FP16 files
Hello, first I would like to say thank you for the speedy implementation in ComfyUI. Are FP16 weights planned for these files?
Thanks in advance!
EDIT:
I've found and downloaded supposed Qwen-Image-FP16 from another website. Is this normal behavior in ComfyUI? (dtype bf16 -> manual cast to fp32) Feels awfully slow, forcing fp16 through the manager results in black outputs. Wan2.2 FP16 correctly inferences fp16 without manual cast to fp32.
You said you downloaded fp16, but that model isn't fp16, its bf16 (from your screenshot: model weight dtype torch.bfloat16). Does your gpu support bfloat? if it doesn't the bfloat support will be emulated in software and computation done in fp32. You should get an actual fp16 model if that's what you need.
You said you downloaded fp16, but that model isn't fp16, its bf16 (from your screenshot: model weight dtype torch.bfloat16). Does your gpu support bfloat? if it doesn't the bfloat support will be emulated in software and computation done in fp32. You should get an actual fp16 model if that's what you need.
I was wondering if it was a correct fp16 version I downloaded. My gpu doesn’t support bfloat16 (20 series), I was going to upgrade to a 50 series by the end of the year anyway. I assumed it would support fp16 because wan2.2 does.
I don't understand why they forcing us to use FP8 clip model without providing us with additional better FP16 variant I though original was in FP16 but I have no clue where to download it so it would work in ComfyUI without: clip missing: ['model.layers.5.mlp.up_proj.weight', messages because its original checkpoints not working properly when I trying to load them downloaded straight from HugginFace.
Some info about Qwen Image model training from official technical documentation:
(During training, all-gather operations are performed in bfloat16, while gradient reduce-scatter operations utilize float32,)
Edit: Looks like they added FP16 variant of clip model so will download it straight away.
You said you downloaded fp16, but that model isn't fp16, its bf16 (from your screenshot: model weight dtype torch.bfloat16). Does your gpu support bfloat? if it doesn't the bfloat support will be emulated in software and computation done in fp32. You should get an actual fp16 model if that's what you need.
I was wondering if it was a correct fp16 version I downloaded. My gpu doesn’t support bfloat16 (20 series), I was going to upgrade to a 50 series by the end of the year anyway. I assumed it would support fp16 because wan2.2 does.
Wish they release FP16 Qwen Image model variant too I prefer FP16 because I usually run those large models on my CPU instead GPU and my GPU don't support BF16.
Just don't wanna it to cast to FP32 instead FP16 when I run it in CPU mode on my DDR4 memory.
Update:
Found distilled from the BF16 model FP16 model variant here:
https://civitai.com/models/1841092/qwen-image-fp16