ModuleNotFoundError: No module named 'triton' in ComfyUI

Fix ModuleNotFoundError: No module named 'triton' in ComfyUI by separating harmless comfy_kitchen logs from blocking SageAttention, WanVideoWrapper, or acceleration-stack failures.

For ModuleNotFoundError: No module named 'triton' or the long Found comfy_kitchen backend triton: {'available': False, ...} log, first decide whether it is an optional backend report or a workflow-blocking import error.

30-second decision

If the only message is comfy_kitchen backend triton and ComfyUI opens, leave it alone. If Easy-SAM3, Trellis, WanVideoWrapper, SageAttention, or another workflow fails with ModuleNotFoundError: No module named 'triton', repair Triton in the same Python environment that starts ComfyUI. On Windows, start with triton-windows, not a blind pip install triton.

If your blocking message is only No module named 'sageattention' and Triton is not mentioned, use SageAttention missing first.

The two Triton patterns

Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}

ModuleNotFoundError: No module named 'triton'

The first is usually a capability report, not a crash. The second means a node actually tried to import Triton and stopped.

Triage

What you see	Meaning	Best action
`Found comfy_kitchen backend triton: {'available': False, ...}` and ComfyUI opens	Optional backend unavailable	Usually ignore it
`ModuleNotFoundError: No module named 'triton'` during workflow import or queue	A custom node tried to import Triton	Repair Triton in the same Python environment
`No module named 'triton'` during a SageAttention node	SageAttention or a node needs Triton	Fix Triton in the same Python environment
`No module named 'flash_attn'` near SageAttention/Triton setup	Adjacent attention package missing	Treat as part of the same acceleration stack; check whether the workflow truly needs FlashAttention
`pip install triton` fails on Windows	The normal `triton` path is not the Windows fix	Use `triton-windows` instead
Installed Triton but ComfyUI still cannot import it	Installed into the wrong Python	Use the Python that launches ComfyUI
`tcc.exe`, `triton_key`, or compiler errors after installing	Version or runtime mismatch	Check PyTorch/Triton mapping and rollback if needed

If the same startup warning repeats but your workflow still runs, it is noise. If a specific node stops because it imports triton, it is actionable.

What Triton does in ComfyUI

Triton is a GPU programming language and compiler used by PyTorch compiler paths, SageAttention, and some newer acceleration backends. ComfyUI does not need it for ordinary image generation. It matters when you use SageAttention (or a node wrapping it), torch.compile or compiled kernels, comfy_kitchen acceleration backends, or video/large-model workflows that ask for Triton-based acceleration.

SageAttention and Triton often appear as the same repair path: SageAttention is the feature; Triton is a lower-level package it may need for compiled kernels. The safe order is: confirm ComfyUI starts, confirm the exact Python, fix Triton only if the failing workflow needs it, then verify SageAttention after Triton works.

Step 1: Decide whether this is fatal

If the terminal reaches To see the GUI go to: http://127.0.0.1:8188 and a small default workflow runs, treat the Triton line as optional. If a specific workflow fails, read the stack trace and identify which package asked for Triton (often SageAttention nodes, Wan/LTX video workflows, or optimization nodes).

Step 2: Check the exact Python ComfyUI uses

Most failed repairs install into system Python while ComfyUI uses embedded Python. For the GitHub Windows portable package, run from the extracted portable root (the folder with run_*.bat, python_embeded, and ComfyUI):

.\python_embeded\python.exe -s -c "import sys; print(sys.executable); print(sys.version)"
.\python_embeded\python.exe -s -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
.\python_embeded\python.exe -s -m pip show triton triton-windows

For a manual install, activate the launch venv and use python -c ... / python -m pip show .... If torch.cuda.is_available() is False, fix PyTorch/CUDA first. For ComfyUI Desktop, use the app's terminal; do not assume the portable python_embeded path exists.

Step 3: On Windows, do not use the normal `triton` package

Use triton-windows, not a blind pip install triton. Remove an old/wrong triton from the same environment first:

.\python_embeded\python.exe -s -m pip uninstall -y triton

Then install triton-windows:

.\python_embeded\python.exe -s -m pip install -U "triton-windows<3.7"

For a manual venv:

python -m pip uninstall -y triton
python -m pip install -U "triton-windows<3.7"

The version cap avoids pulling a future Triton minor before your PyTorch stack is ready.

Step 4: Match PyTorch and Triton versions

Each PyTorch minor works with a matching Triton minor. As a practical guide (use the upstream triton-windows compatibility table when in doubt):

PyTorch	Expected Triton Minor
2.6	3.2
2.7	3.3
2.8	3.4
2.9	3.5
2.10 or 2.11	3.6

Python version alone is not enough — PyTorch, CUDA, GPU, and the package that needs Triton also matter.

Step 5: Embedded Python may need `include` and `libs`

Some portable builds use embedded Python. The triton-windows instructions call out that Python headers and libraries may need to exist beside the embedded python.exe, in folders named include and libs. Do not confuse libs with Lib (the normal library folder, which should not be modified). If Triton imports but compilation fails or the error mentions missing headers/libraries, follow the triton-windows embedded Python section for the exact include and libs files that match your Python minor version.

Step 6: Verify before restarting

.\python_embeded\python.exe -s -c "import triton; print(triton.__version__)"
.\python_embeded\python.exe -s -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"

If the import test fails, ComfyUI will too. For kernel-level validation, use the test script from the triton-windows repository (a plain import proves the package is installed; a kernel test proves it can compile and run a CUDA kernel).

Common failure cases

Error	Likely cause	Fix
`No module named 'triton'` after install	Wrong Python environment	Use the portable bundled Python, launch venv, or app-managed terminal
`No module named 'triton.language'; 'triton' is not a package`	A local file/folder named `triton` shadows the package	Rename the file/folder
`DLL load failed`	Missing runtime dependency or incompatible wheel	Install Visual C++ Redistributable and check package versions
`tcc.exe` compile failure	Triton runtime/compiler mismatch	Check `triton-windows` and PyTorch versions
`cannot import triton_key`	Package version mismatch	Remove stale Triton packages and install a matching range
ComfyUI worked before installing acceleration packages	Optional package destabilized the environment	Uninstall and return to PyTorch attention

What not to do

Do not install triton into system Python and expect ComfyUI portable to see it.
Do not install Triton only because a startup capability line mentions it.
Do not mix random SageAttention, PyTorch, CUDA, and Triton wheels from different guides.
Do not run package repairs as Administrator unless you intentionally installed ComfyUI in a protected system path.
Do not modify python_embeded\Lib when the instructions say libs.

FAQ

Is `comfy_kitchen backend triton` a fatal error?

Not always. If ComfyUI opens and your workflow runs, it is usually an optional capability report. Treat it as fatal only when a workflow or node fails because it tried to import Triton.

Should I run `pip install triton` on Windows?

Usually no. Use the triton-windows path and install into the exact Python that launches ComfyUI.

Why do Easy-SAM3, Trellis, or WanVideoWrapper mention Triton?

They can use acceleration-heavy nodes. Triton may be lower in the dependency chain than the visible node, so the workflow can fail even though ordinary image generation still works.

Source References

Start free with Wonderful Launcher if this affects your real ComfyUI environment. It keeps launcher-native repair, task logs, and runtime checks in one place; credits are only for image generation and metered tools.

Download Wonderful Launcher See credit packages

Did this fix your issue?

Your answer helps prioritize verified ComfyUI repairs.

30-second decision

If your blocking message is only No module named 'sageattention' and Triton is not mentioned, use SageAttention missing first.

The two Triton patterns

Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}

ModuleNotFoundError: No module named 'triton'

The first is usually a capability report, not a crash. The second means a node actually tried to import Triton and stopped.

Triage

What you see	Meaning	Best action
`Found comfy_kitchen backend triton: {'available': False, ...}` and ComfyUI opens	Optional backend unavailable	Usually ignore it
`ModuleNotFoundError: No module named 'triton'` during workflow import or queue	A custom node tried to import Triton	Repair Triton in the same Python environment
`No module named 'triton'` during a SageAttention node	SageAttention or a node needs Triton	Fix Triton in the same Python environment
`No module named 'flash_attn'` near SageAttention/Triton setup	Adjacent attention package missing	Treat as part of the same acceleration stack; check whether the workflow truly needs FlashAttention
`pip install triton` fails on Windows	The normal `triton` path is not the Windows fix	Use `triton-windows` instead
Installed Triton but ComfyUI still cannot import it	Installed into the wrong Python	Use the Python that launches ComfyUI
`tcc.exe`, `triton_key`, or compiler errors after installing	Version or runtime mismatch	Check PyTorch/Triton mapping and rollback if needed

If the same startup warning repeats but your workflow still runs, it is noise. If a specific node stops because it imports triton, it is actionable.

What Triton does in ComfyUI

Step 1: Decide whether this is fatal

Step 2: Check the exact Python ComfyUI uses

.\python_embeded\python.exe -s -c "import sys; print(sys.executable); print(sys.version)"
.\python_embeded\python.exe -s -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
.\python_embeded\python.exe -s -m pip show triton triton-windows

Step 3: On Windows, do not use the normal `triton` package

Use triton-windows, not a blind pip install triton. Remove an old/wrong triton from the same environment first:

.\python_embeded\python.exe -s -m pip uninstall -y triton

Then install triton-windows:

.\python_embeded\python.exe -s -m pip install -U "triton-windows<3.7"

For a manual venv:

python -m pip uninstall -y triton
python -m pip install -U "triton-windows<3.7"

The version cap avoids pulling a future Triton minor before your PyTorch stack is ready.

Step 4: Match PyTorch and Triton versions

Each PyTorch minor works with a matching Triton minor. As a practical guide (use the upstream triton-windows compatibility table when in doubt):

PyTorch	Expected Triton Minor
2.6	3.2
2.7	3.3
2.8	3.4
2.9	3.5
2.10 or 2.11	3.6

Python version alone is not enough — PyTorch, CUDA, GPU, and the package that needs Triton also matter.

Step 5: Embedded Python may need `include` and `libs`

Step 6: Verify before restarting

.\python_embeded\python.exe -s -c "import triton; print(triton.__version__)"
.\python_embeded\python.exe -s -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"

Common failure cases

Error	Likely cause	Fix
`No module named 'triton'` after install	Wrong Python environment	Use the portable bundled Python, launch venv, or app-managed terminal
`No module named 'triton.language'; 'triton' is not a package`	A local file/folder named `triton` shadows the package	Rename the file/folder
`DLL load failed`	Missing runtime dependency or incompatible wheel	Install Visual C++ Redistributable and check package versions
`tcc.exe` compile failure	Triton runtime/compiler mismatch	Check `triton-windows` and PyTorch versions
`cannot import triton_key`	Package version mismatch	Remove stale Triton packages and install a matching range
ComfyUI worked before installing acceleration packages	Optional package destabilized the environment	Uninstall and return to PyTorch attention

What not to do

Do not install triton into system Python and expect ComfyUI portable to see it.
Do not install Triton only because a startup capability line mentions it.
Do not mix random SageAttention, PyTorch, CUDA, and Triton wheels from different guides.
Do not run package repairs as Administrator unless you intentionally installed ComfyUI in a protected system path.
Do not modify python_embeded\Lib when the instructions say libs.

Did this fix your issue?

Your answer helps prioritize verified ComfyUI repairs.

ModuleNotFoundError: No module named 'triton' in ComfyUI

The two Triton patterns

Triage

What Triton does in ComfyUI

Step 1: Decide whether this is fatal

Step 2: Check the exact Python ComfyUI uses

Step 3: On Windows, do not use the normal `triton` package

Step 4: Match PyTorch and Triton versions

Step 5: Embedded Python may need `include` and `libs`

Step 6: Verify before restarting

Common failure cases

What not to do

FAQ

Is `comfy_kitchen backend triton` a fatal error?

Should I run `pip install triton` on Windows?

Why do Easy-SAM3, Trellis, or WanVideoWrapper mention Triton?

Source References

Table of Contents

ModuleNotFoundError: No module named 'triton' in ComfyUI

The two Triton patterns

Triage

What Triton does in ComfyUI

Step 1: Decide whether this is fatal

Step 2: Check the exact Python ComfyUI uses

Step 3: On Windows, do not use the normal `triton` package

Step 4: Match PyTorch and Triton versions

Step 5: Embedded Python may need `include` and `libs`

Step 6: Verify before restarting

Common failure cases

What not to do

FAQ

Is `comfy_kitchen backend triton` a fatal error?

Should I run `pip install triton` on Windows?

Why do Easy-SAM3, Trellis, or WanVideoWrapper mention Triton?

Source References

Table of Contents