Views
No views yet
Qwen3VLModel
(no LM head, tie_word_embeddings=False), so it can only produce hidden states. Graft this
single tensor onto that encoder to make it generative (e.g. for prompt upsampling) without
re-downloading the full model.1from safetensors.torch import load_file
2head = load_file("lm_head.safetensors")["lm_head.weight"]
3# attach to a Qwen3VLForConditionalGeneration shell whose .model is the loaded encoder body