Is there an existing issue for this?
What should this feature add?
Three quantized single-file combinations have neither a load path nor a named refusal:
- nvfp4 FLUX.1 and FLUX.2 transformers. The FLUX loaders pass
nvfp4_payloads={} (load/model_loaders/flux.py),
so packed nvfp4 layers are not decoded.
- int8 (
int8_convrot) Qwen-Image transformer. qwen_image.py has no int8 path for the transformer.
Either load them (the nvfp4 and int8 helpers already serve Z-Image, Krea-2, Qwen-Image nvfp4 and others), or refuse
them at install with a message naming the format, as Ideogram 4 does for nvfp4.
Alternatives
Use the fp8 or GGUF builds of these models.
Additional Content
What such a file does today has not been traced; it may fail with an unclear error. Refusing at install is the
minimum so a multi-GB download does not register as a model that cannot load.
The full matrix of what loads today is in #9726 (Model Format Support page).
Is there an existing issue for this?
What should this feature add?
Three quantized single-file combinations have neither a load path nor a named refusal:
nvfp4_payloads={}(load/model_loaders/flux.py),so packed nvfp4 layers are not decoded.
int8_convrot) Qwen-Image transformer.qwen_image.pyhas no int8 path for the transformer.Either load them (the nvfp4 and int8 helpers already serve Z-Image, Krea-2, Qwen-Image nvfp4 and others), or refuse
them at install with a message naming the format, as Ideogram 4 does for nvfp4.
Alternatives
Use the fp8 or GGUF builds of these models.
Additional Content
What such a file does today has not been traced; it may fail with an unclear error. Refusing at install is the
minimum so a multi-GB download does not register as a model that cannot load.
The full matrix of what loads today is in #9726 (Model Format Support page).