How to force int8 in the latest update?

[okolenmi]

Custom Node Testing

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

Your question

I have tried to load int8 model and I have this output in console as result:

[INFO] Requested to load AnimaTEModel_
[INFO] loaded completely; 10996.80 MB usable, 1136.88 MB loaded, full load: True
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: int8_tensorwise , emulated ops: mxfp8, nvfp4, float8_e5m2, float8_e4m3fn
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLOW

There are no int8 option in unet node to force it and there are no cli parameter that may force it (despite all other exist):

Console says that "int8_tensorwise" available natively, so why is it casted to bf16? And why is it marked as "manual cast"? I have no cli parameters that may force bf16 precision.

Other

I have Intel Arc B580 GPU. It should support int8 natively (it doesn't support fp8).

[georgearchviz-maker]

it doesn't work on turing GPUs either.

[Aamir3d]

From what I understand, int8 and convrot isn't available for AMD (and probably Intel) cards just yet. Perhaps one of the devs can clarify.

[0xDELUXA]

From what I understand, int8 and convrot isn't available for AMD (and probably Intel) cards just yet. Perhaps one of the devs can clarify.

It is (at least on AMD with triton-windows installed), using this custom node: https://github.com/patientx/ComfyUI-INT8-Fast-ROCM.
Native INT8 convrot support for AMD in ComfyUI isn't ready yet.

[okolenmi]

The problem is not support. I can force fp8 which is not supported (even if it's slow), but there are no such option for int8 which is supported for my GPU as log says. Log is a more credible source than random people here.

[0xDELUXA]

The problem is not support. I can force fp8 which is not supported (even if it's slow), but there are no such option for int8 which is supported for my GPU as log says. Log is a more credible source than random people here.

Which part of the log do you want to see that int8 is being used? Can you explain? ComfyUI doesn't print anything like "Using int8" or similar. If a model is quantized to int8 format and you load it in the Load Diffusion Model node, Comfy will automatically use int8, no additional configuration is needed.

[github-actions]

This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically.