Issue with zimage model

[chirusandeep]

Hi @robertvoy
Found this issue of tensor size with z-image model.

RuntimeError: The size of tensor a (96) must match the size of tensor b (104) at non-singleton dimension 3

[robertvoy]

Can you send me the workflow you get that error with?

[chirusandeep]

Can you send me the workflow you get that error with?

zimagedistributed.json

[robertvoy]

I finally got a chance to try the workflow, and I'm not sure if it's a GGUF issue, but using just standard safetensors, everything works fine with Z-image.

Also, please note that in your workflow, the distributed collector is not necessary when using the Distributed Upscaler.

[chirusandeep]

I finally got a chance to try the workflow, and I'm not sure if it's a GGUF issue, but using just standard safetensors, everything works fine with Z-image.

Also, please note that in your workflow, the distributed collector is not necessary when using the Distributed Upscaler.

Yeah got it, there was something just want to request if possible.

1. What if like "distributed clean vram"?
2. when ran wan2.2, can we reuse same cache stored in ram by all VRAM GPUs? because the ram cache being created multiple times for each GPU.

Waiting for your reply, Thanks.

[robertvoy]

Currently, each local worker creates their own ram cache.

Clean vram, just unloads the models from the workers and into ram.

[chirusandeep]

Tried clean vram after node of distributed collector, it just clean up GPU:0 not GPU:1. any sample workflow, could you please try that?

[robertvoy]

There is a clean vram button in the distributed UI panel...

[chirusandeep]

I got that it was manual trigger, my case was using comfyui workflow by api. I have used "clean vram" that node only cleans the VRAM of GPU:0 and not rest.
if possible? "distributed clean vram" node, would be helpful!
Thanks @robertvoy

[robertvoy]

Got it, alright, I will consider adding this. Thank you.

[robertvoy]

image_z_image_turbo-distributed-clearVRAM.json

Just had a chance to check this and you can use a clear VRAM node to clear VRAM on all workers.