Triton Sageattention

Triton + SageAttention (ComfyUI acceleration)

See also comfyui-launch-flags for the full attention / VRAM / cache flag matrix. Note the Z-Image exception: Z-Image is broken under --use-sage-attention, so launch it with --use-pytorch-cross-attention instead.

Prefer kitchen INT8 attention when it is available

If kitchen action:"status" (or panel_kitchen) reports kitchen present and int8_attention_is_available on this GPU, launch with --use-ck-attention and skip the sageattention wheel dance. Kitchen INT8 attention is a ComfyUI flag; it does not need a version-matched sageattention wheel. Restart required, consent-gated like every restart.

Only fall through to the Triton + SageAttention install below when kitchen INT8 is unknown or not available. A failed kitchen probe is unknown, not a no.

Overview

Two optional accelerators that many modern video graphs (especially kijai's ComfyUI-WanVideoWrapper) reference by default:

• SageAttention (import sageattention), a quantized attention kernel. Selected via a node's attention_mode = sageattn (WanVideoWrapper) or ComfyUI's --use-sage-attention startup flag. ~20 to 40% faster sampling on supported NVIDIA GPUs.
• Triton, the GPU kernel compiler that inductor torch.compile needs. WanVideoWrapper's WanVideoTorchCompileSettings (and any torch.compile/ inductor node) compiles the model through Triton for another speedup.

The risk. Both are version-locked to your exact torch + CUDA + python. A wrong wheel does worse than fail to install. It can break the torch install (mismatched CUDA DLLs, ImportError, or silent NaNs). And the failure mode of not having them is a hard crash before any sampling: ValueError: Can't import SageAttention: No module named 'sageattention', or compile errors / triton: unavailable in the startup log. This is exactly the video-extend TRAP 5.

Therefore the default is to get a working render FIRST with the sdpa / no-compile fallback, then OFFER to install acceleration for speed. Never run a torch-breaking install unannounced to "fix" a workflow. Fall back, render, then ask.

Verification note (June 2026). Wheel sources, the triton↔torch table, and the live attention_mode enum below were verified against woct0rdho/triton-windows, woct0rdho/SageAttention releases, and WanVideoWrapper's nodes (see Sources). Versions move fast, so always re-read the live torch/CUDA/python first (commands below) and pick the wheel that matches. Flag anything you can't confirm rather than guessing.

Decide first: do you even need them?

Workflow crashes "No module named 'sageattention'"  ──┐
  or "triton: unavailable" / torch.compile error     ─┤
                                                       ▼
              1. APPLY THE SDPA / NO-COMPILE FALLBACK  → render works now
                                                       ▼
              2. OFFER acceleration, in this order:
                 a. If kitchen INT8 attention is available:
                    "Want --use-ck-attention? No sageattention wheel."
                 b. Else:
                    "Want me to install Triton + SageAttention for ~20–40%
                     faster sampling? It's a version-matched install that
                     touches your torch env — I'll verify torch/CUDA/python
                     first and can roll back."
                                                       ▼
              3. Only on YES → install per-OS below → verify → re-enable
                 sageattn + torch.compile in the workflow.

Mac (no CUDA): skip the install entirely. The answer is always sdpa/MPS.

The safe sdpa / no-compile fallback (DO THIS FIRST)

When Triton/SageAttention aren't installed, make the workflow run unaccelerated but correct by switching attention to sdpa (PyTorch's built-in scaled dot-product attention, always available, no extra deps) and removing the torch.compile/inductor wiring.

WanVideoWrapper (the common case):

1. On every WanVideoModelLoader set attention_mode to sdpa.
   • Confirmed enum values: sdpa, flash_attn_2, flash_attn_3, sageattn, sparse_sage_attention. The examples ship with sageattn; sdpa is the universal safe one.
2. Disconnect WanVideoTorchCompileSettings from each loader's compile_args input (or delete/bypass the node). No compile = no Triton needed.
3. (If present) bypass any WanVideoSetRadialAttention / sparse_sage_attention node. Those also route through SageAttention.

Generic ComfyUI: don't launch with --use-sage-attention; bypass any TorchCompileModel / inductor node.

This costs you speed, not quality. Use create_workflow (action:"modify") / the panel's strip-and-re-point flow to flip the widget and drop the link, then enqueue. Once it renders, offer the install.

Cross-ref: video-extend documents this exact fix as TRAP 5 for the Pusa extension graph (both WanVideoModelLoaders → attention_mode=sdpa, disconnect WanVideoTorchCompileSettings).

Windows install (the priority)

Windows has no official Triton or SageAttention build. You use community prebuilt wheels, and they must match torch + CUDA + python exactly. The panel agent has a shell (Bash for Claude / exec for Codex). Use it to run these in the correct python, never the system python.

Step 1 — find the RIGHT python (NOT system python)

ComfyUI on Windows comes in three flavors; each has its own python whose pip you must target:

Detect it from the live server, the surest way to hit the same python ComfyUI runs on:

• install_comfyui (action:"environment") / get_system_stats report embedded_python (true → Portable), the python version and the pytorch_version (e.g. 2.10.0+cu130).
• Inspect the running process's argv (from get_system_stats). The path to main.py reveals the install root; its sibling standalone-env / python_embeded holds the python.
• Last resort, ask the user for their ComfyUI folder.

Installing into the wrong python (e.g. a global pip install) is the #1 Windows mistake. The package lands somewhere ComfyUI never imports from, so the loader still crashes "No module named 'sageattention'". Always use "<that python>" -m pip.

Step 2 — read the installed torch + CUDA + python

Run with the python you found:

"<python>" -c "import sys, torch; print(sys.version.split()[0], torch.__version__, torch.version.cuda)"

Example live output on this machine: 3.13.12 2.10.0+cu130 13.0, meaning python 3.13, torch 2.10, CUDA line cu130. You'll pick wheels for that triple.

Step 3 — install triton-windows (matched to torch)

Source: woct0rdho/triton-windows (the canonical Windows Triton fork; also on PyPI as triton-windows). The pin is an upper bound. pip resolves the right build for your torch:

"<python>" -m pip install -U "triton-windows<3.7"

Why <3.7: each torch minor pins a Triton minor. Verified table:

(torch 2.6 or older → triton 3.2 or earlier.) Pick the row for your torch.

• CUDA toolkit: since triton-windows 3.2.0.post11 a minimal CUDA toolchain is bundled in the wheel, so you do NOT need a separate CUDA Toolkit install for Triton itself. (Triton 3.3 through 3.6 bundle the CUDA 12.8 line; works against cu12x/cu13x torch.)
• MSVC / vcredist: Triton compiles C++ at runtime, so it needs the MSVC toolchain and "Visual C++ Redistributable 2015-2022" present. A TinyCC is bundled (since 3.2.0.post13) which covers many cases, but installing the Visual Studio Build Tools (C++ workload) plus the latest vcredist is the reliable fix if you hit compiler errors (see Traps).
• Embedded/Portable python only: the embedded distro ships without C headers, so Triton can't compile. Download the matching python_<ver>_include_libs.zip from the triton-windows releases and copy its include and libs (note: libs, not lib) folders into python_embeded\. The Desktop standalone-env usually already has these.

Step 4 — install SageAttention (prebuilt wheel, matched to torch+CUDA)

Prefer the prebuilt wheel. Building from source needs the full CUDA Toolkit (nvcc) plus MSVC and often fails on Windows. Source: woct0rdho/SageAttention releases (Windows wheels; v2 = SageAttention 2.x).

Latest verified tag: v2.2.0-windows.post5, with these four wheels (all cp310-abi3, so they work on python 3.10 through 3.13+ via the stable ABI; one wheel covers all those pythons):

Pick by your CUDA line (cu128 vs cu130, from torch.version.cuda: 12.8 → cu128, 13.0 → cu130) and torch minor. For the live machine above (torch 2.10.0+cu130, py3.13) that is the last wheel. Install by full URL:

"<python>" -m pip install "https://github.com/woct0rdho/SageAttention/releases/download/v2.2.0-windows.post5/sageattention-2.2.0+cu130torch2.10.0andhigher.post5-cp310-abi3-win_amd64.whl"

• The cpXXX-abi3 tag means one wheel works across python ≥ its base (3.10+), so py3.13 is covered even though there's no cp313-specific wheel. This is expected, not a mismatch.
• Always check the releases page for a newer tag than .post5 and newer torch variants. The filename pattern is stable (+cu<line>torch<minor>...abi3).
• Don't build from source unless no wheel matches your torch/CUDA at all (then you need CUDA Toolkit plus MSVC; flag the cost to the user first).

Step 5 — verify (Windows)

"<python>" -c "import triton; print('triton', triton.__version__)"
"<python>" -c "import sageattention; print('sageattention OK')"
"<python>" -c "import torch; print('torch still ok', torch.__version__, torch.cuda.is_available())"

All three must succeed and torch must still import with CUDA. If the third line now fails, the install clobbered torch (see Traps, roll back). Then restart ComfyUI and confirm the startup log no longer prints Could not load sageattention / triton: unavailable. Finally re-enable in the workflow: WanVideoModelLoader.attention_mode = sageattn and reconnect WanVideoTorchCompileSettings, enqueue, and confirm it samples (a torch.compile node will spend extra time on the first run compiling, which is normal).