nano-gv

Random-init ~16 MB CI fixture for the sendnn-inference plugin's async multimodal-encoder tests. Generated by tools/build_nano_gv.py from ibm-granite/granite-vision-3.2-2b's config.

Not a functional model — outputs are meaningless. Do not use for inference. The architecture identifiers (LlavaNextForConditionalGeneration

• granite text + siglip vision) match the source, but every dim is shrunk to the minimum that still exercises the plugin's MM code paths:

component	source	nano
text hidden	2048	64
text layers	40	2
text heads	32	2
text head_dim	64	32
vision hidden	1152	64
vision layers	27	4
image size	384	112

The tokenizer + processor are copied from the source verbatim (with image-size fields rescaled to 112) so image_token_index and the chat template stay valid.