feat(wan): run Wan 2.2 attention on SageAttention when enabled - #9665

Summary

With attention_backend: sage, Wan 2.2 denoising ("Denoise - Wan 2.2" and "Denoise Video - Wan 2.2") enters the SageAttention scope from #9657, so its eligible attention calls run on SageAttention. Everything else stays on PyTorch SDPA, and the default auto is unchanged. The docs page lists Wan 2.2 under Supported models.

Related Issues / Discussions

Stacked on #9657 (SageAttention backend) and #9664, the PR before it. Related: #9529, which asks for SageAttention to speed up Wan 2.2 A14B at 1280×720 on a 24 GB GPU. This PR does not close it: at that size a 24 GB GPU is already full with PyTorch's attention, and SageAttention made it slower (see QA). #9831 tracks that, together with Wan's missing working-memory estimate.

QA Instructions

Wan 2.2 TI2V-5B GGUF Q8 through "Denoise Video - Wan 2.2", model loaded, three arms per seed as described in #9657:

• All 120 attention calls per step ran on SageAttention, including the text cross-attention with 512 text tokens; none fell back.

• Accuracy on real activations, every call of step 0 (120 calls): worst relative L2 0.076, in the text cross-attention; the heads with low cosine there contribute almost nothing (output norm 0.08 against a median of 5.7).

• Wan 2.2 T2V A14B (GGUF Q4_K_M, both experts, 4-step Lightning LoRA V1.1, 81 frames, partial loading on, device_working_mem_gb: 10, which A14B needs on 24 GB without SageAttention too, see [bug]: Wan 2.2 A14B at 720p does not fit 24 GB (no working-memory estimate); SageAttention makes the overflow slower #9831), denoising only:

  	Default SDPA	SageAttention	Peak VRAM
  832×480	121.5 s	60.9 s (−50 %)	18.0 / 18.1 GB
  1280×720	293 s	662 s (+126 %)	24.1 / 24.0 GB

  At 1280×720 the GPU is already full with SDPA; SageAttention's quantized copies (~1.2 GB more than SDPA per call) push it over, and Windows moves memory to shared system memory instead of raising out-of-memory, so its fallback never triggers. The docs row states both results.

• "Denoise - Wan 2.2" (image): TI2V-5B at 1280×704 compresses each image to 880 tokens, below the 1536-token threshold, so all 2400 calls of a measured generation stayed on SDPA.

• The timings and LPIPS 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 Wan 2.2. 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

No material findings for this call site. Remaining: Wan 2.2 A14B at 1280×720 on a 24 GB GPU is slower with SageAttention (#9831).

Compatibility / Rollout

Merge after #9664. 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)