Using the now mandated "triton-backend" completely freezes PC with AMD rx 6800 on windows

[patientx]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

I expect it run without problems with my amd gpu like before triton-backend was enabled forcefully

Actual Behavior

After this #14862 I can't use rx 6800 -rdna2- with latest triton-windows 3.7.1.post27 , it locks up PC when I try to load an int8-convrot model with the native comfyui "load diffusion model" node. The thing is , it was this way before, I even tried to solve it with this PR Comfy-Org/comfy-kitchen#54 , in there I use triton-windows 3.6 but later I updated to 3.7.1 and this freezing up the pc so only reset works was still a thing, and now I am on latest one (3.7.1.post27) YET it is still locking up pc.

Before making a toggle mandatory, look if someone is having problems with it. Not everyone uses linux or latest rdna4 gpus.

Steps to Reproduce

on windows - rdna2 gpu - latest rocm & pytorch - latest triton-windows - load an int8-convrot model with native loader and try to generate , if it somehow does the first generation, it locks up eventually.

Debug Logs

I can't post debug logs because PC locks up as soon as I begin generation with native nodes and triton-backend is active.

Other

This also causes problems on linux with offficial amd triton , added details below.

No response

[patientx]

I tried using the custom node for int8 models, that also locks up too... So essentially I am now unable to use int8-convrot with comfyui.

[0xDELUXA]

@liminfei-amd Have you run any tests on Windows? The workflows I usually use fail with an "unspecified launch failure" after this PR. Reverting it makes everything work again.

[patientx]

OK, it is also not working on linux... With official amd triton installed there.

with this comfyui setup :

with this workflow :

triton is 3.7 and kitchen gets enabled :

getting this error and crash :

[liminfei-amd]

Thanks @patientx — confirmed, this is #14862 auto-enabling the comfy-kitchen Triton backend on ROCm. RDNA2 has no WMMA, so the INT8 kernels hang instead of falling back to eager.

Fix is up: #14869 — it only auto-enables on WMMA/MFMA GPUs, so RDNA1/RDNA2 stay on the eager INT8 path (and adds --disable-triton-backend).

For now the v0.27.0 release works: it predates that PR and keeps the Triton backend opt-in. Thanks for flagging it!

[patientx]

Thanks @patientx — confirmed, this is #14862 auto-enabling the comfy-kitchen Triton backend on ROCm. RDNA2 has no WMMA, so the INT8 kernels hang instead of falling back to eager.

Fix is up: #14869 — it only auto-enables on WMMA/MFMA GPUs, so RDNA1/RDNA2 stay on the eager INT8 path (and adds --disable-triton-backend).

For now the v0.27.0 release works: it predates that PR and keeps the Triton backend opt-in. Thanks for flagging it!

hopefully it fixes this, but how does the custom node work when the kitchen doesn't , because fyi , in eager mode it is waay slower than using my fork of bob's int8 node https://github.com/patientx/ComfyUI-INT8-Fast-ROCM . It is the same when I manage to run with triton-backend before it freezes once or twice, native is slower than that node.

[patientx]

it doesn't enable triton-backend of rdna2 now, same as before essentially.

[liminfei-amd]

it doesn't enable triton-backend of rdna2 now, same as before essentially.

Right, that's intended for now, @patientx — RDNA3/4 have the WMMA matrix cores
the Triton INT8 kernels need, RDNA2 doesn't, so #14869 keeps it on the stable
eager path instead of hanging. We'll open a separate issue to track
RDNA2-and-older support as a follow-up.