please - do the same to TE
can you, please also convert the Text Encoder into this quant?
I don't think we should in this case, the current nvfp4_awq is a good quant and this would be worse.
I ran into attempts, but the models lacked currently necessary identifiable metadata that would cause inoperable errors (at least with an RDNA4 AMD GPU, comfyui 0.31.0 and after compiling https://github.com/0xDELUXA/comfy-kitchen_win-rocm). But the qwen3vl_32b_minimax_h3-w4a8_convrot over at https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/ works.
Am definitely a fan of w4a8.
Btw, it seems you can use Qwen3 VL 4B as text encoder too, according to this https://www.reddit.com/r/StableDiffusion/s/0JOjvi7TbG
There's this one, and it works fine for me, but I didn't compare it to ComfyUI's NVFP4 checkpoint, so use it at your own discretion:
I ran into attempts, but the models lacked currently necessary identifiable metadata that would cause inoperable errors (at least with an RDNA4 AMD GPU, comfyui 0.31.0 and after compiling https://github.com/0xDELUXA/comfy-kitchen_win-rocm). But the qwen3vl_32b_minimax_h3-w4a8_convrot over at https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/ works.
Am definitely a fan of w4a8.
We don't need to build that repo from source. All the changes are upstream as well. Any INT8 ConvRot model can be converted to W4A8 and will work on AMD with the latest ComfyUI using https://github.com/0xDELUXA/comfy-int8-convrot-to-w4a8.
I ran into attempts, but the models lacked currently necessary identifiable metadata that would cause inoperable errors (at least with an RDNA4 AMD GPU, comfyui 0.31.0 and after compiling https://github.com/0xDELUXA/comfy-kitchen_win-rocm). But the qwen3vl_32b_minimax_h3-w4a8_convrot over at https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/ works.
Am definitely a fan of w4a8.
We don't need to build that repo from source. All the changes are upstream as well. Any INT8 ConvRot model can be converted to W4A8 and will work on AMD with the latest ComfyUI using https://github.com/0xDELUXA/comfy-int8-convrot-to-w4a8.
i do it to avoid settling for any generic jazz, to get the most dialed in kernels and to try new dev branches with a re-used batch file i have been using since minimax h3 publically launched. If you or whoever with a fairly recent AMD GPU happen to catch this post, then youre welcome to consider the quickly created batch file I have been using to compile, which in my case leads to less than a couple minutes to complete the compiling https://pastebin.com/00FaWB8j . I am already checking out the sol-attn addition via this branch https://github.com/0xDELUXA/comfy-kitchen_win-rocm/tree/amd/hip-sol-attn and utilizing this https://github.com/DrBearJew/ComfyUI-SolAttn_triton/tree/h3-community-r01 . It's worth compiling is the point, and also all the work 0xDeluxa does for us RDNA4 and AMD GPU users is awesome and I greatly appreciate it. Same goes for all that Kijai does in general for the whole open source community.
Would be great to see some info about the new comfy-kitchen options for AMD in https://rocm.blogs.amd.com/software-tools-optimization/comfyui-fa-backends/README.html. Just commented on it.