token self-attention of the inner blocks and all 77-token text cross-attention are below the break-even and stay on SDPA. At 2048² the inner blocks qualify too.
• Blind side-by-side review: 7 of 8 pairs equal, 1 judged better with SageAttention.
• Accuracy on real activations, every call of step 0 (10 calls): cosine 1.0000, relative L2 0.003.
• The timings, LPIPS and blind review come from the measurement prototype, which still used sageattn()'s default K smoothing; the selection feat(attention): add opt-in SageAttention backend for diffusion models #9657 ships is faster per call.

To try it: install SageAttention as the docs page from #9657 describes, set attention_backend: sage, restart and generate with SDXL. The first generation logs SageAttention ... serves diffusion-model attention on cuda:0 (...) with the GPU and kernel name, and with log_level: debug every generation logs how many calls ran on SageAttention and why the others did not. With attention_backend: auto the output is bit-identical to main.

Checks: ruff clean; pytest tests/app/invocations passes with the whole stack applied (2090 passed). No call-site test was added: this change only enters the scope, whose behavior #9657 tests.

Review

Resolved: the scope first opened for every UNet base, including the unmeasured SD2; it now opens for SDXL only.

Compatibility / Rollout

Merge after #9662. No API or persisted-state change; only attention_backend: sage behaves differently.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)