Is there an existing issue for this?
What should this feature add?
Load (or clearly refuse) nvfp4 builds of the Qwen3-VL 4B/8B encoder. The checkpoint loader passes
nvfp4_payloads={} (load/model_loaders/krea2.py), so packed layers are neither decoded nor refused by name,
while bf16, fp8, int8, MXFP8 and GGUF builds load.
Alternatives
Use the fp8, int8 or GGUF encoder.
Additional Content
The Qwen3 and Mistral encoder loaders already keep nvfp4 packed (install_nvfp4_layers) and are the template.
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?
Load (or clearly refuse) nvfp4 builds of the Qwen3-VL 4B/8B encoder. The checkpoint loader passes
nvfp4_payloads={}(load/model_loaders/krea2.py), so packed layers are neither decoded nor refused by name,while bf16, fp8, int8, MXFP8 and GGUF builds load.
Alternatives
Use the fp8, int8 or GGUF encoder.
Additional Content
The Qwen3 and Mistral encoder loaders already keep nvfp4 packed (
install_nvfp4_layers) and are the template.The full matrix of what loads today is in #9726 (Model Format Support page).