bf16 safetensors crash ComfyUI 0.24 (leftover comfy_quant tensors) - cause + fix confirmed
Thanks for getting those quants up so fast.
Found a problem with the two bf16 safetensors (ideogram4_bf16 and ideogram4_unconditional_bf16): they crash ComfyUI 0.24's UNETLoader with
AttributeError: 'NoneType' object has no attribute 'clone'
(raised inside comfy_kitchen during load_state_dict)
Cause, as far as I can tell: the exports still contain the per-layer quantization descriptor tensors from the fp8 source. There are 211 small U8 tensors named *.comfy_quant (27 bytes each) alongside the bf16 weights. ComfyUI sees those and takes its quantized loading path, then fails because the fp8 payload and scales they describe no longer exist. The weights themselves look fine, full bf16, and account for the entire file size.
Confirmed by fixing it locally: re-saving both files without the *.comfy_quant tensors makes them load and generate normally on ComfyUI 0.24.0. So a re-export that just drops those descriptor tensors should fix it for everyone.
Happy to share the small python script I used to strip them if that helps. Let me know!
Ah, thank you for the information.
The BF16 safetensors are mostly there to be used as a source for re-quatization. They are converted from the scaled fp8 weights (the highest precision we've got from Ideogram-ai), unless your hardware doesn't support fp8, there's not really a point in using these for inferenece, as the resulting quality should be almost identical to fp8, mabe even a tiny bit worse.
If you make a PR here to replace them with your fixed BF16 safetensors I might accept them, but I don't think it's worth it on my end to re-upload 36GB just for that.
the gguf doesn't work on the standard workflow (replaced load diffusion for load gguf) it says this:
ValueError: This model is not currently supported - (Unknown model architecture!)
@DarienXccs there are multiple different custom nodes for loeading GGUFs, maybe at least one of them would work?
Probably would have to ask City96 to add Ideogram to the Comfy GGUF node.
https://github.com/city96/ComfyUI-GGUF/issues
(many other nodes that use gguf are based on this node)