search

kijai / ComfyUI-CogVideoXWrapper

750 stars 45 forks source link

LCM support #41

Closed MGTRIDER closed 2 months ago

MGTRIDER commented 2 months ago

Hi there sir, is it possible to add the LCM sampler as an alternative sampler for faster convergence, thereby speeding up rendering? And maybe even something like HyperTile. But i think that LCM sampler support would already give a big boost. Thank you for your time.

kijai commented 2 months ago

The LCM sampler itself is not faster, it's just the sampling method needed to sample LCM distilled models, and there's nothing like that available for CogVideoX.

MGTRIDER commented 2 months ago

The LCM sampler itself is not faster, it's just the sampling method needed to sample LCM distilled models, and there's nothing like that available for CogVideoX.

I see, thanks for the explanation. Can you help me with the following issue ? Whenever i try to use the fp8 transformer on the CogVideo 2b model, i get the following error: Promotion for Float8 Types is not supported, attempted to promote Half and Float8_e4m3fn

kijai commented 2 months ago

The LCM sampler itself is not faster, it's just the sampling method needed to sample LCM distilled models, and there's nothing like that available for CogVideoX.

I see, thanks for the explanation. Can you help me with the following issue ? Whenever i try to use the fp8 transformer on the CogVideo 2b model, i get the following error: Promotion for Float8 Types is not supported, attempted to promote Half and Float8_e4m3fn

It's not currently supported on the 2b model, there's a workaround but I haven't applied it yet.

MGTRIDER commented 2 months ago

The LCM sampler itself is not faster, it's just the sampling method needed to sample LCM distilled models, and there's nothing like that available for CogVideoX.

I see, thanks for the explanation. Can you help me with the following issue ? Whenever i try to use the fp8 transformer on the CogVideo 2b model, i get the following error: Promotion for Float8 Types is not supported, attempted to promote Half and Float8_e4m3fn

It's not currently supported on the 2b model, there's a workaround but I haven't applied it yet.

Alright, thank you very much, then i will close the issue.

kijai commented 2 months ago

The LCM sampler itself is not faster, it's just the sampling method needed to sample LCM distilled models, and there's nothing like that available for CogVideoX.

I see, thanks for the explanation. Can you help me with the following issue ? Whenever i try to use the fp8 transformer on the CogVideo 2b model, i get the following error: Promotion for Float8 Types is not supported, attempted to promote Half and Float8_e4m3fn

Using fp8 with the 2b model should now be working.