Just did a quick test
Everything works!
Motion is perhaps slightly less smooth compared to fp8 - more testing needed.
Performance is about 5% faster than fp8 on my 5090.
Many thanks for quanting this.
You're welcome. I did about 10 sample videos and found it to be pretty good.
For me with my 4060 ti (16gb) your wan int8 convrot conversion gave around a 9-10% speed up compared to the standard fp8_scaled. I'll take it! 😄 The int8 text encoder however doesn't seem to help with the speed, at least not for me.
I'm thinking a comfyui update could maybe give int8 convrot a bit more speed. Not sure if the lora slowdown related to int8 has been fully solved yet. I do however know that bypassing my lora loaders (and loras) gave an additional speed boost of around 10-15% but this could just be because of the freed up system resources.
Torch compiling also did nothing for me but it seldom does.. Maybe it's my entry level pc specs that's just unable to take full advantage of it.
Changing to comfyui's default lora loader maybe gave me a couple of seconds faster inference but that could just be within the margin of error. I usually use the power lora loader.
Alright then. 😀 I really only wanted to say thank you for what you are doing, I really appreciate it!
Is there any technical problem with T2V quantization? There has never been a T2V model in INT-8. (Even the old official ones only had i2V)