comfyui

Will there be FP8 and NVFP4 of the models?

#5
by VivekThePreceptor - opened

As title suggests will there be release of the FP8 and NVFP4 of the models? Blackwell GPU can get good use of it.

I want to specify that AMD runs INT8. Is the speed the same as FP8? Previously, INT8 acceleration was only supported on NVIDIA cards, so will running INT8 on AMD cards be slower, or will it be roughly comparable to FP8?

I want to specify that AMD runs INT8. Is the speed the same as FP8? Previously, INT8 acceleration was only supported on NVIDIA cards, so will running INT8 on AMD cards be slower, or will it be roughly comparable to FP8?

I'm not an Expert, But that's not True. INT8 is 15% or more faster than FP8 in AMD. Either you're missing something. Check if you have triton enabled or you have a proper setup for ROCm. As per my knowledge this is the case.

I want to specify that AMD runs INT8. Is the speed the same as FP8? Previously, INT8 acceleration was only supported on NVIDIA cards, so will running INT8 on AMD cards be slower, or will it be roughly comparable to FP8?

I'm not an Expert, But that's not True. INT8 is 15% or more faster than FP8 in AMD. Either you're missing something. Check if you have triton enabled or you have a proper setup for ROCm. As per my knowledge this is the case.

Thank you for your reply. I did make a lot of modifications earlier for adjustments, and perhaps the test environments for the two are different. I will test it again later.

Comfy Org org

Plan is to add a 4bit version yes, we're working on something new, something else than nvfp4 most likely.

int4 please

Plan is to add a 4bit version yes, we're working on something new, something else than nvfp4 most likely.

Okay, Now I'm more interested these something else.

It’s been highly optimized by now, but 20 steps is still a bit slow; it would be great if a 4-step acceleration LoRA could be released.

Fp8 version please on my end fp8 versions are near 2x faster

fp8 version please, as our A100:SM80 is approximately three times slower in int8 convRot compared to fp8.
https://github.com/Comfy-Org/ComfyUI/issues/14824.
Or is there any solution to solve the problem on A100?

I want to specify that AMD runs INT8. Is the speed the same as FP8? Previously, INT8 acceleration was only supported on NVIDIA cards, so will running INT8 on AMD cards be slower, or will it be roughly comparable to FP8?

I'm not an Expert, But that's not True. INT8 is 15% or more faster than FP8 in AMD. Either you're missing something. Check if you have triton enabled or you have a proper setup for ROCm. As per my knowledge this is the case.

Faster? On my 9060XT INT8 is (still) waaaay slower than FP8...

As title suggests will there be release of the FP8 and NVFP4 of the models? Blackwell GPU can get good use of it.

It'll probably look like dogtrash compared to the likes of an Unsloth Q8 GGUF for LTX 2.3. NVFP4 models for me always come out looking like crap. I'd rather wait for an MXFP8 which didn't look half bad.

Sign up or log in to comment