Repository navigation
feat(qwen-image): run Qwen-Image attention on SageAttention when enabled - #9661
Merged
lstein merged 31 commits intoOct 16, 2026
Merged
Conversation
New `attention_backend: auto | sage` setting; with `sage`, eligible PyTorch SDPA calls inside a denoising scope run on SageAttention 2.2 and every other call keeps SDPA unchanged. Each GPU's first SageAttention call is checked against SDPA and a failing GPU falls back for the session; no model family enters the scope yet. Adds the SageAttention docs page and the regenerated config schema.
With `attention_backend: sage`, FLUX.1 denoising, including ControlNets inside the loop, enters the SageAttention scope. RTX 4090, FLUX.1 dev FP8: 16.1 s → 14.0 s at 1024², 84.2 s → 54.8 s at 2048², peak VRAM unchanged; a blind review judged no image worse.
With `attention_backend: sage`, FLUX.2 denoising enters the SageAttention scope. RTX 4090, FLUX.2 Klein 9B FP8: 15.6 s → 15.1 s at 1024², 27.1 s → 23.1 s at 2048², peak VRAM unchanged; a blind review judged no image worse.
With `attention_backend: sage`, Z-Image denoising enters the SageAttention scope; Z-Image Control masks its attention and keeps SDPA. RTX 4090, Z-Image FP8: 10.4 s → 10.0 s at 1024², 33.7 s → 25.9 s at 2048², peak VRAM unchanged; a blind review judged no image worse.
With `attention_backend: sage`, Qwen-Image denoising enters the SageAttention scope. RTX 4090, Qwen-Image 2512 GGUF Q4_K_M: 78.8 s → 73.5 s at 1024², 208.9 s → 150.3 s at 2048², peak VRAM unchanged; a blind review judged no image worse.
This was referenced Oct 3, 2026
github-actions
Bot
added
python
…ent SageAttention failures sm86 now runs the Triton kernel upstream's sageattn picks there, and sm87 gets none; the Windows build's CUDA routing for both is unexplained and unmeasured here. Out-of-memory and Triton cache errors fall back to SDPA for that call (three retire the device), and the first-call check runs per precision and head size. The kernel runs with the tensors' device current, so a single-device install on cuda:1 uses SageAttention too.
An out-of-memory error now moves only the rest of that generation to SDPA, logged once, instead of counting toward retiring the device; out-of-memory is recognised by the shared is_oom_error. Only Triton cache failures in a row retire a device, and a success resets the count.
# Conflicts: # invokeai/frontend/api/openapi.json
# Conflicts: # invokeai/frontend/api/openapi.json # tests/test_config.py
Pfannkuchensack
marked this pull request as ready for review
October 12, 2026 20:44
…lower with SageAttention On Windows the driver moves memory to shared system memory instead of raising out-of-memory, so the fallback never triggers and the extra memory per call slows every step; issue invoke-ai#9831 tracks a fix.
# Conflicts: # invokeai/app/services/config/config_default.py # invokeai/frontend/api/openapi.json # invokeai/frontend/api/schema.ts
lstein
added a commit
that referenced
this pull request
Oct 16, 2026
> [!NOTE] > Stacked PR 5 of 8: merge after #9661. 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, [`482e7f348f`](Pfannkuchensack@482e7f3). ## Summary With `attention_backend: sage`, Krea-2 denoising enters the SageAttention scope from #9657, so its eligible attention calls run on SageAttention; its grouped-query attention (48 query and 12 K/V heads) runs natively, without expanding K/V. While `INVOKE_KREA2_SDPA_BACKEND` pins a PyTorch kernel for measurement, the scope stays closed so the pinned kernel is what runs. The default `auto` is unchanged. The docs page lists Krea-2 under Supported models. ## Related Issues / Discussions Stacked on #9657 (SageAttention backend) and #9661, the PR before it. ## QA Instructions Krea-2 Turbo GGUF Q4_K_M, model loaded, arms as described in #9657: | | 1024² | 2048² | | --- | --- | --- | | Step, SDPA (Krea-2's own cuDNN-first policy) | 1254 ms | 4976 ms | | Step, SageAttention | 1214 ms (−3 %) | 4186 ms (−16 %) | | Whole generation | 11.5 → 11.0 s (−4 %) | 42.7 → 36.2 s (−15 %) | | Peak VRAM in the denoise scope | unchanged | unchanged | | LPIPS against SDPA, median (no floor, see below) | 0.058 | 0.238 | - Krea-2 already ranks cuDNN first in its own SDPA policy, so the cuDNN-first arm is identical to the default and gives no noise floor: every composition change shows up in LPIPS. The pairs with the largest LPIPS (0.305 at 2048², 0.192 at 1024²) show the same scene with a shifted crop or a missing stack of books; hands and fabric are equally sharp at 1:1 and the lettering is correct. - Blind side-by-side review: 8 of 8 pairs equal. - Accuracy on real activations: block 0 has the same Q/K channel outlier as Qwen-Image (cosine 0.53 with K smoothing, 0.988 without, as #9657 ships). - Per step, 28 of the 32 attention calls run on SageAttention; the other 4 are below the size threshold. - The timings, LPIPS and blind review come from the measurement prototype, which still had K smoothing on; with the shipped selection one generation measured LPIPS 0.055 against SDPA. To try it: install SageAttention as the docs page from #9657 describes, set `attention_backend: sage`, restart and generate with Krea-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 Resolved: with `INVOKE_KREA2_SDPA_BACKEND` set, SageAttention served the calls the override was meant to measure; the scope now stays closed while it is set. ## Compatibility / Rollout Merge after #9661. No API or persisted-state change; only `attention_backend: sage` behaves differently. ## Checklist - [x] _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_ - [x] _Relevant performance/efficiency opportunities considered; material claims have evidence_ - [x] _Material review findings resolved and relevant checks rerun_ - [x] _Documentation added / updated (if applicable)_ - [ ] _Updated `What's New` copy (if doing a release after this PR)_
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Note
Stacked PR 4 of 8: merge after #9660. 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,4abdcd48dd.Summary
With
attention_backend: sage, Qwen-Image denoising enters the SageAttention scope from #9657, so its eligible attention calls run on SageAttention. Everything else stays on PyTorch SDPA, and the defaultautois unchanged. The docs page lists Qwen-Image under Supported models.Related Issues / Discussions
Stacked on #9657 (SageAttention backend) and #9660, the PR before it.
QA Instructions
Qwen-Image 2512 GGUF Q4_K_M, model loaded, three arms per seed as described in #9657:
To try it: install SageAttention as the docs page from #9657 describes, set
attention_backend: sage, restart and generate with Qwen-Image. The first generation logsSageAttention ... serves diffusion-model attention on cuda:0 (...)with the GPU and kernel name, and withlog_level: debugevery generation logs how many calls ran on SageAttention and why the others did not. Withattention_backend: autothe output is bit-identical tomain.Checks: ruff clean;
pytest tests/app/invocationspasses 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: the block-52 error described above; the images were still judged equivalent.
Compatibility / Rollout
Merge after #9660. No API or persisted-state change; only
attention_backend: sagebehaves differently.Checklist
What's Newcopy (if doing a release after this PR)