A DataForward[Tensor] object is not a real tensor but rather a descriptor of it. But, users may want to manipulate the materialized tensor with generic operations like slicing or aggregating. A concrete case of this is for Qwen-Image where the first 34 token positions are discarded from the embeddings created by the text encoder (Qwen2.5 VL). Right now this was implemented as a simple parameter skip_tokens in GeneratorInput but this is a temporary hack.
A
DataForward[Tensor]object is not a real tensor but rather a descriptor of it. But, users may want to manipulate the materialized tensor with generic operations like slicing or aggregating. A concrete case of this is for Qwen-Image where the first 34 token positions are discarded from the embeddings created by the text encoder (Qwen2.5 VL). Right now this was implemented as a simple parameterskip_tokensinGeneratorInputbut this is a temporary hack.